Project import generated by Copybara.

GitOrigin-RevId: 796203faee20d7aae2876aac8ca5a1827dee4fe3

WORKSPACE

@@ -64,6 +64,12 @@ http_archive(
     sha256 = "267103f8a1e9578978aa1dc256001e6529ef593e5aea38193d31c2872ee025e8",
     strip_prefix = "glog-0.3.5",
     build_file = "@//third_party:glog.BUILD",
+    patches = [
+        "@//third_party:com_github_glog_glog_9779e5ea6ef59562b030248947f787d1256132ae.diff"
+    ],
+    patch_args = [
+        "-p1",
+    ],
 )
 
 # libyuv

mediapipe/calculators/audio/audio_decoder_calculator.cc

@@ -61,7 +61,9 @@ class AudioDecoderCalculator : public CalculatorBase {
 ::mediapipe::Status AudioDecoderCalculator::GetContract(
     CalculatorContract* cc) {
   cc->InputSidePackets().Tag("INPUT_FILE_PATH").Set<std::string>();
-
+  if (cc->InputSidePackets().HasTag("OPTIONS")) {
+    cc->InputSidePackets().Tag("OPTIONS").Set<mediapipe::AudioDecoderOptions>();
+  }
   cc->Outputs().Tag("AUDIO").Set<Matrix>();
   if (cc->Outputs().HasTag("AUDIO_HEADER")) {
     cc->Outputs().Tag("AUDIO_HEADER").SetNone();
@@ -72,7 +74,9 @@ class AudioDecoderCalculator : public CalculatorBase {
 ::mediapipe::Status AudioDecoderCalculator::Open(CalculatorContext* cc) {
   const std::string& input_file_path =
       cc->InputSidePackets().Tag("INPUT_FILE_PATH").Get<std::string>();
-  const auto& decoder_options = cc->Options<mediapipe::AudioDecoderOptions>();
+  const auto& decoder_options =
+      tool::RetrieveOptions(cc->Options<mediapipe::AudioDecoderOptions>(),
+                            cc->InputSidePackets(), "OPTIONS");
   decoder_ = absl::make_unique<AudioDecoder>();
   MP_RETURN_IF_ERROR(decoder_->Initialize(input_file_path, decoder_options));
   std::unique_ptr<mediapipe::TimeSeriesHeader> header =

mediapipe/calculators/audio/stabilized_log_calculator.cc

@@ -75,8 +75,13 @@ class StabilizedLogCalculator : public CalculatorBase {
 
   ::mediapipe::Status Process(CalculatorContext* cc) override {
     auto input_matrix = cc->Inputs().Index(0).Get<Matrix>();
+    if (input_matrix.array().isNaN().any()) {
+      return ::mediapipe::InvalidArgumentError("NaN input to log operation.");
+    }
     if (check_nonnegativity_) {
-      CHECK_GE(input_matrix.minCoeff(), 0);
+      if (input_matrix.minCoeff() < 0.0) {
+        return ::mediapipe::OutOfRangeError("Negative input to log operation.");
+      }
     }
     std::unique_ptr<Matrix> output_frame(new Matrix(
         output_scale_ * (input_matrix.array() + stabilizer_).log().matrix()));

mediapipe/calculators/audio/stabilized_log_calculator_test.cc

@@ -11,6 +11,7 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
+#include <cmath>
 
 #include "Eigen/Core"
 #include "mediapipe/calculators/audio/stabilized_log_calculator.pb.h"
@@ -108,13 +109,22 @@ TEST_F(StabilizedLogCalculatorTest, ZerosAreStabilized) {
       runner_->Outputs().Index(0).packets[0].Get<Matrix>());
 }
 
-TEST_F(StabilizedLogCalculatorTest, NegativeValuesCheckFail) {
+TEST_F(StabilizedLogCalculatorTest, NanValuesReturnError) {
+  InitializeGraph();
+  FillInputHeader();
+  AppendInputPacket(
+      new Matrix(Matrix::Constant(kNumChannels, kNumSamples, std::nanf(""))),
+      0 /* timestamp */);
+  ASSERT_FALSE(RunGraph().ok());
+}
+
+TEST_F(StabilizedLogCalculatorTest, NegativeValuesReturnError) {
   InitializeGraph();
   FillInputHeader();
   AppendInputPacket(
       new Matrix(Matrix::Constant(kNumChannels, kNumSamples, -1.0)),
       0 /* timestamp */);
-  ASSERT_DEATH(RunGraphNoReturn(), "");
+  ASSERT_FALSE(RunGraph().ok());
 }
 
 TEST_F(StabilizedLogCalculatorTest, NegativeValuesDoNotCheckFailIfCheckIsOff) {

mediapipe/calculators/audio/time_series_framer_calculator.cc

@@ -56,6 +56,14 @@ namespace mediapipe {
 // If pad_final_packet is true, all input samples will be emitted and the final
 // packet will be zero padded as necessary.  If pad_final_packet is false, some
 // samples may be dropped at the end of the stream.
+//
+// If use_local_timestamp is true, the output packet's timestamp is based on the
+// last sample of the packet. The timestamp of this sample is inferred by
+// input_packet_timesamp + local_sample_index / sampling_rate_. If false, the
+// output packet's timestamp is based on the cumulative timestamping, which is
+// done by adopting the timestamp of the first sample of the packet and this
+// sample's timestamp is inferred by initial_input_timestamp_ +
+// cumulative_completed_samples / sample_rate_.
 class TimeSeriesFramerCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
@@ -86,11 +94,26 @@ class TimeSeriesFramerCalculator : public CalculatorBase {
   void FrameOutput(CalculatorContext* cc);
 
   Timestamp CurrentOutputTimestamp() {
+    if (use_local_timestamp_) {
+      return current_timestamp_;
+    }
+    return CumulativeOutputTimestamp();
+  }
+
+  Timestamp CumulativeOutputTimestamp() {
     return initial_input_timestamp_ +
            round(cumulative_completed_samples_ / sample_rate_ *
                  Timestamp::kTimestampUnitsPerSecond);
   }
 
+  // Returns the timestamp of a sample on a base, which is usually the time
+  // stamp of a packet.
+  Timestamp CurrentSampleTimestamp(const Timestamp& timestamp_base,
+                                   int64 number_of_samples) {
+    return timestamp_base + round(number_of_samples / sample_rate_ *
+                                  Timestamp::kTimestampUnitsPerSecond);
+  }
+
   // The number of input samples to advance after the current output frame is
   // emitted.
   int next_frame_step_samples() const {
@@ -118,14 +141,18 @@ class TimeSeriesFramerCalculator : public CalculatorBase {
   // any overlap).
   int64 cumulative_completed_samples_;
   Timestamp initial_input_timestamp_;
+  // The current timestamp is updated along with the incoming packets.
+  Timestamp current_timestamp_;
   int num_channels_;
 
   // Each entry in this deque consists of a single sample, i.e. a
-  // single column vector.
+  // single column vector, and its timestamp.
-  std::deque<Matrix> sample_buffer_;
+  std::deque<std::pair<Matrix, Timestamp>> sample_buffer_;
 
   bool use_window_;
   Matrix window_;
+
+  bool use_local_timestamp_;
 };
 REGISTER_CALCULATOR(TimeSeriesFramerCalculator);
 
@@ -133,7 +160,8 @@ void TimeSeriesFramerCalculator::EnqueueInput(CalculatorContext* cc) {
   const Matrix& input_frame = cc->Inputs().Index(0).Get<Matrix>();
 
   for (int i = 0; i < input_frame.cols(); ++i) {
-    sample_buffer_.emplace_back(input_frame.col(i));
+    sample_buffer_.emplace_back(std::make_pair(
+        input_frame.col(i), CurrentSampleTimestamp(cc->InputTimestamp(), i)));
   }
 
   cumulative_input_samples_ += input_frame.cols();
@@ -151,14 +179,16 @@ void TimeSeriesFramerCalculator::FrameOutput(CalculatorContext* cc) {
         new Matrix(num_channels_, frame_duration_samples_));
     for (int i = 0; i < std::min(frame_step_samples, frame_duration_samples_);
          ++i) {
-      output_frame->col(i) = sample_buffer_.front();
+      output_frame->col(i) = sample_buffer_.front().first;
+      current_timestamp_ = sample_buffer_.front().second;
       sample_buffer_.pop_front();
     }
     const int frame_overlap_samples =
         frame_duration_samples_ - frame_step_samples;
     if (frame_overlap_samples > 0) {
       for (int i = 0; i < frame_overlap_samples; ++i) {
-        output_frame->col(i + frame_step_samples) = sample_buffer_[i];
+        output_frame->col(i + frame_step_samples) = sample_buffer_[i].first;
+        current_timestamp_ = sample_buffer_[i].second;
       }
     } else {
       samples_still_to_drop_ = -frame_overlap_samples;
@@ -178,6 +208,7 @@ void TimeSeriesFramerCalculator::FrameOutput(CalculatorContext* cc) {
 ::mediapipe::Status TimeSeriesFramerCalculator::Process(CalculatorContext* cc) {
   if (initial_input_timestamp_ == Timestamp::Unstarted()) {
     initial_input_timestamp_ = cc->InputTimestamp();
+    current_timestamp_ = initial_input_timestamp_;
   }
 
   EnqueueInput(cc);
@@ -195,7 +226,8 @@ void TimeSeriesFramerCalculator::FrameOutput(CalculatorContext* cc) {
     std::unique_ptr<Matrix> output_frame(new Matrix);
     output_frame->setZero(num_channels_, frame_duration_samples_);
     for (int i = 0; i < sample_buffer_.size(); ++i) {
-      output_frame->col(i) = sample_buffer_[i];
+      output_frame->col(i) = sample_buffer_[i].first;
+      current_timestamp_ = sample_buffer_[i].second;
     }
 
     cc->Outputs().Index(0).Add(output_frame.release(),
@@ -258,6 +290,7 @@ void TimeSeriesFramerCalculator::FrameOutput(CalculatorContext* cc) {
   cumulative_output_frames_ = 0;
   samples_still_to_drop_ = 0;
   initial_input_timestamp_ = Timestamp::Unstarted();
+  current_timestamp_ = Timestamp::Unstarted();
 
   std::vector<double> window_vector;
   use_window_ = false;
@@ -282,6 +315,7 @@ void TimeSeriesFramerCalculator::FrameOutput(CalculatorContext* cc) {
                                           frame_duration_samples_)
                   .cast<float>();
   }
+  use_local_timestamp_ = framer_options.use_local_timestamp();
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/audio/time_series_framer_calculator.proto

@@ -62,4 +62,11 @@ message TimeSeriesFramerCalculatorOptions {
     HANN = 2;
   }
   optional WindowFunction window_function = 4 [default = NONE];
+
+  // If use_local_timestamp is true, the output packet's timestamp is based on
+  // the last sample of the packet and it's inferred from the latest input
+  // packet's timestamp.  If false, the output packet's timestamp is based on
+  // the cumulative timestamping, which is inferred from the intial input
+  // timestamp and the cumulative number of samples.
+  optional bool use_local_timestamp = 6 [default = false];
 }

mediapipe/calculators/audio/time_series_framer_calculator_test.cc

@@ -35,6 +35,8 @@ namespace mediapipe {
 namespace {
 
 const int kInitialTimestampOffsetMicroseconds = 4;
+const int kGapBetweenPacketsInSeconds = 1;
+const int kUniversalInputPacketSize = 50;
 
 class TimeSeriesFramerCalculatorTest
     : public TimeSeriesCalculatorTest<TimeSeriesFramerCalculatorOptions> {
@@ -391,5 +393,93 @@ TEST_F(TimeSeriesFramerCalculatorWindowingSanityTest, HannWindowSanityCheck) {
   RunAndTestSinglePacketAverage(0.5f);
 }
 
-}  // anonymous namespace
+// A simple test class that checks the local packet time stamp. This class
+// generate a series of packets with and without gaps between packets and tests
+// the behavior with cumulative timestamping and local packet timestamping.
+class TimeSeriesFramerCalculatorTimestampingTest
+    : public TimeSeriesFramerCalculatorTest {
+ protected:
+  // Creates test input and saves a reference copy.
+  void InitializeInputForTimeStampingTest() {
+    concatenated_input_samples_.resize(0, num_input_channels_);
+    num_input_samples_ = 0;
+    for (int i = 0; i < 10; ++i) {
+      // This range of packet sizes was chosen such that some input
+      // packets will be smaller than the output packet size and other
+      // input packets will be larger.
+      int packet_size = kUniversalInputPacketSize;
+      double timestamp_seconds = kInitialTimestampOffsetMicroseconds * 1.0e-6 +
+                                 num_input_samples_ / input_sample_rate_;
+      if (options_.use_local_timestamp()) {
+        timestamp_seconds += kGapBetweenPacketsInSeconds * i;
+      }
+
+      Matrix* data_frame =
+          NewTestFrame(num_input_channels_, packet_size, timestamp_seconds);
+
+      AppendInputPacket(data_frame, round(timestamp_seconds *
+                                          Timestamp::kTimestampUnitsPerSecond));
+      num_input_samples_ += packet_size;
+    }
+  }
+
+  void CheckOutputTimestamps() {
+    int num_full_packets = output().packets.size();
+    if (options_.pad_final_packet()) {
+      num_full_packets -= 1;
+    }
+
+    int64 num_samples = 0;
+    for (int packet_num = 0; packet_num < num_full_packets; ++packet_num) {
+      const Packet& packet = output().packets[packet_num];
+      num_samples += FrameDurationSamples();
+      double expected_timestamp =
+          options_.use_local_timestamp()
+              ? GetExpectedLocalTimestampForSample(num_samples - 1)
+              : GetExpectedCumulativeTimestamp(num_samples - 1);
+      ASSERT_NEAR(packet.Timestamp().Seconds(), expected_timestamp, 1e-10);
+    }
+  }
+
+  ::mediapipe::Status RunTimestampTest() {
+    InitializeGraph();
+    InitializeInputForTimeStampingTest();
+    FillInputHeader();
+    return RunGraph();
+  }
+
+ private:
+  // Returns the timestamp in seconds based on local timestamping.
+  double GetExpectedLocalTimestampForSample(int sample_index) {
+    return kInitialTimestampOffsetMicroseconds * 1.0e-6 +
+           sample_index / input_sample_rate_ +
+           (sample_index / kUniversalInputPacketSize) *
+               kGapBetweenPacketsInSeconds;
+  }
+
+  // Returns the timestamp inseconds based on cumulative timestamping.
+  double GetExpectedCumulativeTimestamp(int sample_index) {
+    return kInitialTimestampOffsetMicroseconds * 1.0e-6 +
+           sample_index / FrameDurationSamples() * FrameDurationSamples() /
+               input_sample_rate_;
+  }
+};
+
+TEST_F(TimeSeriesFramerCalculatorTimestampingTest, UseLocalTimeStamp) {
+  options_.set_frame_duration_seconds(100.0 / input_sample_rate_);
+  options_.set_use_local_timestamp(true);
+
+  MP_ASSERT_OK(RunTimestampTest());
+  CheckOutputTimestamps();
+}
+
+TEST_F(TimeSeriesFramerCalculatorTimestampingTest, UseCumulativeTimeStamp) {
+  options_.set_frame_duration_seconds(100.0 / input_sample_rate_);
+  options_.set_use_local_timestamp(false);
+
+  MP_ASSERT_OK(RunTimestampTest());
+  CheckOutputTimestamps();
+}
+
+}  // namespace
 }  // namespace mediapipe

mediapipe/calculators/core/BUILD

@@ -166,7 +166,13 @@ cc_library(
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
         "@org_tensorflow//tensorflow/lite:framework",
+    ] + select({
+        "//mediapipe/gpu:disable_gpu": [],
+        "//mediapipe:ios": [],
+        "//conditions:default": [
+            "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_buffer",
         ],
+    }),
     alwayslink = 1,
 )
 

mediapipe/calculators/core/concatenate_vector_calculator.cc

@@ -19,6 +19,10 @@
 #include "mediapipe/framework/formats/landmark.pb.h"
 #include "tensorflow/lite/interpreter.h"
 
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+#include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+
 namespace mediapipe {
 
 // Example config:
@@ -45,4 +49,11 @@ REGISTER_CALCULATOR(ConcatenateTfLiteTensorVectorCalculator);
 typedef ConcatenateVectorCalculator<::mediapipe::NormalizedLandmark>
     ConcatenateLandmarkVectorCalculator;
 REGISTER_CALCULATOR(ConcatenateLandmarkVectorCalculator);
+
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+typedef ConcatenateVectorCalculator<::tflite::gpu::gl::GlBuffer>
+    ConcatenateGlBufferVectorCalculator;
+REGISTER_CALCULATOR(ConcatenateGlBufferVectorCalculator);
+#endif
+
 }  // namespace mediapipe

mediapipe/calculators/core/concatenate_vector_calculator.h

@@ -15,6 +15,7 @@
 #ifndef MEDIAPIPE_CALCULATORS_CORE_CONCATENATE_VECTOR_CALCULATOR_H_
 #define MEDIAPIPE_CALCULATORS_CORE_CONCATENATE_VECTOR_CALCULATOR_H_
 
+#include <type_traits>
 #include <vector>
 
 #include "mediapipe/calculators/core/concatenate_vector_calculator.pb.h"
@@ -59,16 +60,58 @@ class ConcatenateVectorCalculator : public CalculatorBase {
         if (cc->Inputs().Index(i).IsEmpty()) return ::mediapipe::OkStatus();
       }
     }
-    auto output = absl::make_unique<std::vector<T>>();
+
+    return ConcatenateVectors<T>(std::is_copy_constructible<T>(), cc);
+  }
+
+  template <typename U>
+  ::mediapipe::Status ConcatenateVectors(std::true_type,
+                                         CalculatorContext* cc) {
+    auto output = absl::make_unique<std::vector<U>>();
     for (int i = 0; i < cc->Inputs().NumEntries(); ++i) {
       if (cc->Inputs().Index(i).IsEmpty()) continue;
-      const std::vector<T>& input = cc->Inputs().Index(i).Get<std::vector<T>>();
+      const std::vector<U>& input = cc->Inputs().Index(i).Get<std::vector<U>>();
       output->insert(output->end(), input.begin(), input.end());
     }
     cc->Outputs().Index(0).Add(output.release(), cc->InputTimestamp());
     return ::mediapipe::OkStatus();
   }
 
+  template <typename U>
+  ::mediapipe::Status ConcatenateVectors(std::false_type,
+                                         CalculatorContext* cc) {
+    return ConsumeAndConcatenateVectors<T>(std::is_move_constructible<U>(), cc);
+  }
+
+  template <typename U>
+  ::mediapipe::Status ConsumeAndConcatenateVectors(std::true_type,
+                                                   CalculatorContext* cc) {
+    auto output = absl::make_unique<std::vector<U>>();
+    for (int i = 0; i < cc->Inputs().NumEntries(); ++i) {
+      if (cc->Inputs().Index(i).IsEmpty()) continue;
+      ::mediapipe::StatusOr<std::unique_ptr<std::vector<U>>> input_status =
+          cc->Inputs().Index(i).Value().Consume<std::vector<U>>();
+      if (input_status.ok()) {
+        std::unique_ptr<std::vector<U>> input_vector =
+            std::move(input_status).ValueOrDie();
+        output->insert(output->end(),
+                       std::make_move_iterator(input_vector->begin()),
+                       std::make_move_iterator(input_vector->end()));
+      } else {
+        return input_status.status();
+      }
+    }
+    cc->Outputs().Index(0).Add(output.release(), cc->InputTimestamp());
+    return ::mediapipe::OkStatus();
+  }
+
+  template <typename U>
+  ::mediapipe::Status ConsumeAndConcatenateVectors(std::false_type,
+                                                   CalculatorContext* cc) {
+    return ::mediapipe::InternalError(
+        "Cannot copy or move input vectors to concatenate them");
+  }
+
  private:
   bool only_emit_if_all_present_;
 };

mediapipe/calculators/core/concatenate_vector_calculator_test.cc

@@ -235,4 +235,167 @@ TEST(ConcatenateFloatVectorCalculatorTest, OneEmptyStreamNoOutput) {
   EXPECT_EQ(0, outputs.size());
 }
 
+typedef ConcatenateVectorCalculator<std::unique_ptr<int>>
+    TestConcatenateUniqueIntPtrCalculator;
+REGISTER_CALCULATOR(TestConcatenateUniqueIntPtrCalculator);
+
+TEST(TestConcatenateUniqueIntVectorCalculatorTest, ConsumeOneTimestamp) {
+  /* Note: We don't use CalculatorRunner for this test because it keeps copies
+   * of input packets, so packets sent to the graph don't have sole ownership.
+   * The test needs to send packets that own the data.
+   */
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "in_1"
+        input_stream: "in_2"
+        input_stream: "in_3"
+        node {
+          calculator: "TestConcatenateUniqueIntPtrCalculator"
+          input_stream: "in_1"
+          input_stream: "in_2"
+          input_stream: "in_3"
+          output_stream: "out"
+        }
+      )");
+
+  std::vector<Packet> outputs;
+  tool::AddVectorSink("out", &graph_config, &outputs);
+
+  CalculatorGraph graph;
+  MP_EXPECT_OK(graph.Initialize(graph_config));
+  MP_EXPECT_OK(graph.StartRun({}));
+
+  // input1 : {0, 1, 2}
+  std::unique_ptr<std::vector<std::unique_ptr<int>>> input_1 =
+      absl::make_unique<std::vector<std::unique_ptr<int>>>(3);
+  for (int i = 0; i < 3; ++i) {
+    input_1->at(i) = absl::make_unique<int>(i);
+  }
+  // input2: {3}
+  std::unique_ptr<std::vector<std::unique_ptr<int>>> input_2 =
+      absl::make_unique<std::vector<std::unique_ptr<int>>>(1);
+  input_2->at(0) = absl::make_unique<int>(3);
+  // input3: {4, 5}
+  std::unique_ptr<std::vector<std::unique_ptr<int>>> input_3 =
+      absl::make_unique<std::vector<std::unique_ptr<int>>>(2);
+  input_3->at(0) = absl::make_unique<int>(4);
+  input_3->at(1) = absl::make_unique<int>(5);
+
+  MP_EXPECT_OK(graph.AddPacketToInputStream(
+      "in_1", Adopt(input_1.release()).At(Timestamp(1))));
+  MP_EXPECT_OK(graph.AddPacketToInputStream(
+      "in_2", Adopt(input_2.release()).At(Timestamp(1))));
+  MP_EXPECT_OK(graph.AddPacketToInputStream(
+      "in_3", Adopt(input_3.release()).At(Timestamp(1))));
+
+  MP_EXPECT_OK(graph.WaitUntilIdle());
+  MP_EXPECT_OK(graph.CloseAllPacketSources());
+  MP_EXPECT_OK(graph.WaitUntilDone());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  const std::vector<std::unique_ptr<int>>& result =
+      outputs[0].Get<std::vector<std::unique_ptr<int>>>();
+  EXPECT_EQ(6, result.size());
+  for (int i = 0; i < 6; ++i) {
+    const std::unique_ptr<int>& v = result[i];
+    EXPECT_EQ(i, *v);
+  }
+}
+
+TEST(TestConcatenateUniqueIntVectorCalculatorTest, OneEmptyStreamStillOutput) {
+  /* Note: We don't use CalculatorRunner for this test because it keeps copies
+   * of input packets, so packets sent to the graph don't have sole ownership.
+   * The test needs to send packets that own the data.
+   */
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "in_1"
+        input_stream: "in_2"
+        node {
+          calculator: "TestConcatenateUniqueIntPtrCalculator"
+          input_stream: "in_1"
+          input_stream: "in_2"
+          output_stream: "out"
+        }
+      )");
+
+  std::vector<Packet> outputs;
+  tool::AddVectorSink("out", &graph_config, &outputs);
+
+  CalculatorGraph graph;
+  MP_EXPECT_OK(graph.Initialize(graph_config));
+  MP_EXPECT_OK(graph.StartRun({}));
+
+  // input1 : {0, 1, 2}
+  std::unique_ptr<std::vector<std::unique_ptr<int>>> input_1 =
+      absl::make_unique<std::vector<std::unique_ptr<int>>>(3);
+  for (int i = 0; i < 3; ++i) {
+    input_1->at(i) = absl::make_unique<int>(i);
+  }
+
+  MP_EXPECT_OK(graph.AddPacketToInputStream(
+      "in_1", Adopt(input_1.release()).At(Timestamp(1))));
+
+  MP_EXPECT_OK(graph.WaitUntilIdle());
+  MP_EXPECT_OK(graph.CloseAllPacketSources());
+  MP_EXPECT_OK(graph.WaitUntilDone());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  const std::vector<std::unique_ptr<int>>& result =
+      outputs[0].Get<std::vector<std::unique_ptr<int>>>();
+  EXPECT_EQ(3, result.size());
+  for (int i = 0; i < 3; ++i) {
+    const std::unique_ptr<int>& v = result[i];
+    EXPECT_EQ(i, *v);
+  }
+}
+
+TEST(TestConcatenateUniqueIntVectorCalculatorTest, OneEmptyStreamNoOutput) {
+  /* Note: We don't use CalculatorRunner for this test because it keeps copies
+   * of input packets, so packets sent to the graph don't have sole ownership.
+   * The test needs to send packets that own the data.
+   */
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "in_1"
+        input_stream: "in_2"
+        node {
+          calculator: "TestConcatenateUniqueIntPtrCalculator"
+          input_stream: "in_1"
+          input_stream: "in_2"
+          output_stream: "out"
+          options {
+            [mediapipe.ConcatenateVectorCalculatorOptions.ext] {
+              only_emit_if_all_present: true
+            }
+          }
+        }
+      )");
+
+  std::vector<Packet> outputs;
+  tool::AddVectorSink("out", &graph_config, &outputs);
+
+  CalculatorGraph graph;
+  MP_EXPECT_OK(graph.Initialize(graph_config));
+  MP_EXPECT_OK(graph.StartRun({}));
+
+  // input1 : {0, 1, 2}
+  std::unique_ptr<std::vector<std::unique_ptr<int>>> input_1 =
+      absl::make_unique<std::vector<std::unique_ptr<int>>>(3);
+  for (int i = 0; i < 3; ++i) {
+    input_1->at(i) = absl::make_unique<int>(i);
+  }
+
+  MP_EXPECT_OK(graph.AddPacketToInputStream(
+      "in_1", Adopt(input_1.release()).At(Timestamp(1))));
+
+  MP_EXPECT_OK(graph.WaitUntilIdle());
+  MP_EXPECT_OK(graph.CloseAllPacketSources());
+  MP_EXPECT_OK(graph.WaitUntilDone());
+
+  EXPECT_EQ(0, outputs.size());
+}
+
 }  // namespace mediapipe

mediapipe/calculators/image/BUILD

@@ -19,7 +19,6 @@ package(default_visibility = ["//visibility:private"])
 exports_files(["LICENSE"])
 
 load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-load("@bazel_skylib//lib:selects.bzl", "selects")
 
 proto_library(
     name = "opencv_image_encoder_calculator_proto",
@@ -227,19 +226,13 @@ cc_library(
         "//mediapipe/framework/port:status",
         "//mediapipe/framework/port:vector",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
-            "//mediapipe/gpu:gpu_buffer",
+            "//mediapipe/gpu:gl_quad_renderer",
             "//mediapipe/gpu:shader_util",
         ],
-        "//mediapipe:ios": [
-            "//mediapipe/gpu:gl_calculator_helper",
-            "//mediapipe/gpu:gl_simple_shaders",
-            "//mediapipe/gpu:gpu_buffer",
-            "//mediapipe/gpu:shader_util",
-        ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -263,13 +256,13 @@ cc_library(
         "//mediapipe/framework/port:status",
         "//mediapipe/framework/port:vector",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
-            "//mediapipe/gpu:gpu_buffer",
+            "//mediapipe/gpu:gl_quad_renderer",
             "//mediapipe/gpu:shader_util",
         ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -322,14 +315,14 @@ cc_library(
         "//mediapipe/framework/port:opencv_imgproc",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
-    ] + selects.with_or({
+    ] + select({
-        ("//mediapipe:android", "//mediapipe:ios"): [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
             "//mediapipe/gpu:gl_quad_renderer",
             "//mediapipe/gpu:shader_util",
         ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -363,14 +356,15 @@ cc_library(
         "//mediapipe/framework/port:opencv_imgproc",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
-    ] + selects.with_or({
+        "//mediapipe/gpu:gpu_buffer",
-        ("//mediapipe:android", "//mediapipe:ios"): [
+    ] + select({
+        "//mediapipe/gpu:disable_gpu": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
             "//mediapipe/gpu:gl_quad_renderer",
             "//mediapipe/gpu:shader_util",
         ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -415,19 +409,13 @@ cc_library(
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/util:color_cc_proto",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
-            "//mediapipe/gpu:gpu_buffer",
+            "//mediapipe/gpu:gl_quad_renderer",
             "//mediapipe/gpu:shader_util",
         ],
-        "//mediapipe:ios": [
-            "//mediapipe/gpu:gl_calculator_helper",
-            "//mediapipe/gpu:gl_simple_shaders",
-            "//mediapipe/gpu:gpu_buffer",
-            "//mediapipe/gpu:shader_util",
-        ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -486,11 +474,11 @@ cc_library(
         "//mediapipe/framework/formats:image_frame",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
-    ] + selects.with_or({
+    ] + select({
-        ("//mediapipe:android", "//mediapipe:ios"): [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gpu_buffer",
         ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )

mediapipe/calculators/image/bilateral_filter_calculator.cc

@@ -27,11 +27,11 @@
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/framework/port/vector.h"
 
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
 #include "mediapipe/gpu/shader_util.h"
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace mediapipe {
 
@@ -101,11 +101,11 @@ class BilateralFilterCalculator : public CalculatorBase {
 
   bool use_gpu_ = false;
   bool gpu_initialized_ = false;
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   mediapipe::GlCalculatorHelper gpu_helper_;
   GLuint program_ = 0;
   GLuint program_joint_ = 0;
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 };
 REGISTER_CALCULATOR(BilateralFilterCalculator);
 
@@ -122,39 +122,46 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
     return ::mediapipe::InternalError("GPU output must have GPU input.");
   }
 
+  bool use_gpu = false;
+
   // Input image to filter.
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag(kInputFrameTagGpu)) {
     cc->Inputs().Tag(kInputFrameTagGpu).Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag(kInputFrameTag)) {
     cc->Inputs().Tag(kInputFrameTag).Set<ImageFrame>();
   }
 
   // Input guide image mask (optional)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag(kInputGuideTagGpu)) {
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
     cc->Inputs().Tag(kInputGuideTagGpu).Set<mediapipe::GpuBuffer>();
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+    use_gpu |= true;
   }
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag(kInputGuideTag)) {
     cc->Inputs().Tag(kInputGuideTag).Set<ImageFrame>();
   }
 
   // Output image.
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Outputs().HasTag(kOutputFrameTagGpu)) {
     cc->Outputs().Tag(kOutputFrameTagGpu).Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Outputs().HasTag(kOutputFrameTag)) {
     cc->Outputs().Tag(kOutputFrameTag).Set<ImageFrame>();
   }
 
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -166,11 +173,11 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
 
   if (cc->Inputs().HasTag(kInputFrameTagGpu) &&
       cc->Outputs().HasTag(kOutputFrameTagGpu)) {
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     use_gpu_ = true;
 #else
-    RET_CHECK_FAIL() << "GPU processing on non-Android not supported yet.";
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif
   }
 
   sigma_color_ = options_.sigma_color();
@@ -180,9 +187,9 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
   if (!use_gpu_) sigma_color_ *= 255.0;
 
   if (use_gpu_) {
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
-#endif
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   return ::mediapipe::OkStatus();
@@ -190,7 +197,7 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
 
 ::mediapipe::Status BilateralFilterCalculator::Process(CalculatorContext* cc) {
   if (use_gpu_) {
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
           if (!gpu_initialized_) {
@@ -200,7 +207,7 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
           MP_RETURN_IF_ERROR(RenderGpu(cc));
           return ::mediapipe::OkStatus();
         }));
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     MP_RETURN_IF_ERROR(RenderCpu(cc));
   }
@@ -209,14 +216,14 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
 }
 
 ::mediapipe::Status BilateralFilterCalculator::Close(CalculatorContext* cc) {
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   gpu_helper_.RunInGlContext([this] {
     if (program_) glDeleteProgram(program_);
     program_ = 0;
     if (program_joint_) glDeleteProgram(program_joint_);
     program_joint_ = 0;
   });
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
@@ -263,7 +270,7 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
   if (cc->Inputs().Tag(kInputFrameTagGpu).IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   const auto& input_frame =
       cc->Inputs().Tag(kInputFrameTagGpu).Get<mediapipe::GpuBuffer>();
   auto input_texture = gpu_helper_.CreateSourceTexture(input_frame);
@@ -321,13 +328,13 @@ REGISTER_CALCULATOR(BilateralFilterCalculator);
   // Cleanup
   input_texture.Release();
   output_texture.Release();
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 void BilateralFilterCalculator::GlRender(CalculatorContext* cc) {
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   static const GLfloat square_vertices[] = {
       -1.0f, -1.0f,  // bottom left
       1.0f,  -1.0f,  // bottom right
@@ -373,11 +380,11 @@ void BilateralFilterCalculator::GlRender(CalculatorContext* cc) {
   glDeleteVertexArrays(1, &vao);
   glDeleteBuffers(2, vbo);
 
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 }
 
 ::mediapipe::Status BilateralFilterCalculator::GlSetup(CalculatorContext* cc) {
-#if defined(__ANDROID__) || defined(__EMSCRIPTEN__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   const GLint attr_location[NUM_ATTRIBUTES] = {
       ATTRIB_VERTEX,
       ATTRIB_TEXTURE_POSITION,
@@ -545,7 +552,7 @@ void BilateralFilterCalculator::GlRender(CalculatorContext* cc) {
   glUniform1i(glGetUniformLocation(program_joint_, "input_frame"), 1);
   glUniform1i(glGetUniformLocation(program_joint_, "guide_frame"), 2);
 
-#endif  // __ANDROID__ || __EMSCRIPTEN__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/image/image_cropping_calculator.cc

@@ -24,12 +24,12 @@
 #include "mediapipe/framework/port/ret_check.h"
 #include "mediapipe/framework/port/status.h"
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
 #include "mediapipe/gpu/gpu_buffer.h"
 #include "mediapipe/gpu/shader_util.h"
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace {
 enum { ATTRIB_VERTEX, ATTRIB_TEXTURE_POSITION, NUM_ATTRIBUTES };
@@ -37,9 +37,20 @@ enum { ATTRIB_VERTEX, ATTRIB_TEXTURE_POSITION, NUM_ATTRIBUTES };
 
 namespace mediapipe {
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+namespace {
 
-#endif  // __ANDROID__ or iOS
+#if !defined(MEDIAPIPE_DISABLE_GPU)
+
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+
+constexpr char kRectTag[] = "RECT";
+constexpr char kNormRectTag[] = "NORM_RECT";
+constexpr char kHeightTag[] = "HEIGHT";
+constexpr char kImageTag[] = "IMAGE";
+constexpr char kImageGpuTag[] = "IMAGE_GPU";
+constexpr char kWidthTag[] = "WIDTH";
+
+}  // namespace
 
 // Crops the input texture to the given rectangle region. The rectangle can
 // be at arbitrary location on the image with rotation. If there's rotation, the
@@ -91,48 +102,55 @@ class ImageCroppingCalculator : public CalculatorBase {
   bool use_gpu_ = false;
   // Output texture corners (4) after transoformation in normalized coordinates.
   float transformed_points_[8];
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   bool gpu_initialized_ = false;
   mediapipe::GlCalculatorHelper gpu_helper_;
   GLuint program_ = 0;
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 };
 REGISTER_CALCULATOR(ImageCroppingCalculator);
 
 ::mediapipe::Status ImageCroppingCalculator::GetContract(
     CalculatorContract* cc) {
-  RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
+  RET_CHECK(cc->Inputs().HasTag(kImageTag) ^ cc->Inputs().HasTag(kImageGpuTag));
-  RET_CHECK(cc->Outputs().HasTag("IMAGE") ^ cc->Outputs().HasTag("IMAGE_GPU"));
+  RET_CHECK(cc->Outputs().HasTag(kImageTag) ^
+            cc->Outputs().HasTag(kImageGpuTag));
 
-  if (cc->Inputs().HasTag("IMAGE")) {
+  bool use_gpu = false;
-    RET_CHECK(cc->Outputs().HasTag("IMAGE"));
-    cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
-    cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
-  }
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
-    RET_CHECK(cc->Outputs().HasTag("IMAGE_GPU"));
-    cc->Inputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
-    cc->Outputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
-  }
-#endif  // __ANDROID__ or iOS
 
-  if (cc->Inputs().HasTag("RECT")) {
+  if (cc->Inputs().HasTag(kImageTag)) {
-    cc->Inputs().Tag("RECT").Set<Rect>();
+    RET_CHECK(cc->Outputs().HasTag(kImageTag));
+    cc->Inputs().Tag(kImageTag).Set<ImageFrame>();
+    cc->Outputs().Tag(kImageTag).Set<ImageFrame>();
   }
-  if (cc->Inputs().HasTag("NORM_RECT")) {
+#if !defined(MEDIAPIPE_DISABLE_GPU)
-    cc->Inputs().Tag("NORM_RECT").Set<NormalizedRect>();
+  if (cc->Inputs().HasTag(kImageGpuTag)) {
+    RET_CHECK(cc->Outputs().HasTag(kImageGpuTag));
+    cc->Inputs().Tag(kImageGpuTag).Set<GpuBuffer>();
+    cc->Outputs().Tag(kImageGpuTag).Set<GpuBuffer>();
+    use_gpu |= true;
   }
-  if (cc->Inputs().HasTag("WIDTH")) {
+#endif  //  !MEDIAPIPE_DISABLE_GPU
-    cc->Inputs().Tag("WIDTH").Set<int>();
+
+  RET_CHECK(cc->Inputs().HasTag(kRectTag) ^ cc->Inputs().HasTag(kNormRectTag));
+  if (cc->Inputs().HasTag(kRectTag)) {
+    cc->Inputs().Tag(kRectTag).Set<Rect>();
   }
-  if (cc->Inputs().HasTag("HEIGHT")) {
+  if (cc->Inputs().HasTag(kNormRectTag)) {
-    cc->Inputs().Tag("HEIGHT").Set<int>();
+    cc->Inputs().Tag(kNormRectTag).Set<NormalizedRect>();
+  }
+  if (cc->Inputs().HasTag(kWidthTag)) {
+    cc->Inputs().Tag(kWidthTag).Set<int>();
+  }
+  if (cc->Inputs().HasTag(kHeightTag)) {
+    cc->Inputs().Tag(kHeightTag).Set<int>();
   }
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -140,26 +158,35 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
 ::mediapipe::Status ImageCroppingCalculator::Open(CalculatorContext* cc) {
   cc->SetOffset(TimestampDiff(0));
 
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
+  if (cc->Inputs().HasTag(kImageGpuTag)) {
     use_gpu_ = true;
   }
 
   options_ = cc->Options<mediapipe::ImageCroppingCalculatorOptions>();
 
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
 #else
     RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status ImageCroppingCalculator::Process(CalculatorContext* cc) {
+  if (cc->Inputs().HasTag(kRectTag) && cc->Inputs().Tag(kRectTag).IsEmpty()) {
+    VLOG(1) << "RECT is empty for timestamp: " << cc->InputTimestamp();
+    return ::mediapipe::OkStatus();
+  }
+  if (cc->Inputs().HasTag(kNormRectTag) &&
+      cc->Inputs().Tag(kNormRectTag).IsEmpty()) {
+    VLOG(1) << "NORM_RECT is empty for timestamp: " << cc->InputTimestamp();
+    return ::mediapipe::OkStatus();
+  }
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
           if (!gpu_initialized_) {
@@ -169,7 +196,7 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
           MP_RETURN_IF_ERROR(RenderGpu(cc));
           return ::mediapipe::OkStatus();
         }));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     MP_RETURN_IF_ERROR(RenderCpu(cc));
   }
@@ -177,19 +204,22 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
 }
 
 ::mediapipe::Status ImageCroppingCalculator::Close(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   gpu_helper_.RunInGlContext([this] {
     if (program_) glDeleteProgram(program_);
     program_ = 0;
   });
   gpu_initialized_ = false;
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status ImageCroppingCalculator::RenderCpu(CalculatorContext* cc) {
-  const auto& input_img = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
+  if (cc->Inputs().Tag(kImageTag).IsEmpty()) {
+    return ::mediapipe::OkStatus();
+  }
+  const auto& input_img = cc->Inputs().Tag(kImageTag).Get<ImageFrame>();
   cv::Mat input_mat = formats::MatView(&input_img);
 
   float rect_center_x = input_img.Width() / 2.0f;
@@ -197,8 +227,8 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
   float rotation = 0.0f;
   int target_width = input_img.Width();
   int target_height = input_img.Height();
-  if (cc->Inputs().HasTag("RECT")) {
+  if (cc->Inputs().HasTag(kRectTag)) {
-    const auto& rect = cc->Inputs().Tag("RECT").Get<Rect>();
+    const auto& rect = cc->Inputs().Tag(kRectTag).Get<Rect>();
     if (rect.width() > 0 && rect.height() > 0 && rect.x_center() >= 0 &&
         rect.y_center() >= 0) {
       rect_center_x = rect.x_center();
@@ -207,8 +237,8 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
       target_height = rect.height();
       rotation = rect.rotation();
     }
-  } else if (cc->Inputs().HasTag("NORM_RECT")) {
+  } else if (cc->Inputs().HasTag(kNormRectTag)) {
-    const auto& rect = cc->Inputs().Tag("NORM_RECT").Get<NormalizedRect>();
+    const auto& rect = cc->Inputs().Tag(kNormRectTag).Get<NormalizedRect>();
     if (rect.width() > 0.0 && rect.height() > 0.0 && rect.x_center() >= 0.0 &&
         rect.y_center() >= 0.0) {
       rect_center_x = std::round(rect.x_center() * input_img.Width());
@@ -218,9 +248,9 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
       rotation = rect.rotation();
     }
   } else {
-    if (cc->Inputs().HasTag("WIDTH") && cc->Inputs().HasTag("HEIGHT")) {
+    if (cc->Inputs().HasTag(kWidthTag) && cc->Inputs().HasTag(kHeightTag)) {
-      target_width = cc->Inputs().Tag("WIDTH").Get<int>();
+      target_width = cc->Inputs().Tag(kWidthTag).Get<int>();
-      target_height = cc->Inputs().Tag("HEIGHT").Get<int>();
+      target_height = cc->Inputs().Tag(kHeightTag).Get<int>();
     } else if (options_.has_width() && options_.has_height()) {
       target_width = options_.width();
       target_height = options_.height();
@@ -253,16 +283,17 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
       input_img.Format(), cropped_image.cols, cropped_image.rows));
   cv::Mat output_mat = formats::MatView(output_frame.get());
   cropped_image.copyTo(output_mat);
-  cc->Outputs().Tag("IMAGE").Add(output_frame.release(), cc->InputTimestamp());
+  cc->Outputs().Tag(kImageTag).Add(output_frame.release(),
+                                   cc->InputTimestamp());
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status ImageCroppingCalculator::RenderGpu(CalculatorContext* cc) {
-  if (cc->Inputs().Tag("IMAGE_GPU").IsEmpty()) {
+  if (cc->Inputs().Tag(kImageGpuTag).IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
-  const Packet& input_packet = cc->Inputs().Tag("IMAGE_GPU").Value();
+  const Packet& input_packet = cc->Inputs().Tag(kImageGpuTag).Value();
   const auto& input_buffer = input_packet.Get<mediapipe::GpuBuffer>();
   auto src_tex = gpu_helper_.CreateSourceTexture(input_buffer);
 
@@ -287,18 +318,18 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
 
   // Send result image in GPU packet.
   auto output = dst_tex.GetFrame<mediapipe::GpuBuffer>();
-  cc->Outputs().Tag("IMAGE_GPU").Add(output.release(), cc->InputTimestamp());
+  cc->Outputs().Tag(kImageGpuTag).Add(output.release(), cc->InputTimestamp());
 
   // Cleanup
   src_tex.Release();
   dst_tex.Release();
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 void ImageCroppingCalculator::GlRender() {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   static const GLfloat square_vertices[] = {
       -1.0f, -1.0f,  // bottom left
       1.0f,  -1.0f,  // bottom right
@@ -342,11 +373,11 @@ void ImageCroppingCalculator::GlRender() {
   glDeleteVertexArrays(1, &vao);
   glDeleteBuffers(2, vbo);
 
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 }
 
 ::mediapipe::Status ImageCroppingCalculator::InitGpu(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   const GLint attr_location[NUM_ATTRIBUTES] = {
       ATTRIB_VERTEX,
       ATTRIB_TEXTURE_POSITION,
@@ -392,7 +423,7 @@ void ImageCroppingCalculator::GlRender() {
   // Parameters
   glUseProgram(program_);
   glUniform1i(glGetUniformLocation(program_, "input_frame"), 1);
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
@@ -410,8 +441,8 @@ void ImageCroppingCalculator::GetOutputDimensions(CalculatorContext* cc,
   int y_center = src_height / 2;
   // Get the rotation of the cropping box.
   float rotation = 0.0f;
-  if (cc->Inputs().HasTag("RECT")) {
+  if (cc->Inputs().HasTag(kRectTag)) {
-    const auto& rect = cc->Inputs().Tag("RECT").Get<Rect>();
+    const auto& rect = cc->Inputs().Tag(kRectTag).Get<Rect>();
     // Only use the rect if it is valid.
     if (rect.width() > 0 && rect.height() > 0 && rect.x_center() >= 0 &&
         rect.y_center() >= 0) {
@@ -421,8 +452,8 @@ void ImageCroppingCalculator::GetOutputDimensions(CalculatorContext* cc,
       crop_height = rect.height();
       rotation = rect.rotation();
     }
-  } else if (cc->Inputs().HasTag("NORM_RECT")) {
+  } else if (cc->Inputs().HasTag(kNormRectTag)) {
-    const auto& rect = cc->Inputs().Tag("NORM_RECT").Get<NormalizedRect>();
+    const auto& rect = cc->Inputs().Tag(kNormRectTag).Get<NormalizedRect>();
     // Only use the rect if it is valid.
     if (rect.width() > 0.0 && rect.height() > 0.0 && rect.x_center() >= 0.0 &&
         rect.y_center() >= 0.0) {
@@ -433,9 +464,9 @@ void ImageCroppingCalculator::GetOutputDimensions(CalculatorContext* cc,
       rotation = rect.rotation();
     }
   } else {
-    if (cc->Inputs().HasTag("WIDTH") && cc->Inputs().HasTag("HEIGHT")) {
+    if (cc->Inputs().HasTag(kWidthTag) && cc->Inputs().HasTag(kHeightTag)) {
-      crop_width = cc->Inputs().Tag("WIDTH").Get<int>();
+      crop_width = cc->Inputs().Tag(kWidthTag).Get<int>();
-      crop_height = cc->Inputs().Tag("HEIGHT").Get<int>();
+      crop_height = cc->Inputs().Tag(kHeightTag).Get<int>();
     } else if (options_.has_width() && options_.has_height()) {
       crop_width = options_.width();
       crop_height = options_.height();

mediapipe/calculators/image/image_properties_calculator.cc

@@ -15,9 +15,9 @@
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame.h"
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 #include "mediapipe/gpu/gpu_buffer.h"
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace mediapipe {
 
@@ -44,11 +44,11 @@ class ImagePropertiesCalculator : public CalculatorBase {
     if (cc->Inputs().HasTag("IMAGE")) {
       cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
     }
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     if (cc->Inputs().HasTag("IMAGE_GPU")) {
       cc->Inputs().Tag("IMAGE_GPU").Set<::mediapipe::GpuBuffer>();
     }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
     if (cc->Outputs().HasTag("SIZE")) {
       cc->Outputs().Tag("SIZE").Set<std::pair<int, int>>();
@@ -71,7 +71,7 @@ class ImagePropertiesCalculator : public CalculatorBase {
       width = image.Width();
       height = image.Height();
     }
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     if (cc->Inputs().HasTag("IMAGE_GPU") &&
         !cc->Inputs().Tag("IMAGE_GPU").IsEmpty()) {
       const auto& image =
@@ -79,7 +79,7 @@ class ImagePropertiesCalculator : public CalculatorBase {
       width = image.width();
       height = image.height();
     }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
     cc->Outputs().Tag("SIZE").AddPacket(
         MakePacket<std::pair<int, int>>(width, height)

mediapipe/calculators/image/image_transformation_calculator.cc

@@ -22,12 +22,12 @@
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/gpu/scale_mode.pb.h"
 
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_quad_renderer.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
 #include "mediapipe/gpu/shader_util.h"
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 #if defined(__ANDROID__)
 // The size of Java arrays is dynamic, which makes it difficult to
@@ -42,9 +42,9 @@ typedef int DimensionsPacketType[2];
 
 namespace mediapipe {
 
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace {
 int RotationModeToDegrees(mediapipe::RotationMode_Mode rotation) {
@@ -170,12 +170,12 @@ class ImageTransformationCalculator : public CalculatorBase {
   mediapipe::ScaleMode_Mode scale_mode_;
 
   bool use_gpu_ = false;
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   GlCalculatorHelper helper_;
   std::unique_ptr<QuadRenderer> rgb_renderer_;
   std::unique_ptr<QuadRenderer> yuv_renderer_;
   std::unique_ptr<QuadRenderer> ext_rgb_renderer_;
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 };
 REGISTER_CALCULATOR(ImageTransformationCalculator);
 
@@ -185,18 +185,22 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
   RET_CHECK(cc->Outputs().HasTag("IMAGE") ^ cc->Outputs().HasTag("IMAGE_GPU"));
 
+  bool use_gpu = false;
+
   if (cc->Inputs().HasTag("IMAGE")) {
     RET_CHECK(cc->Outputs().HasTag("IMAGE"));
     cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
     cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
   }
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag("IMAGE_GPU")) {
     RET_CHECK(cc->Outputs().HasTag("IMAGE_GPU"));
     cc->Inputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
     cc->Outputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+
   if (cc->Inputs().HasTag("ROTATION_DEGREES")) {
     cc->Inputs().Tag("ROTATION_DEGREES").Set<int>();
   }
@@ -212,9 +216,11 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
     cc->Outputs().Tag("LETTERBOX_PADDING").Set<std::array<float, 4>>();
   }
 
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(GlCalculatorHelper::UpdateContract(cc));
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -250,12 +256,12 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   scale_mode_ = ParseScaleMode(options_.scale_mode(), DEFAULT_SCALE_MODE);
 
   if (use_gpu_) {
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     // Let the helper access the GL context information.
     MP_RETURN_IF_ERROR(helper_.Open(cc));
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   return ::mediapipe::OkStatus();
@@ -264,10 +270,10 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 ::mediapipe::Status ImageTransformationCalculator::Process(
     CalculatorContext* cc) {
   if (use_gpu_) {
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     return helper_.RunInGlContext(
         [this, cc]() -> ::mediapipe::Status { return RenderGpu(cc); });
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     return RenderCpu(cc);
   }
@@ -277,7 +283,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 ::mediapipe::Status ImageTransformationCalculator::Close(
     CalculatorContext* cc) {
   if (use_gpu_) {
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     QuadRenderer* rgb_renderer = rgb_renderer_.release();
     QuadRenderer* yuv_renderer = yuv_renderer_.release();
     QuadRenderer* ext_rgb_renderer = ext_rgb_renderer_.release();
@@ -295,8 +301,9 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
         delete yuv_renderer;
       }
     });
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
+
   return ::mediapipe::OkStatus();
 }
 
@@ -371,7 +378,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 
 ::mediapipe::Status ImageTransformationCalculator::RenderGpu(
     CalculatorContext* cc) {
-#if defined(__ANDROID__) || defined(__APPLE__) && !TARGET_OS_OSX
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   int input_width = cc->Inputs().Tag("IMAGE_GPU").Get<GpuBuffer>().width();
   int input_height = cc->Inputs().Tag("IMAGE_GPU").Get<GpuBuffer>().height();
 
@@ -408,7 +415,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 #endif    // iOS
   {
     src1 = helper_.CreateSourceTexture(input);
-#if defined(__ANDROID__)
+#if defined(TEXTURE_EXTERNAL_OES)
     if (src1.target() == GL_TEXTURE_EXTERNAL_OES) {
       if (!ext_rgb_renderer_) {
         ext_rgb_renderer_ = absl::make_unique<QuadRenderer>();
@@ -417,7 +424,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
       }
       renderer = ext_rgb_renderer_.get();
     } else  // NOLINT(readability/braces)
-#endif      // __ANDROID__
+#endif      // TEXTURE_EXTERNAL_OES
     {
       if (!rgb_renderer_) {
         rgb_renderer_ = absl::make_unique<QuadRenderer>();
@@ -460,7 +467,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   auto output = dst.GetFrame<GpuBuffer>();
   cc->Outputs().Tag("IMAGE_GPU").Add(output.release(), cc->InputTimestamp());
 
-#endif  // __ANDROID__ || iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/image/recolor_calculator.cc

@@ -21,12 +21,11 @@
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/util/color.pb.h"
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
-#include "mediapipe/gpu/gpu_buffer.h"
 #include "mediapipe/gpu/shader_util.h"
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace {
 enum { ATTRIB_VERTEX, ATTRIB_TEXTURE_POSITION, NUM_ATTRIBUTES };
@@ -95,10 +94,10 @@ class RecolorCalculator : public CalculatorBase {
   mediapipe::RecolorCalculatorOptions::MaskChannel mask_channel_;
 
   bool use_gpu_ = false;
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   mediapipe::GlCalculatorHelper gpu_helper_;
   GLuint program_ = 0;
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 };
 REGISTER_CALCULATOR(RecolorCalculator);
 
@@ -107,36 +106,43 @@ REGISTER_CALCULATOR(RecolorCalculator);
   RET_CHECK(!cc->Inputs().GetTags().empty());
   RET_CHECK(!cc->Outputs().GetTags().empty());
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+  bool use_gpu = false;
+
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag("IMAGE_GPU")) {
     cc->Inputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag("IMAGE")) {
     cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
   }
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag("MASK_GPU")) {
     cc->Inputs().Tag("MASK_GPU").Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag("MASK")) {
     cc->Inputs().Tag("MASK").Set<ImageFrame>();
   }
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Outputs().HasTag("IMAGE_GPU")) {
     cc->Outputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Outputs().HasTag("IMAGE")) {
     cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
   }
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -146,9 +152,9 @@ REGISTER_CALCULATOR(RecolorCalculator);
 
   if (cc->Inputs().HasTag("IMAGE_GPU")) {
     use_gpu_ = true;
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   MP_RETURN_IF_ERROR(LoadOptions(cc));
@@ -158,7 +164,7 @@ REGISTER_CALCULATOR(RecolorCalculator);
 
 ::mediapipe::Status RecolorCalculator::Process(CalculatorContext* cc) {
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, &cc]() -> ::mediapipe::Status {
           if (!initialized_) {
@@ -168,7 +174,7 @@ REGISTER_CALCULATOR(RecolorCalculator);
           MP_RETURN_IF_ERROR(RenderGpu(cc));
           return ::mediapipe::OkStatus();
         }));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     MP_RETURN_IF_ERROR(RenderCpu(cc));
   }
@@ -176,12 +182,12 @@ REGISTER_CALCULATOR(RecolorCalculator);
 }
 
 ::mediapipe::Status RecolorCalculator::Close(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   gpu_helper_.RunInGlContext([this] {
     if (program_) glDeleteProgram(program_);
     program_ = 0;
   });
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
@@ -194,7 +200,7 @@ REGISTER_CALCULATOR(RecolorCalculator);
   if (cc->Inputs().Tag("MASK_GPU").IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   // Get inputs and setup output.
   const Packet& input_packet = cc->Inputs().Tag("IMAGE_GPU").Value();
   const Packet& mask_packet = cc->Inputs().Tag("MASK_GPU").Value();
@@ -233,13 +239,13 @@ REGISTER_CALCULATOR(RecolorCalculator);
   img_tex.Release();
   mask_tex.Release();
   dst_tex.Release();
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 void RecolorCalculator::GlRender() {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   static const GLfloat square_vertices[] = {
       -1.0f, -1.0f,  // bottom left
       1.0f,  -1.0f,  // bottom right
@@ -287,7 +293,7 @@ void RecolorCalculator::GlRender() {
   glBindVertexArray(0);
   glDeleteVertexArrays(1, &vao);
   glDeleteBuffers(2, vbo);
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 }
 
 ::mediapipe::Status RecolorCalculator::LoadOptions(CalculatorContext* cc) {
@@ -305,7 +311,7 @@ void RecolorCalculator::GlRender() {
 }
 
 ::mediapipe::Status RecolorCalculator::InitGpu(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   const GLint attr_location[NUM_ATTRIBUTES] = {
       ATTRIB_VERTEX,
       ATTRIB_TEXTURE_POSITION,
@@ -374,7 +380,7 @@ void RecolorCalculator::GlRender() {
   glUniform1i(glGetUniformLocation(program_, "mask"), 2);
   glUniform3f(glGetUniformLocation(program_, "recolor"), color_[0], color_[1],
               color_[2]);
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/image/set_alpha_calculator.cc

@@ -25,12 +25,11 @@
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/framework/port/vector.h"
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
-#include "mediapipe/gpu/gpu_buffer.h"
 #include "mediapipe/gpu/shader_util.h"
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace mediapipe {
 
@@ -107,16 +106,18 @@ class SetAlphaCalculator : public CalculatorBase {
 
   bool use_gpu_ = false;
   bool gpu_initialized_ = false;
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   mediapipe::GlCalculatorHelper gpu_helper_;
   GLuint program_ = 0;
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 };
 REGISTER_CALCULATOR(SetAlphaCalculator);
 
 ::mediapipe::Status SetAlphaCalculator::GetContract(CalculatorContract* cc) {
   CHECK_GE(cc->Inputs().NumEntries(), 1);
 
+  bool use_gpu = false;
+
   if (cc->Inputs().HasTag(kInputFrameTag) &&
       cc->Inputs().HasTag(kInputFrameTagGpu)) {
     return ::mediapipe::InternalError("Cannot have multiple input images.");
@@ -127,38 +128,43 @@ REGISTER_CALCULATOR(SetAlphaCalculator);
   }
 
   // Input image to add/edit alpha channel.
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag(kInputFrameTagGpu)) {
     cc->Inputs().Tag(kInputFrameTagGpu).Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag(kInputFrameTag)) {
     cc->Inputs().Tag(kInputFrameTag).Set<ImageFrame>();
   }
 
   // Input alpha image mask (optional)
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag(kInputAlphaTagGpu)) {
     cc->Inputs().Tag(kInputAlphaTagGpu).Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag(kInputAlphaTag)) {
     cc->Inputs().Tag(kInputAlphaTag).Set<ImageFrame>();
   }
 
   // RGBA output image.
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Outputs().HasTag(kOutputFrameTagGpu)) {
     cc->Outputs().Tag(kOutputFrameTagGpu).Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Outputs().HasTag(kOutputFrameTag)) {
     cc->Outputs().Tag(kOutputFrameTag).Set<ImageFrame>();
   }
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -170,11 +176,11 @@ REGISTER_CALCULATOR(SetAlphaCalculator);
 
   if (cc->Inputs().HasTag(kInputFrameTagGpu) &&
       cc->Outputs().HasTag(kOutputFrameTagGpu)) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     use_gpu_ = true;
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   // Get global value from options (-1 if not set).
@@ -187,17 +193,17 @@ REGISTER_CALCULATOR(SetAlphaCalculator);
     RET_CHECK_FAIL() << "Must use either image mask or options alpha value.";
 
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
 #endif
-  }
+  }  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status SetAlphaCalculator::Process(CalculatorContext* cc) {
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
           if (!gpu_initialized_) {
@@ -207,7 +213,7 @@ REGISTER_CALCULATOR(SetAlphaCalculator);
           MP_RETURN_IF_ERROR(RenderGpu(cc));
           return ::mediapipe::OkStatus();
         }));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     MP_RETURN_IF_ERROR(RenderCpu(cc));
   }
@@ -216,12 +222,12 @@ REGISTER_CALCULATOR(SetAlphaCalculator);
 }
 
 ::mediapipe::Status SetAlphaCalculator::Close(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   gpu_helper_.RunInGlContext([this] {
     if (program_) glDeleteProgram(program_);
     program_ = 0;
   });
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
@@ -295,7 +301,7 @@ REGISTER_CALCULATOR(SetAlphaCalculator);
   if (cc->Inputs().Tag(kInputFrameTagGpu).IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   // Setup source texture.
   const auto& input_frame =
       cc->Inputs().Tag(kInputFrameTagGpu).Get<mediapipe::GpuBuffer>();
@@ -348,13 +354,13 @@ REGISTER_CALCULATOR(SetAlphaCalculator);
   // Cleanup
   input_texture.Release();
   output_texture.Release();
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 void SetAlphaCalculator::GlRender(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   static const GLfloat square_vertices[] = {
       -1.0f, -1.0f,  // bottom left
       1.0f,  -1.0f,  // bottom right
@@ -403,11 +409,11 @@ void SetAlphaCalculator::GlRender(CalculatorContext* cc) {
   glDeleteVertexArrays(1, &vao);
   glDeleteBuffers(2, vbo);
 
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 }
 
 ::mediapipe::Status SetAlphaCalculator::GlSetup(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   const GLint attr_location[NUM_ATTRIBUTES] = {
       ATTRIB_VERTEX,
       ATTRIB_TEXTURE_POSITION,
@@ -460,7 +466,7 @@ void SetAlphaCalculator::GlRender(CalculatorContext* cc) {
   glUniform1i(glGetUniformLocation(program_, "alpha_mask"), 2);
   glUniform1f(glGetUniformLocation(program_, "alpha_value"), alpha_value_);
 
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/tensorflow/BUILD

@@ -255,6 +255,7 @@ mediapipe_cc_proto_library(
     cc_deps = [
         "//mediapipe/calculators/core:packet_resampler_calculator_cc_proto",
         "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util:audio_decoder_cc_proto",
     ],
     visibility = ["//visibility:public"],
     deps = [":unpack_media_sequence_calculator_proto"],
@@ -653,6 +654,7 @@ cc_library(
         "//mediapipe/framework/formats:location",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
+        "//mediapipe/util:audio_decoder_cc_proto",
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/strings",
         "@org_tensorflow//tensorflow/core:protos_all_cc",
@@ -769,7 +771,6 @@ cc_test(
         "//mediapipe/framework/formats:location",
         "//mediapipe/framework/port:gtest_main",
         "//mediapipe/framework/port:opencv_imgcodecs",
-        "//mediapipe/framework/port:status",
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",
@@ -971,6 +972,7 @@ cc_test(
         "//mediapipe/framework/formats:location",
         "//mediapipe/framework/port:gtest_main",
         "//mediapipe/framework/port:rectangle",
+        "//mediapipe/util:audio_decoder_cc_proto",
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",

mediapipe/calculators/tensorflow/pack_media_sequence_calculator.cc

@@ -285,6 +285,10 @@ class PackMediaSequenceCalculator : public CalculatorBase {
   }
 
   ::mediapipe::Status Process(CalculatorContext* cc) override {
+    int image_height = -1;
+    int image_width = -1;
+    // Because the tag order may vary, we need to loop through tags to get
+    // image information before processing other tag types.
     for (const auto& tag : cc->Inputs().GetTags()) {
       if (!cc->Inputs().Tag(tag).IsEmpty()) {
         features_present_[tag] = true;
@@ -306,14 +310,21 @@ class PackMediaSequenceCalculator : public CalculatorBase {
           return ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
                  << "No encoded image";
         }
+        image_height = image.height();
+        image_width = image.width();
         mpms::AddImageTimestamp(key, cc->InputTimestamp().Value(),
                                 sequence_.get());
         mpms::AddImageEncoded(key, image.encoded_image(), sequence_.get());
       }
+    }
+    for (const auto& tag : cc->Inputs().GetTags()) {
+      if (!cc->Inputs().Tag(tag).IsEmpty()) {
+        features_present_[tag] = true;
+      }
       if (absl::StartsWith(tag, kKeypointsTag) &&
           !cc->Inputs().Tag(tag).IsEmpty()) {
         std::string key = "";
-        if (tag != kImageTag) {
+        if (tag != kKeypointsTag) {
           int tag_length = sizeof(kKeypointsTag) / sizeof(*kKeypointsTag) - 1;
           if (tag[tag_length] == '_') {
             key = tag.substr(tag_length + 1);
@@ -363,11 +374,20 @@ class PackMediaSequenceCalculator : public CalculatorBase {
                   LocationData::BOUNDING_BOX ||
               detection.location_data().format() ==
                   LocationData::RELATIVE_BOUNDING_BOX) {
-            int height = mpms::GetImageHeight(*sequence_);
+            if (mpms::HasImageHeight(*sequence_) &&
-            int width = mpms::GetImageWidth(*sequence_);
+                mpms::HasImageWidth(*sequence_)) {
+              image_height = mpms::GetImageHeight(*sequence_);
+              image_width = mpms::GetImageWidth(*sequence_);
+            }
+            if (image_height == -1 || image_width == -1) {
+              return ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
+                     << "Images must be provided with bounding boxes or the "
+                        "image "
+                     << "height and width must already be in the example.";
+            }
             Location relative_bbox = Location::CreateRelativeBBoxLocation(
                 Location(detection.location_data())
-                    .ConvertToRelativeBBox(width, height));
+                    .ConvertToRelativeBBox(image_width, image_height));
             predicted_locations.push_back(relative_bbox);
             if (detection.label_size() > 0) {
               predicted_class_strings.push_back(detection.label(0));

mediapipe/calculators/tensorflow/pack_media_sequence_calculator_test.cc

@@ -357,6 +357,148 @@ TEST_F(PackMediaSequenceCalculatorTest, PacksTwoBBoxDetections) {
   }
 }
 
+TEST_F(PackMediaSequenceCalculatorTest, PacksBBoxWithoutImageDims) {
+  SetUpCalculator({"BBOX_PREDICTED:detections"}, {}, false, true);
+  auto input_sequence = ::absl::make_unique<tf::SequenceExample>();
+  std::string test_video_id = "test_video_id";
+  mpms::SetClipMediaId(test_video_id, input_sequence.get());
+  int height = 480;
+  int width = 640;
+  int num_vectors = 2;
+  for (int i = 0; i < num_vectors; ++i) {
+    auto detections = ::absl::make_unique<::std::vector<Detection>>();
+    Detection detection;
+    detection.add_label("absolute bbox");
+    detection.add_label_id(0);
+    detection.add_score(0.5);
+    Location::CreateBBoxLocation(0, height / 2, width / 2, height / 2)
+        .ConvertToProto(detection.mutable_location_data());
+    detections->push_back(detection);
+
+    detection = Detection();
+    detection.add_label("relative bbox");
+    detection.add_label_id(1);
+    detection.add_score(0.75);
+    Location::CreateRelativeBBoxLocation(0, 0.5, 0.5, 0.5)
+        .ConvertToProto(detection.mutable_location_data());
+    detections->push_back(detection);
+
+    // The mask detection should be ignored in the output.
+    detection = Detection();
+    detection.add_label("mask");
+    detection.add_score(1.0);
+    cv::Mat image(2, 3, CV_8UC1, cv::Scalar(0));
+    Location::CreateCvMaskLocation<uint8>(image).ConvertToProto(
+        detection.mutable_location_data());
+    detections->push_back(detection);
+
+    runner_->MutableInputs()
+        ->Tag("BBOX_PREDICTED")
+        .packets.push_back(Adopt(detections.release()).At(Timestamp(i)));
+  }
+
+  runner_->MutableSidePackets()->Tag("SEQUENCE_EXAMPLE") =
+      Adopt(input_sequence.release());
+
+  auto status = runner_->Run();
+  EXPECT_EQ(::mediapipe::StatusCode::kInvalidArgument, status.code());
+}
+
+TEST_F(PackMediaSequenceCalculatorTest, PacksBBoxWithImages) {
+  SetUpCalculator({"BBOX_PREDICTED:detections", "IMAGE:images"}, {}, false,
+                  true);
+  auto input_sequence = ::absl::make_unique<tf::SequenceExample>();
+  std::string test_video_id = "test_video_id";
+  mpms::SetClipMediaId(test_video_id, input_sequence.get());
+  int height = 480;
+  int width = 640;
+  int num_vectors = 2;
+  for (int i = 0; i < num_vectors; ++i) {
+    auto detections = ::absl::make_unique<::std::vector<Detection>>();
+    Detection detection;
+    detection.add_label("absolute bbox");
+    detection.add_label_id(0);
+    detection.add_score(0.5);
+    Location::CreateBBoxLocation(0, height / 2, width / 2, height / 2)
+        .ConvertToProto(detection.mutable_location_data());
+    detections->push_back(detection);
+
+    detection = Detection();
+    detection.add_label("relative bbox");
+    detection.add_label_id(1);
+    detection.add_score(0.75);
+    Location::CreateRelativeBBoxLocation(0, 0.5, 0.5, 0.5)
+        .ConvertToProto(detection.mutable_location_data());
+    detections->push_back(detection);
+
+    // The mask detection should be ignored in the output.
+    detection = Detection();
+    detection.add_label("mask");
+    detection.add_score(1.0);
+    cv::Mat image(2, 3, CV_8UC1, cv::Scalar(0));
+    Location::CreateCvMaskLocation<uint8>(image).ConvertToProto(
+        detection.mutable_location_data());
+    detections->push_back(detection);
+
+    runner_->MutableInputs()
+        ->Tag("BBOX_PREDICTED")
+        .packets.push_back(Adopt(detections.release()).At(Timestamp(i)));
+  }
+  cv::Mat image(height, width, CV_8UC3, cv::Scalar(0, 0, 255));
+  std::vector<uchar> bytes;
+  ASSERT_TRUE(cv::imencode(".jpg", image, bytes, {80}));
+  std::string test_image_string(bytes.begin(), bytes.end());
+  OpenCvImageEncoderCalculatorResults encoded_image;
+  encoded_image.set_encoded_image(test_image_string);
+  encoded_image.set_width(width);
+  encoded_image.set_height(height);
+
+  int num_images = 2;
+  for (int i = 0; i < num_images; ++i) {
+    auto image_ptr =
+        ::absl::make_unique<OpenCvImageEncoderCalculatorResults>(encoded_image);
+    runner_->MutableInputs()->Tag("IMAGE").packets.push_back(
+        Adopt(image_ptr.release()).At(Timestamp(i)));
+  }
+  runner_->MutableSidePackets()->Tag("SEQUENCE_EXAMPLE") =
+      Adopt(input_sequence.release());
+
+  MP_ASSERT_OK(runner_->Run());
+
+  const std::vector<Packet>& output_packets =
+      runner_->Outputs().Tag("SEQUENCE_EXAMPLE").packets;
+  ASSERT_EQ(1, output_packets.size());
+  const tf::SequenceExample& output_sequence =
+      output_packets[0].Get<tf::SequenceExample>();
+
+  ASSERT_EQ(test_video_id, mpms::GetClipMediaId(output_sequence));
+  ASSERT_EQ(height, mpms::GetImageHeight(output_sequence));
+  ASSERT_EQ(width, mpms::GetImageWidth(output_sequence));
+  ASSERT_EQ(num_vectors, mpms::GetPredictedBBoxSize(output_sequence));
+  ASSERT_EQ(num_vectors, mpms::GetPredictedBBoxTimestampSize(output_sequence));
+  ASSERT_EQ(0, mpms::GetClassSegmentationEncodedSize(output_sequence));
+  ASSERT_EQ(0, mpms::GetClassSegmentationTimestampSize(output_sequence));
+  for (int i = 0; i < num_vectors; ++i) {
+    ASSERT_EQ(i, mpms::GetPredictedBBoxTimestampAt(output_sequence, i));
+    auto bboxes = mpms::GetPredictedBBoxAt(output_sequence, i);
+    ASSERT_EQ(2, bboxes.size());
+    for (int j = 0; j < bboxes.size(); ++j) {
+      auto rect = bboxes[j].GetRelativeBBox();
+      ASSERT_NEAR(0, rect.xmin(), 0.001);
+      ASSERT_NEAR(0.5, rect.ymin(), 0.001);
+      ASSERT_NEAR(0.5, rect.xmax(), 0.001);
+      ASSERT_NEAR(1.0, rect.ymax(), 0.001);
+    }
+    auto class_strings =
+        mpms::GetPredictedBBoxLabelStringAt(output_sequence, i);
+    ASSERT_EQ("absolute bbox", class_strings[0]);
+    ASSERT_EQ("relative bbox", class_strings[1]);
+    auto class_indices = mpms::GetPredictedBBoxLabelIndexAt(output_sequence, i);
+    ASSERT_EQ(0, class_indices[0]);
+    ASSERT_EQ(1, class_indices[1]);
+  }
+}
+
 TEST_F(PackMediaSequenceCalculatorTest, PacksTwoKeypoints) {
   SetUpCalculator({"KEYPOINTS_TEST:keypoints"}, {}, false, true);
   auto input_sequence = ::absl::make_unique<tf::SequenceExample>();

mediapipe/calculators/tensorflow/unpack_media_sequence_calculator.cc

@@ -19,6 +19,7 @@
 #include "mediapipe/framework/formats/location.h"
 #include "mediapipe/framework/port/ret_check.h"
 #include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/audio_decoder.pb.h"
 #include "mediapipe/util/sequence/media_sequence.h"
 #include "tensorflow/core/example/example.pb.h"
 #include "tensorflow/core/example/feature.pb.h"
@@ -37,6 +38,7 @@ const char kDatasetRootDirTag[] = "DATASET_ROOT";
 const char kDataPath[] = "DATA_PATH";
 const char kPacketResamplerOptions[] = "RESAMPLER_OPTIONS";
 const char kImagesFrameRateTag[] = "IMAGE_FRAME_RATE";
+const char kAudioDecoderOptions[] = "AUDIO_DECODER_OPTIONS";
 
 namespace tf = ::tensorflow;
 namespace mpms = ::mediapipe::mediasequence;
@@ -126,6 +128,11 @@ class UnpackMediaSequenceCalculator : public CalculatorBase {
     if (cc->OutputSidePackets().HasTag(kDataPath)) {
       cc->OutputSidePackets().Tag(kDataPath).Set<std::string>();
     }
+    if (cc->OutputSidePackets().HasTag(kAudioDecoderOptions)) {
+      cc->OutputSidePackets()
+          .Tag(kAudioDecoderOptions)
+          .Set<AudioDecoderOptions>();
+    }
     if (cc->OutputSidePackets().HasTag(kImagesFrameRateTag)) {
       cc->OutputSidePackets().Tag(kImagesFrameRateTag).Set<double>();
     }
@@ -136,10 +143,11 @@ class UnpackMediaSequenceCalculator : public CalculatorBase {
     }
     if ((options.has_padding_before_label() ||
          options.has_padding_after_label()) &&
-        !(cc->OutputSidePackets().HasTag(kPacketResamplerOptions))) {
+        !(cc->OutputSidePackets().HasTag(kAudioDecoderOptions) ||
+          cc->OutputSidePackets().HasTag(kPacketResamplerOptions))) {
       return ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
-             << "If specifying padding, must output "
+             << "If specifying padding, must output " << kPacketResamplerOptions
-             << kPacketResamplerOptions;
+             << "or" << kAudioDecoderOptions;
     }
 
     // Optional streams.
@@ -260,7 +268,8 @@ class UnpackMediaSequenceCalculator : public CalculatorBase {
     // Set the start and end of the clip in the appropriate options protos.
     double start_time = 0;
     double end_time = 0;
-    if (cc->OutputSidePackets().HasTag(kPacketResamplerOptions)) {
+    if (cc->OutputSidePackets().HasTag(kAudioDecoderOptions) ||
+        cc->OutputSidePackets().HasTag(kPacketResamplerOptions)) {
       if (mpms::HasClipStartTimestamp(sequence)) {
         start_time =
             Timestamp(mpms::GetClipStartTimestamp(sequence)).Seconds() -
@@ -271,6 +280,27 @@ class UnpackMediaSequenceCalculator : public CalculatorBase {
                    options.padding_after_label();
       }
     }
+    if (cc->OutputSidePackets().HasTag(kAudioDecoderOptions)) {
+      auto audio_decoder_options = absl::make_unique<AudioDecoderOptions>(
+          options.base_audio_decoder_options());
+      if (mpms::HasClipStartTimestamp(sequence)) {
+        if (options.force_decoding_from_start_of_media()) {
+          audio_decoder_options->set_start_time(0);
+        } else {
+          audio_decoder_options->set_start_time(
+              start_time - options.extra_padding_from_media_decoder());
+        }
+      }
+      if (mpms::HasClipEndTimestamp(sequence)) {
+        audio_decoder_options->set_end_time(
+            end_time + options.extra_padding_from_media_decoder());
+      }
+      LOG(INFO) << "Created AudioDecoderOptions:\n"
+                << audio_decoder_options->DebugString();
+      cc->OutputSidePackets()
+          .Tag(kAudioDecoderOptions)
+          .Set(Adopt(audio_decoder_options.release()));
+    }
     if (cc->OutputSidePackets().HasTag(kPacketResamplerOptions)) {
       auto resampler_options = absl::make_unique<CalculatorOptions>();
       *(resampler_options->MutableExtension(

mediapipe/calculators/tensorflow/unpack_media_sequence_calculator.proto

@@ -18,6 +18,7 @@ package mediapipe;
 
 import "mediapipe/calculators/core/packet_resampler_calculator.proto";
 import "mediapipe/framework/calculator.proto";
+import "mediapipe/util/audio_decoder.proto";
 
 message UnpackMediaSequenceCalculatorOptions {
   extend mediapipe.CalculatorOptions {
@@ -49,4 +50,10 @@ message UnpackMediaSequenceCalculatorOptions {
   // parameters for the MediaDecoderCalculator. End time parameters are still
   // respected.
   optional bool force_decoding_from_start_of_media = 7;
+
+  // Stores the audio decoder settings for the graph. (e.g. which audio
+  // stream to pull from the video.) The sequence's metadata overrides
+  // the clip start and end times and outputs these for the
+  // AudioDecoderCalculator to consume.
+  optional AudioDecoderOptions base_audio_decoder_options = 9;
 }

mediapipe/calculators/tensorflow/unpack_media_sequence_calculator_test.cc

@@ -23,6 +23,7 @@
 #include "mediapipe/framework/port/gtest.h"
 #include "mediapipe/framework/port/rectangle.h"
 #include "mediapipe/framework/port/status_matchers.h"
+#include "mediapipe/util/audio_decoder.pb.h"
 #include "mediapipe/util/sequence/media_sequence.h"
 #include "tensorflow/core/example/example.pb.h"
 
@@ -459,6 +460,62 @@ TEST_F(UnpackMediaSequenceCalculatorTest, GetDatasetFromExample) {
             data_path_);
 }
 
+TEST_F(UnpackMediaSequenceCalculatorTest, GetAudioDecoderOptions) {
+  CalculatorOptions options;
+  options.MutableExtension(UnpackMediaSequenceCalculatorOptions::ext)
+      ->set_padding_before_label(1);
+  options.MutableExtension(UnpackMediaSequenceCalculatorOptions::ext)
+      ->set_padding_after_label(2);
+  SetUpCalculator({}, {"AUDIO_DECODER_OPTIONS:audio_decoder_options"}, {},
+                  &options);
+  runner_->MutableSidePackets()->Tag("SEQUENCE_EXAMPLE") =
+      Adopt(sequence_.release());
+  MP_ASSERT_OK(runner_->Run());
+
+  MP_EXPECT_OK(runner_->OutputSidePackets()
+                   .Tag("AUDIO_DECODER_OPTIONS")
+                   .ValidateAsType<AudioDecoderOptions>());
+  EXPECT_NEAR(runner_->OutputSidePackets()
+                  .Tag("AUDIO_DECODER_OPTIONS")
+                  .Get<AudioDecoderOptions>()
+                  .start_time(),
+              2.0, 1e-5);
+  EXPECT_NEAR(runner_->OutputSidePackets()
+                  .Tag("AUDIO_DECODER_OPTIONS")
+                  .Get<AudioDecoderOptions>()
+                  .end_time(),
+              7.0, 1e-5);
+}
+
+TEST_F(UnpackMediaSequenceCalculatorTest, GetAudioDecoderOptionsOverride) {
+  CalculatorOptions options;
+  options.MutableExtension(UnpackMediaSequenceCalculatorOptions::ext)
+      ->set_padding_before_label(1);
+  options.MutableExtension(UnpackMediaSequenceCalculatorOptions::ext)
+      ->set_padding_after_label(2);
+  options.MutableExtension(UnpackMediaSequenceCalculatorOptions::ext)
+      ->set_force_decoding_from_start_of_media(true);
+  SetUpCalculator({}, {"AUDIO_DECODER_OPTIONS:audio_decoder_options"}, {},
+                  &options);
+  runner_->MutableSidePackets()->Tag("SEQUENCE_EXAMPLE") =
+      Adopt(sequence_.release());
+  MP_ASSERT_OK(runner_->Run());
+
+  MP_EXPECT_OK(runner_->OutputSidePackets()
+                   .Tag("AUDIO_DECODER_OPTIONS")
+                   .ValidateAsType<AudioDecoderOptions>());
+  EXPECT_NEAR(runner_->OutputSidePackets()
+                  .Tag("AUDIO_DECODER_OPTIONS")
+                  .Get<AudioDecoderOptions>()
+                  .start_time(),
+              0.0, 1e-5);
+  EXPECT_NEAR(runner_->OutputSidePackets()
+                  .Tag("AUDIO_DECODER_OPTIONS")
+                  .Get<AudioDecoderOptions>()
+                  .end_time(),
+              7.0, 1e-5);
+}
+
 TEST_F(UnpackMediaSequenceCalculatorTest, GetPacketResamplingOptions) {
   // TODO: Suport proto3 proto.Any in CalculatorOptions.
   // TODO: Avoid proto2 extensions in "RESAMPLER_OPTIONS".

mediapipe/calculators/tflite/BUILD

@@ -195,6 +195,12 @@ cc_test(
     ],
 )
 
+cc_library(
+    name = "util",
+    hdrs = ["util.h"],
+    alwayslink = 1,
+)
+
 cc_library(
     name = "tflite_inference_calculator",
     srcs = ["tflite_inference_calculator.cc"],
@@ -214,6 +220,7 @@ cc_library(
     }),
     visibility = ["//visibility:public"],
     deps = [
+        ":util",
         ":tflite_inference_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/util:resource_util",
@@ -222,20 +229,25 @@ cc_library(
         "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
         "//mediapipe/framework/port:ret_check",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//mediapipe:ios": [
+            "//mediapipe/gpu:MPPMetalHelper",
+            "//mediapipe/gpu:MPPMetalUtil",
+            "//mediapipe/gpu:gpu_buffer",
+            "//mediapipe/objc:mediapipe_framework_ios",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu/common:shape",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu/metal:buffer_convert",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate",
+        ],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gpu_buffer",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu/common:shape",
             "@org_tensorflow//tensorflow/lite/delegates/gpu:gl_delegate",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_buffer",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_program",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_shader",
         ],
-        "//mediapipe:ios": [
-            "//mediapipe/gpu:MPPMetalHelper",
-            "//mediapipe/objc:mediapipe_framework_ios",
-            "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate",
-        ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -259,33 +271,33 @@ cc_library(
     }),
     visibility = ["//visibility:public"],
     deps = [
+        ":util",
         ":tflite_converter_calculator_cc_proto",
         "//mediapipe/util:resource_util",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:image_frame",
         "//mediapipe/framework/formats:matrix",
         "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
-        "//mediapipe/framework/tool:status_util",
-        "//mediapipe/framework/port:status",
         "//mediapipe/framework/port:ret_check",
         "@org_tensorflow//tensorflow/lite:framework",
         "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
-            "//mediapipe/gpu:gl_calculator_helper",
-            "//mediapipe/gpu:gpu_buffer",
-            "@org_tensorflow//tensorflow/lite/delegates/gpu:gl_delegate",
-            "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_buffer",
-            "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_program",
-            "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_shader",
-        ],
         "//mediapipe:ios": [
+            "//mediapipe/gpu:MPPMetalUtil",
             "//mediapipe/gpu:gpu_buffer",
             "//mediapipe/gpu:MPPMetalHelper",
             "//mediapipe/objc:mediapipe_framework_ios",
             "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate",
         ],
-        "//conditions:default": [],
+        "//conditions:default": [
+            "//mediapipe/gpu:gpu_buffer",
+            "//mediapipe/gpu:gl_calculator_helper",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu:gl_delegate",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_buffer",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_program",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_shader",
+        ],
     }),
     alwayslink = 1,
 )
@@ -295,6 +307,7 @@ cc_library(
     srcs = ["tflite_tensors_to_segmentation_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
+        ":util",
         ":tflite_tensors_to_segmentation_calculator_cc_proto",
         "@com_google_absl//absl/strings:str_format",
         "@com_google_absl//absl/types:span",
@@ -308,7 +321,9 @@ cc_library(
         "//mediapipe/util:resource_util",
         "@org_tensorflow//tensorflow/lite:framework",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//mediapipe:ios": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
             "//mediapipe/gpu:gpu_buffer",
@@ -319,7 +334,6 @@ cc_library(
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_shader",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_texture",
         ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -346,8 +360,23 @@ cc_test(
 cc_library(
     name = "tflite_tensors_to_detections_calculator",
     srcs = ["tflite_tensors_to_detections_calculator.cc"],
+    copts = select({
+        "//mediapipe:ios": [
+            "-x objective-c++",
+            "-fobjc-arc",  # enable reference-counting
+        ],
+        "//conditions:default": [],
+    }),
+    linkopts = select({
+        "//mediapipe:ios": [
+            "-framework CoreVideo",
+            "-framework MetalKit",
+        ],
+        "//conditions:default": [],
+    }),
     visibility = ["//visibility:public"],
     deps = [
+        ":util",
         ":tflite_tensors_to_detections_calculator_cc_proto",
         "//mediapipe/framework/formats:detection_cc_proto",
         "@com_google_absl//absl/strings:str_format",
@@ -359,14 +388,21 @@ cc_library(
         "//mediapipe/framework/port:ret_check",
         "@org_tensorflow//tensorflow/lite:framework",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//mediapipe:ios": [
+            "//mediapipe/gpu:MPPMetalUtil",
+            "//mediapipe/gpu:gpu_buffer",
+            "//mediapipe/gpu:MPPMetalHelper",
+            "//mediapipe/objc:mediapipe_framework_ios",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate",
+        ],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "@org_tensorflow//tensorflow/lite/delegates/gpu:gl_delegate",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_buffer",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_program",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_shader",
         ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )

mediapipe/calculators/tflite/tflite_converter_calculator.cc

@@ -16,23 +16,23 @@
 #include <vector>
 
 #include "mediapipe/calculators/tflite/tflite_converter_calculator.pb.h"
+#include "mediapipe/calculators/tflite/util.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame.h"
 #include "mediapipe/framework/formats/matrix.h"
-#include "mediapipe/framework/port/canonical_errors.h"
 #include "mediapipe/framework/port/ret_check.h"
 #include "mediapipe/util/resource_util.h"
 #include "tensorflow/lite/error_reporter.h"
 #include "tensorflow/lite/interpreter.h"
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gpu_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_program.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_shader.h"
 #include "tensorflow/lite/delegates/gpu/gl_delegate.h"
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 #if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
 #import <CoreVideo/CoreVideo.h>
@@ -40,11 +40,12 @@
 #import <MetalKit/MetalKit.h>
 
 #import "mediapipe/gpu/MPPMetalHelper.h"
+#include "mediapipe/gpu/MPPMetalUtil.h"
 #include "mediapipe/gpu/gpu_buffer.h"
 #include "tensorflow/lite/delegates/gpu/metal_delegate.h"
 #endif  // iOS
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
 typedef ::tflite::gpu::gl::GlBuffer GpuTensor;
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
 typedef id<MTLBuffer> GpuTensor;
@@ -66,26 +67,27 @@ typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor>
 
 namespace mediapipe {
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
-using ::tflite::gpu::gl::GlBuffer;
+using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
 using ::tflite::gpu::gl::GlProgram;
 using ::tflite::gpu::gl::GlShader;
 struct GPUData {
   int elements = 1;
-  GlBuffer buffer;
+  GpuTensor buffer;
   GlShader shader;
   GlProgram program;
 };
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
 struct GPUData {
   int elements = 1;
-  id<MTLBuffer> buffer;
+  GpuTensor buffer;
   id<MTLComputePipelineState> pipeline_state;
 };
 #endif
 
 // Calculator for normalizing and converting an ImageFrame or Matrix
-// into a TfLiteTensor (float 32) or a GpuBuffer to a tflite::gpu::GlBuffer.
+// into a TfLiteTensor (float 32) or a GpuBuffer to a tflite::gpu::GlBuffer
+// or MTLBuffer.
 //
 // This calculator is designed to be used with the TfLiteInferenceCalcualtor,
 // as a pre-processing step for calculator inputs.
@@ -102,7 +104,7 @@ struct GPUData {
 // Output:
 //  One of the following tags:
 //  TENSORS - Vector of TfLiteTensor of type kTfLiteFloat32, or kTfLiteUint8.
-//  TENSORS_GPU - vector of GlBuffer.
+//  TENSORS_GPU - vector of GlBuffer or MTLBuffer.
 //
 // Example use:
 // node {
@@ -144,7 +146,7 @@ class TfLiteConverterCalculator : public CalculatorBase {
 
   std::unique_ptr<tflite::Interpreter> interpreter_ = nullptr;
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::unique_ptr<GPUData> gpu_data_out_;
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
@@ -175,25 +177,33 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
   RET_CHECK(cc->Outputs().HasTag("TENSORS") ^
             cc->Outputs().HasTag("TENSORS_GPU"));
 
+  bool use_gpu = false;
+
   if (cc->Inputs().HasTag("IMAGE")) cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
   if (cc->Inputs().HasTag("MATRIX")) cc->Inputs().Tag("MATRIX").Set<Matrix>();
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("IMAGE_GPU"))
+  if (cc->Inputs().HasTag("IMAGE_GPU")) {
     cc->Inputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
-#endif
+    use_gpu |= true;
+  }
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   if (cc->Outputs().HasTag("TENSORS"))
     cc->Outputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Outputs().HasTag("TENSORS_GPU"))
+  if (cc->Outputs().HasTag("TENSORS_GPU")) {
     cc->Outputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
-#endif
+    use_gpu |= true;
+  }
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
-#if defined(__ANDROID__)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
     MP_RETURN_IF_ERROR([MPPMetalHelper updateContract:cc]);
 #endif
+  }
 
   // Assign this calculator's default InputStreamHandler.
   cc->SetInputStreamHandler("FixedSizeInputStreamHandler");
@@ -208,10 +218,10 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
   if (cc->Inputs().HasTag("IMAGE_GPU") ||
       cc->Outputs().HasTag("IMAGE_OUT_GPU")) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     use_gpu_ = true;
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
 #endif
   }
 
@@ -221,7 +231,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
               cc->Outputs().HasTag("TENSORS_GPU"));
     // Cannot use quantization.
     use_quantized_tensors_ = false;
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
     gpu_helper_ = [[MPPMetalHelper alloc] initWithCalculatorContext:cc];
@@ -238,6 +248,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
 ::mediapipe::Status TfLiteConverterCalculator::Process(CalculatorContext* cc) {
   if (use_gpu_) {
+    // GpuBuffer to tflite::gpu::GlBuffer conversion.
     if (!initialized_) {
       MP_RETURN_IF_ERROR(InitGpu(cc));
       initialized_ = true;
@@ -253,7 +264,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 }
 
 ::mediapipe::Status TfLiteConverterCalculator::Close(CalculatorContext* cc) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   gpu_helper_.RunInGlContext([this] { gpu_data_out_.reset(); });
 #endif
 #if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
@@ -372,7 +383,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
 ::mediapipe::Status TfLiteConverterCalculator::ProcessGPU(
     CalculatorContext* cc) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   // GpuBuffer to tflite::gpu::GlBuffer conversion.
   const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
   MP_RETURN_IF_ERROR(
@@ -381,17 +392,11 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
         auto src = gpu_helper_.CreateSourceTexture(input);
         glActiveTexture(GL_TEXTURE0 + 0);
         glBindTexture(GL_TEXTURE_2D, src.name());
-        auto status = gpu_data_out_->buffer.BindToIndex(1);
+        RET_CHECK_CALL(gpu_data_out_->buffer.BindToIndex(1));
-        if (!status.ok()) {
-          return ::mediapipe::InternalError(status.error_message());
-        }
         const tflite::gpu::uint3 workgroups = {
             NumGroups(input.width(), kWorkgroupSize),
             NumGroups(input.height(), kWorkgroupSize), 1};
-        status = gpu_data_out_->program.Dispatch(workgroups);
+        RET_CHECK_CALL(gpu_data_out_->program.Dispatch(workgroups));
-        if (!status.ok()) {
-          return ::mediapipe::InternalError(status.error_message());
-        }
         glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
         glBindTexture(GL_TEXTURE_2D, 0);
         src.Release();
@@ -400,17 +405,17 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
   // Copy into outputs.
   auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
+  MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
+      [this, &output_tensors]() -> ::mediapipe::Status {
         output_tensors->resize(1);
         {
-    GlBuffer& tensor = output_tensors->at(0);
+          GpuTensor& tensor = output_tensors->at(0);
-    using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
+          RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-    auto status = CreateReadWriteShaderStorageBuffer<float>(
+              gpu_data_out_->elements, &tensor));
-        gpu_data_out_->elements, &tensor);
+          RET_CHECK_CALL(CopyBuffer(gpu_data_out_->buffer, tensor));
-    if (!status.ok()) {
-      return ::mediapipe::InternalError(status.error_message());
-    }
-    tflite::gpu::gl::CopyBuffer(gpu_data_out_->buffer, tensor);
         }
+        return ::mediapipe::OkStatus();
+      }));
   cc->Outputs()
       .Tag("TENSORS_GPU")
       .Add(output_tensors.release(), cc->InputTimestamp());
@@ -438,61 +443,55 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
   }
 
   // Copy into outputs.
+  // TODO Avoid this copy.
   auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
+  output_tensors->resize(1);
   {
     id<MTLDevice> device = gpu_helper_.mtlDevice;
-    id<MTLCommandBuffer> command_buffer = [gpu_helper_ commandBuffer];
+    output_tensors->at(0) =
-    command_buffer.label = @"TfLiteConverterCalculatorCopy";
-    id<MTLBuffer> tensor =
         [device newBufferWithLength:gpu_data_out_->elements * sizeof(float)
                             options:MTLResourceStorageModeShared];
-    id<MTLBlitCommandEncoder> blit_command =
+    [MPPMetalUtil blitMetalBufferTo:output_tensors->at(0)
-        [command_buffer blitCommandEncoder];
+                               from:gpu_data_out_->buffer
-    [blit_command copyFromBuffer:gpu_data_out_->buffer
+                           blocking:true
-                    sourceOffset:0
+                      commandBuffer:[gpu_helper_ commandBuffer]];
-                        toBuffer:tensor
-               destinationOffset:0
-                            size:gpu_data_out_->elements * sizeof(float)];
-    [blit_command endEncoding];
-    [command_buffer commit];
-    [command_buffer waitUntilCompleted];
-
-    output_tensors->push_back(tensor);
   }
 
   cc->Outputs()
       .Tag("TENSORS_GPU")
       .Add(output_tensors.release(), cc->InputTimestamp());
 #else
-  RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+  RET_CHECK_FAIL() << "GPU processing is not enabled.";
 #endif
 
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status TfLiteConverterCalculator::InitGpu(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
-  // Configure inputs.
+  // Get input image sizes.
   const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
   mediapipe::ImageFormat::Format format =
       mediapipe::ImageFormatForGpuBufferFormat(input.format());
   gpu_data_out_ = absl::make_unique<GPUData>();
   gpu_data_out_->elements = input.height() * input.width() * max_num_channels_;
   const bool include_alpha = (max_num_channels_ == 4);
-  if (!(format == mediapipe::ImageFormat::SRGB ||
+  const bool single_channel = (max_num_channels_ == 1);
+  if (!(format == mediapipe::ImageFormat::GRAY8 ||
+        format == mediapipe::ImageFormat::SRGB ||
         format == mediapipe::ImageFormat::SRGBA))
     RET_CHECK_FAIL() << "Unsupported GPU input format.";
   if (include_alpha && (format != mediapipe::ImageFormat::SRGBA))
     RET_CHECK_FAIL() << "Num input channels is less than desired output.";
-#endif
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+  MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
+      [this, &include_alpha, &input, &single_channel]() -> ::mediapipe::Status {
         // Device memory.
-  auto status = ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer<float>(
+        RET_CHECK_CALL(
-      gpu_data_out_->elements, &gpu_data_out_->buffer);
+            ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer<float>(
-  if (!status.ok()) {
+                gpu_data_out_->elements, &gpu_data_out_->buffer));
-    return ::mediapipe::InternalError(status.error_message());
-  }
 
         // Shader to convert GL Texture to Shader Storage Buffer Object (SSBO),
         // with normalization to either: [0,1] or [-1,1].
@@ -505,33 +504,31 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
           void main() {
             ivec2 gid = ivec2(gl_GlobalInvocationID.xy);
             if (gid.x >= width_height.x || gid.y >= width_height.y) return;
-            $5  // pixel fetch
+            vec4 pixel = texelFetch(input_texture, gid, 0);
             $3  // normalize [-1,1]
             int linear_index = $7 * ($4 * width_height.x + gid.x);
-            output_data.elements[linear_index + 0] = pixel.x;
+            output_data.elements[linear_index + 0] = pixel.x;  // r channel
-            output_data.elements[linear_index + 1] = pixel.y;
+            $5  // g & b channels
-            output_data.elements[linear_index + 2] = pixel.z;
             $6  // alpha channel
           })",
             /*$0=*/kWorkgroupSize, /*$1=*/input.width(), /*$2=*/input.height(),
             /*$3=*/zero_center_ ? "pixel = (pixel - 0.5) * 2.0;" : "",
             /*$4=*/flip_vertically_ ? "(width_height.y - 1 - gid.y)" : "gid.y",
             /*$5=*/
-      include_alpha ? "vec4 pixel = texelFetch(input_texture, gid, 0);"
+            single_channel
-                    : "vec3 pixel = texelFetch(input_texture, gid, 0).xyz;",
+                ? ""
+                : R"(output_data.elements[linear_index + 1] = pixel.y;
+                            output_data.elements[linear_index + 2] = pixel.z;)",
             /*$6=*/
-      include_alpha ? "output_data.elements[linear_index + 3] = pixel.w;" : "",
+            include_alpha ? "output_data.elements[linear_index + 3] = pixel.w;"
-      /*$7=*/include_alpha ? 4 : 3);
+                          : "",
-  status = GlShader::CompileShader(GL_COMPUTE_SHADER, shader_source,
+            /*$7=*/max_num_channels_);
-                                   &gpu_data_out_->shader);
+        RET_CHECK_CALL(GlShader::CompileShader(GL_COMPUTE_SHADER, shader_source,
-  if (!status.ok()) {
+                                               &gpu_data_out_->shader));
-    return ::mediapipe::InternalError(status.error_message());
+        RET_CHECK_CALL(GlProgram::CreateWithShader(gpu_data_out_->shader,
-  }
+                                                   &gpu_data_out_->program));
-  status = GlProgram::CreateWithShader(gpu_data_out_->shader,
+        return ::mediapipe::OkStatus();
-                                       &gpu_data_out_->program);
+      }));
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
   RET_CHECK(include_alpha)
       << "iOS GPU inference currently accepts only RGBA input.";
@@ -546,8 +543,6 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
   // with normalization to either: [0,1] or [-1,1].
   const std::string shader_source = absl::Substitute(
       R"(
-  #include <simd/simd.h>
-
   #include <metal_stdlib>
 
   using namespace metal;
@@ -612,9 +607,9 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
   // Get desired way to handle input channels.
   max_num_channels_ = options.max_num_channels();
-  // Currently only alpha channel toggling is suppored.
+  CHECK_GE(max_num_channels_, 1);
-  CHECK_GE(max_num_channels_, 3);
   CHECK_LE(max_num_channels_, 4);
+  CHECK_NE(max_num_channels_, 2);
 #if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
   if (cc->Inputs().HasTag("IMAGE_GPU"))
     // Currently on iOS, tflite gpu input tensor must be 4 channels,

mediapipe/calculators/tflite/tflite_converter_calculator.proto

@@ -36,8 +36,7 @@ message TfLiteConverterCalculatorOptions {
   optional bool flip_vertically = 2 [default = false];
 
   // Controls how many channels of the input image get passed through to the
-  // tensor. Currently this only controls whether or not to ignore alpha
+  // tensor. Valid values are 1,3,4 only. Ignored for iOS GPU.
-  // channel, so it must be 3 or 4.
   optional int32 max_num_channels = 3 [default = 3];
 
   // The calculator expects Matrix inputs to be in column-major order. Set

mediapipe/calculators/tflite/tflite_inference_calculator.cc

@@ -12,10 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <cstring>
+#include <memory>
 #include <string>
 #include <vector>
 
 #include "mediapipe/calculators/tflite/tflite_inference_calculator.pb.h"
+#include "mediapipe/calculators/tflite/util.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/port/ret_check.h"
 #include "mediapipe/util/resource_util.h"
@@ -24,14 +27,15 @@
 #include "tensorflow/lite/kernels/register.h"
 #include "tensorflow/lite/model.h"
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gpu_buffer.h"
+#include "tensorflow/lite/delegates/gpu/common/shape.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_program.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_shader.h"
 #include "tensorflow/lite/delegates/gpu/gl_delegate.h"
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 #if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
 #import <CoreVideo/CoreVideo.h>
@@ -39,33 +43,42 @@
 #import <MetalKit/MetalKit.h>
 
 #import "mediapipe/gpu/MPPMetalHelper.h"
+#include "mediapipe/gpu/MPPMetalUtil.h"
+#include "mediapipe/gpu/gpu_buffer.h"
+#include "tensorflow/lite/delegates/gpu/common/shape.h"
+#include "tensorflow/lite/delegates/gpu/metal/buffer_convert.h"
 #include "tensorflow/lite/delegates/gpu/metal_delegate.h"
 #endif  // iOS
 
-#if defined(__ANDROID__)
+namespace {
+
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
 typedef ::tflite::gpu::gl::GlBuffer GpuTensor;
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
 typedef id<MTLBuffer> GpuTensor;
 #endif
 
+// Round up n to next multiple of m.
+size_t RoundUp(size_t n, size_t m) { return ((n + m - 1) / m) * m; }  // NOLINT
+}  // namespace
+
 // TfLiteInferenceCalculator File Layout:
 //  * Header
 //  * Core
 //  * Aux
 namespace mediapipe {
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+using ::tflite::gpu::gl::CopyBuffer;
+using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
 using ::tflite::gpu::gl::GlBuffer;
-using ::tflite::gpu::gl::GlProgram;
+#endif
-using ::tflite::gpu::gl::GlShader;
+
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 struct GPUData {
   int elements = 1;
-  GlBuffer buffer;
+  GpuTensor buffer;
-};
+  ::tflite::gpu::BHWC shape;
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
-struct GPUData {
-  int elements = 1;
-  id<MTLBuffer> buffer;
 };
 #endif
 
@@ -134,7 +147,7 @@ class TfLiteInferenceCalculator : public CalculatorBase {
   std::unique_ptr<tflite::FlatBufferModel> model_;
   TfLiteDelegate* delegate_ = nullptr;
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::unique_ptr<GPUData> gpu_data_in_;
   std::vector<std::unique_ptr<GPUData>> gpu_data_out_;
@@ -142,6 +155,7 @@ class TfLiteInferenceCalculator : public CalculatorBase {
   MPPMetalHelper* gpu_helper_ = nullptr;
   std::unique_ptr<GPUData> gpu_data_in_;
   std::vector<std::unique_ptr<GPUData>> gpu_data_out_;
+  TFLBufferConvert* converter_from_BPHWC4_ = nil;
 #endif
 
   std::string model_path_ = "";
@@ -161,19 +175,25 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   RET_CHECK(cc->Outputs().HasTag("TENSORS") ^
             cc->Outputs().HasTag("TENSORS_GPU"));
 
+  bool use_gpu = false;
+
   if (cc->Inputs().HasTag("TENSORS"))
     cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("TENSORS_GPU"))
+  if (cc->Inputs().HasTag("TENSORS_GPU")) {
     cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
-#endif
+    use_gpu |= true;
+  }
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   if (cc->Outputs().HasTag("TENSORS"))
     cc->Outputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Outputs().HasTag("TENSORS_GPU"))
+  if (cc->Outputs().HasTag("TENSORS_GPU")) {
     cc->Outputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
-#endif
+    use_gpu |= true;
+  }
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   if (cc->InputSidePackets().HasTag("CUSTOM_OP_RESOLVER")) {
     cc->InputSidePackets()
@@ -181,11 +201,17 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
         .Set<tflite::ops::builtin::BuiltinOpResolver>();
   }
 
-#if defined(__ANDROID__)
+  const auto& options =
+      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
+  use_gpu |= options.use_gpu();
+
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
     MP_RETURN_IF_ERROR([MPPMetalHelper updateContract:cc]);
 #endif
+  }
 
   // Assign this calculator's default InputStreamHandler.
   cc->SetInputStreamHandler("FixedSizeInputStreamHandler");
@@ -199,35 +225,41 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   MP_RETURN_IF_ERROR(LoadOptions(cc));
 
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     gpu_input_ = true;
     gpu_inference_ = true;  // Inference must be on GPU also.
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK(!cc->Inputs().HasTag("TENSORS_GPU"))
-#endif
+        << "GPU processing not enabled.";
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   if (cc->Outputs().HasTag("TENSORS_GPU")) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     gpu_output_ = true;
     RET_CHECK(cc->Inputs().HasTag("TENSORS_GPU"))
         << "GPU output must also have GPU Input.";
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK(!cc->Inputs().HasTag("TENSORS_GPU"))
-#endif
+        << "GPU processing not enabled.";
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   MP_RETURN_IF_ERROR(LoadModel(cc));
 
   if (gpu_inference_) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
     gpu_helper_ = [[MPPMetalHelper alloc] initWithCalculatorContext:cc];
     RET_CHECK(gpu_helper_);
 #endif
-
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
+        [this, &cc]() -> ::mediapipe::Status { return LoadDelegate(cc); }));
+#else
     MP_RETURN_IF_ERROR(LoadDelegate(cc));
+#endif
   }
 
   return ::mediapipe::OkStatus();
@@ -237,35 +269,27 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   // 1. Receive pre-processed tensor inputs.
   if (gpu_input_) {
     // Read GPU input into SSBO.
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     const auto& input_tensors =
         cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
     RET_CHECK_EQ(input_tensors.size(), 1);
     MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
         [this, &input_tensors]() -> ::mediapipe::Status {
           // Explicit copy input.
-          tflite::gpu::gl::CopyBuffer(input_tensors[0], gpu_data_in_->buffer);
+          RET_CHECK_CALL(CopyBuffer(input_tensors[0], gpu_data_in_->buffer));
           return ::mediapipe::OkStatus();
         }));
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
     const auto& input_tensors =
         cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
     RET_CHECK_EQ(input_tensors.size(), 1);
-    id<MTLCommandBuffer> command_buffer = [gpu_helper_ commandBuffer];
-    command_buffer.label = @"TfLiteInferenceCalculatorInput";
-    id<MTLBlitCommandEncoder> blit_command =
-        [command_buffer blitCommandEncoder];
     // Explicit copy input.
-    [blit_command copyFromBuffer:input_tensors[0]
+    [MPPMetalUtil blitMetalBufferTo:gpu_data_in_->buffer
-                    sourceOffset:0
+                               from:input_tensors[0]
-                        toBuffer:gpu_data_in_->buffer
+                           blocking:true
-               destinationOffset:0
+                      commandBuffer:[gpu_helper_ commandBuffer]];
-                            size:gpu_data_in_->elements * sizeof(float)];
-    [blit_command endEncoding];
-    [command_buffer commit];
-    [command_buffer waitUntilCompleted];
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
 #endif
   } else {
     // Read CPU input into tensors.
@@ -278,18 +302,20 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       if (use_quantized_tensors_) {
         const uint8* input_tensor_buffer = input_tensor->data.uint8;
         uint8* local_tensor_buffer = interpreter_->typed_input_tensor<uint8>(i);
-        memcpy(local_tensor_buffer, input_tensor_buffer, input_tensor->bytes);
+        std::memcpy(local_tensor_buffer, input_tensor_buffer,
+                    input_tensor->bytes);
       } else {
         const float* input_tensor_buffer = input_tensor->data.f;
         float* local_tensor_buffer = interpreter_->typed_input_tensor<float>(i);
-        memcpy(local_tensor_buffer, input_tensor_buffer, input_tensor->bytes);
+        std::memcpy(local_tensor_buffer, input_tensor_buffer,
+                    input_tensor->bytes);
       }
     }
   }
 
   // 2. Run inference.
   if (gpu_inference_) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this]() -> ::mediapipe::Status {
           RET_CHECK_EQ(interpreter_->Invoke(), kTfLiteOk);
@@ -304,52 +330,51 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
   // 3. Output processed tensors.
   if (gpu_output_) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     // Output result tensors (GPU).
     auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
+    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
+        [this, &output_tensors]() -> ::mediapipe::Status {
           output_tensors->resize(gpu_data_out_.size());
           for (int i = 0; i < gpu_data_out_.size(); ++i) {
-      GlBuffer& tensor = output_tensors->at(i);
+            GpuTensor& tensor = output_tensors->at(i);
-      using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
+            RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-      auto status = CreateReadWriteShaderStorageBuffer<float>(
+                gpu_data_out_[i]->elements, &tensor));
-          gpu_data_out_[i]->elements, &tensor);
+            RET_CHECK_CALL(CopyBuffer(gpu_data_out_[i]->buffer, tensor));
-      if (!status.ok()) {
-        return ::mediapipe::InternalError(status.error_message());
-      }
-      tflite::gpu::gl::CopyBuffer(gpu_data_out_[i]->buffer, tensor);
           }
+          return ::mediapipe::OkStatus();
+        }));
     cc->Outputs()
         .Tag("TENSORS_GPU")
         .Add(output_tensors.release(), cc->InputTimestamp());
 #elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
     // Output result tensors (GPU).
     auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
+    output_tensors->resize(gpu_data_out_.size());
     id<MTLDevice> device = gpu_helper_.mtlDevice;
     id<MTLCommandBuffer> command_buffer = [gpu_helper_ commandBuffer];
-    command_buffer.label = @"TfLiteInferenceCalculatorOutput";
+    command_buffer.label = @"TfLiteInferenceBPHWC4Convert";
+    id<MTLComputeCommandEncoder> convert_command =
+        [command_buffer computeCommandEncoder];
     for (int i = 0; i < gpu_data_out_.size(); ++i) {
-      id<MTLBuffer> tensor =
+      output_tensors->at(i) =
           [device newBufferWithLength:gpu_data_out_[i]->elements * sizeof(float)
                               options:MTLResourceStorageModeShared];
-      id<MTLBlitCommandEncoder> blit_command =
+      // Reshape tensor.
-          [command_buffer blitCommandEncoder];
+      [converter_from_BPHWC4_ convertWithEncoder:convert_command
-      // Explicit copy input.
+                                           shape:gpu_data_out_[i]->shape
-      [blit_command copyFromBuffer:gpu_data_out_[i]->buffer
+                                    sourceBuffer:gpu_data_out_[i]->buffer
-                      sourceOffset:0
+                                 convertedBuffer:output_tensors->at(i)];
-                          toBuffer:tensor
+    }
-                 destinationOffset:0
+    [convert_command endEncoding];
-                              size:gpu_data_out_[i]->elements * sizeof(float)];
-      [blit_command endEncoding];
     [command_buffer commit];
     [command_buffer waitUntilCompleted];
-      output_tensors->push_back(tensor);
-    }
     cc->Outputs()
         .Tag("TENSORS_GPU")
         .Add(output_tensors.release(), cc->InputTimestamp());
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
-#endif
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     // Output result tensors (CPU).
     const auto& tensor_indexes = interpreter_->outputs();
@@ -367,7 +392,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
 ::mediapipe::Status TfLiteInferenceCalculator::Close(CalculatorContext* cc) {
   if (delegate_) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]() -> Status {
       TfLiteGpuDelegateDelete(delegate_);
       gpu_data_in_.reset();
@@ -446,7 +471,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
 ::mediapipe::Status TfLiteInferenceCalculator::LoadDelegate(
     CalculatorContext* cc) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   // Configure and create the delegate.
   TfLiteGpuDelegateOptions options = TfLiteGpuDelegateOptionsDefault();
   options.compile_options.precision_loss_allowed = 1;
@@ -466,15 +491,12 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
     for (int d = 0; d < tensor->dims->size; ++d) {
       gpu_data_in_->elements *= tensor->dims->data[d];
     }
-    // Input to model can be either RGB/RGBA only.
+    CHECK_GE(tensor->dims->data[3], 1);
-    RET_CHECK_GE(tensor->dims->data[3], 3);
+    CHECK_LE(tensor->dims->data[3], 4);
-    RET_CHECK_LE(tensor->dims->data[3], 4);
+    CHECK_NE(tensor->dims->data[3], 2);
     // Create and bind input buffer.
-    auto status = ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer<float>(
+    RET_CHECK_CALL(::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer<float>(
-        gpu_data_in_->elements, &gpu_data_in_->buffer);
+        gpu_data_in_->elements, &gpu_data_in_->buffer));
-    if (!status.ok()) {
-      return ::mediapipe::InternalError(status.error_message());
-    }
     RET_CHECK_EQ(TfLiteGpuDelegateBindBufferToTensor(
                      delegate_, gpu_data_in_->buffer.id(),
                      interpreter_->inputs()[0]),  // First tensor only
@@ -496,12 +518,8 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
     // Create and bind output buffers.
     interpreter_->SetAllowBufferHandleOutput(true);
     for (int i = 0; i < gpu_data_out_.size(); ++i) {
-      using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
+      RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-      auto status = CreateReadWriteShaderStorageBuffer<float>(
+          gpu_data_out_[i]->elements, &gpu_data_out_[i]->buffer));
-          gpu_data_out_[i]->elements, &gpu_data_out_[i]->buffer);
-      if (!status.ok()) {
-        return ::mediapipe::InternalError(status.error_message());
-      }
       RET_CHECK_EQ(
           TfLiteGpuDelegateBindBufferToTensor(
               delegate_, gpu_data_out_[i]->buffer.id(), output_indices[i]),
@@ -511,14 +529,15 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
   // Must call this last.
   RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_), kTfLiteOk);
-#endif  // __ANDROID__
+#endif  // OpenGL
 
 #if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
   // Configure and create the delegate.
   GpuDelegateOptions options;
   options.allow_precision_loss = false;  // Must match converter, F=float/T=half
-  options.wait_type = GpuDelegateOptions::WaitType::kActive;
+  options.wait_type = GpuDelegateOptions::WaitType::kPassive;
   if (!delegate_) delegate_ = TFLGpuDelegateCreate(&options);
+  id<MTLDevice> device = gpu_helper_.mtlDevice;
 
   if (gpu_input_) {
     // Get input image sizes.
@@ -539,11 +558,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       LOG(WARNING) << "Please ensure input GPU tensor is 4 channels.";
     }
     // Create and bind input buffer.
-    id<MTLDevice> device = gpu_helper_.mtlDevice;
     gpu_data_in_->buffer =
         [device newBufferWithLength:gpu_data_in_->elements * sizeof(float)
                             options:MTLResourceStorageModeShared];
-    // Must call this before TFLGpuDelegateBindMetalBufferToTensor.
     RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_), kTfLiteOk);
     RET_CHECK_EQ(TFLGpuDelegateBindMetalBufferToTensor(
                      delegate_,
@@ -561,12 +578,33 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       gpu_data_out_[i]->elements = 1;
       // TODO handle *2 properly on some dialated models
       for (int d = 0; d < tensor->dims->size; ++d) {
-        gpu_data_out_[i]->elements *= tensor->dims->data[d];
+        // Pad each dim for BHWC4 conversion inside delegate.
+        gpu_data_out_[i]->elements *= RoundUp(tensor->dims->data[d], 4);
+      }
+      // Save dimensions for reshaping back later.
+      gpu_data_out_[i]->shape.b = tensor->dims->data[0];
+      switch (tensor->dims->size) {
+        case 2:
+          gpu_data_out_[i]->shape.h = 1;
+          gpu_data_out_[i]->shape.w = 1;
+          gpu_data_out_[i]->shape.c = tensor->dims->data[1];
+          break;
+        case 3:
+          gpu_data_out_[i]->shape.h = 1;
+          gpu_data_out_[i]->shape.w = tensor->dims->data[1];
+          gpu_data_out_[i]->shape.c = tensor->dims->data[2];
+          break;
+        case 4:
+          gpu_data_out_[i]->shape.h = tensor->dims->data[1];
+          gpu_data_out_[i]->shape.w = tensor->dims->data[2];
+          gpu_data_out_[i]->shape.c = tensor->dims->data[3];
+          break;
+        default:
+          return mediapipe::InternalError("Unsupported tensor shape.");
       }
     }
     // Create and bind output buffers.
     interpreter_->SetAllowBufferHandleOutput(true);
-    id<MTLDevice> device = gpu_helper_.mtlDevice;
     for (int i = 0; i < gpu_data_out_.size(); ++i) {
       gpu_data_out_[i]->buffer =
           [device newBufferWithLength:gpu_data_out_[i]->elements * sizeof(float)
@@ -575,6 +613,14 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
                        delegate_, output_indices[i], gpu_data_out_[i]->buffer),
                    true);
     }
+    // Create converter for GPU output.
+    converter_from_BPHWC4_ = [[TFLBufferConvert alloc] initWithDevice:device
+                                                            isFloat16:false
+                                                      convertToPBHWC4:false];
+    if (converter_from_BPHWC4_ == nil) {
+      return mediapipe::InternalError(
+          "Error initializating output buffer converter");
+    }
   }
 #endif  // iOS
 

mediapipe/calculators/tflite/tflite_tensors_to_detections_calculator.cc

@@ -18,6 +18,7 @@
 #include "absl/strings/str_format.h"
 #include "absl/types/span.h"
 #include "mediapipe/calculators/tflite/tflite_tensors_to_detections_calculator.pb.h"
+#include "mediapipe/calculators/tflite/util.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/deps/file_path.h"
 #include "mediapipe/framework/formats/detection.pb.h"
@@ -26,28 +27,61 @@
 #include "mediapipe/framework/port/ret_check.h"
 #include "tensorflow/lite/interpreter.h"
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_program.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_shader.h"
 #include "tensorflow/lite/delegates/gpu/gl_delegate.h"
-#endif  // ANDROID
+#endif  // !MEDIAPIPE_DISABLE_GPU
 
-#if defined(__ANDROID__)
+#if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
-using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
+#import <CoreVideo/CoreVideo.h>
-using ::tflite::gpu::gl::GlBuffer;
+#import <Metal/Metal.h>
-using ::tflite::gpu::gl::GlProgram;
+#import <MetalKit/MetalKit.h>
-using ::tflite::gpu::gl::GlShader;
-#endif  // ANDROID
 
-namespace mediapipe {
+#import "mediapipe/gpu/MPPMetalHelper.h"
+#include "mediapipe/gpu/MPPMetalUtil.h"
+#include "mediapipe/gpu/gpu_buffer.h"
+#include "tensorflow/lite/delegates/gpu/metal_delegate.h"
+#endif  // iOS
 
 namespace {
 
 constexpr int kNumInputTensorsWithAnchors = 3;
 constexpr int kNumCoordsPerBox = 4;
 
+}  // namespace
+
+namespace mediapipe {
+
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
+using ::tflite::gpu::gl::GlShader;
+#endif
+
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+typedef ::tflite::gpu::gl::GlBuffer GpuTensor;
+typedef ::tflite::gpu::gl::GlProgram GpuProgram;
+#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+typedef id<MTLBuffer> GpuTensor;
+typedef id<MTLComputePipelineState> GpuProgram;
+#endif
+
+namespace {
+
+#if !defined(MEDIAPIPE_DISABLE_GPU)
+struct GPUData {
+  GpuProgram decode_program;
+  GpuProgram score_program;
+  GpuTensor decoded_boxes_buffer;
+  GpuTensor raw_boxes_buffer;
+  GpuTensor raw_anchors_buffer;
+  GpuTensor scored_boxes_buffer;
+  GpuTensor raw_scores_buffer;
+};
+#endif
+
 void ConvertRawValuesToAnchors(const float* raw_anchors, int num_boxes,
                                std::vector<Anchor>* anchors) {
   anchors->clear();
@@ -88,7 +122,7 @@ void ConvertAnchorsToRawValues(const std::vector<Anchor>& anchors,
 //               optional to pass in a third tensor for anchors (e.g. for SSD
 //               models) depend on the outputs of the detection model. The size
 //               of anchor tensor must be (num_boxes * 4).
-//  TENSORS_GPU - vector of GlBuffer.
+//  TENSORS_GPU - vector of GlBuffer of MTLBuffer.
 // Output:
 //  DETECTIONS - Result MediaPipe detections.
 //
@@ -126,7 +160,7 @@ class TfLiteTensorsToDetectionsCalculator : public CalculatorBase {
                                  std::vector<Detection>* output_detections);
 
   ::mediapipe::Status LoadOptions(CalculatorContext* cc);
-  ::mediapipe::Status GlSetup(CalculatorContext* cc);
+  ::mediapipe::Status GpuInit(CalculatorContext* cc);
   ::mediapipe::Status DecodeBoxes(const float* raw_boxes,
                                   const std::vector<Anchor>& anchors,
                                   std::vector<float>* boxes);
@@ -146,15 +180,12 @@ class TfLiteTensorsToDetectionsCalculator : public CalculatorBase {
   std::vector<Anchor> anchors_;
   bool side_packet_anchors_{};
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
-  std::unique_ptr<GlProgram> decode_program_;
+  std::unique_ptr<GPUData> gpu_data_;
-  std::unique_ptr<GlProgram> score_program_;
+#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
-  std::unique_ptr<GlBuffer> decoded_boxes_buffer_;
+  MPPMetalHelper* gpu_helper_ = nullptr;
-  std::unique_ptr<GlBuffer> raw_boxes_buffer_;
+  std::unique_ptr<GPUData> gpu_data_;
-  std::unique_ptr<GlBuffer> raw_anchors_buffer_;
-  std::unique_ptr<GlBuffer> scored_boxes_buffer_;
-  std::unique_ptr<GlBuffer> raw_scores_buffer_;
 #endif
 
   bool gpu_input_ = false;
@@ -167,15 +198,18 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
   RET_CHECK(!cc->Inputs().GetTags().empty());
   RET_CHECK(!cc->Outputs().GetTags().empty());
 
+  bool use_gpu = false;
+
   if (cc->Inputs().HasTag("TENSORS")) {
     cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
   }
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
-    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GlBuffer>>();
+    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+    use_gpu |= true;
   }
-#endif
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   if (cc->Outputs().HasTag("DETECTIONS")) {
     cc->Outputs().Tag("DETECTIONS").Set<std::vector<Detection>>();
@@ -187,9 +221,13 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
     }
   }
 
-#if defined(__ANDROID__)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
+#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+    MP_RETURN_IF_ERROR([MPPMetalHelper updateContract:cc]);
 #endif
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -200,8 +238,11 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
     gpu_input_ = true;
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
+#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+    gpu_helper_ = [[MPPMetalHelper alloc] initWithCalculatorContext:cc];
+    RET_CHECK(gpu_helper_);
 #endif
   }
 
@@ -209,7 +250,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
   side_packet_anchors_ = cc->InputSidePackets().HasTag("ANCHORS");
 
   if (gpu_input_) {
-    MP_RETURN_IF_ERROR(GlSetup(cc));
+    MP_RETURN_IF_ERROR(GpuInit(cc));
   }
 
   return ::mediapipe::OkStatus();
@@ -228,7 +269,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
     MP_RETURN_IF_ERROR(ProcessGPU(cc, output_detections.get()));
   } else {
     MP_RETURN_IF_ERROR(ProcessCPU(cc, output_detections.get()));
-  }  // if gpu_input_
+  }
 
   // Output
   if (cc->Outputs().HasTag("DETECTIONS")) {
@@ -245,7 +286,8 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
   const auto& input_tensors =
       cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
 
-  if (input_tensors.size() == 2) {
+  if (input_tensors.size() == 2 ||
+      input_tensors.size() == kNumInputTensorsWithAnchors) {
     // Postprocessing on CPU for model without postprocessing op. E.g. output
     // raw score tensor and box tensor. Anchor decoding will be handled below.
     const TfLiteTensor* raw_box_tensor = &input_tensors[0];
@@ -358,13 +400,17 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 }
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::ProcessGPU(
     CalculatorContext* cc, std::vector<Detection>* output_detections) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GlBuffer>>();
+      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+  RET_CHECK_GE(input_tensors.size(), 2);
 
+  MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this, &input_tensors, &cc,
+                                                 &output_detections]()
+                                                    -> ::mediapipe::Status {
     // Copy inputs.
-  tflite::gpu::gl::CopyBuffer(input_tensors[0], *raw_boxes_buffer_.get());
+    RET_CHECK_CALL(CopyBuffer(input_tensors[0], gpu_data_->raw_boxes_buffer));
-  tflite::gpu::gl::CopyBuffer(input_tensors[1], *raw_scores_buffer_.get());
+    RET_CHECK_CALL(CopyBuffer(input_tensors[1], gpu_data_->raw_scores_buffer));
     if (!anchors_init_) {
       if (side_packet_anchors_) {
         CHECK(!cc->InputSidePackets().Tag("ANCHORS").IsEmpty());
@@ -372,44 +418,36 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
             cc->InputSidePackets().Tag("ANCHORS").Get<std::vector<Anchor>>();
         std::vector<float> raw_anchors(num_boxes_ * kNumCoordsPerBox);
         ConvertAnchorsToRawValues(anchors, num_boxes_, raw_anchors.data());
-      raw_anchors_buffer_->Write<float>(absl::MakeSpan(raw_anchors));
+        RET_CHECK_CALL(gpu_data_->raw_anchors_buffer.Write<float>(
+            absl::MakeSpan(raw_anchors)));
       } else {
-      CHECK_EQ(input_tensors.size(), 3);
+        CHECK_EQ(input_tensors.size(), kNumInputTensorsWithAnchors);
-      tflite::gpu::gl::CopyBuffer(input_tensors[2], *raw_anchors_buffer_.get());
+        RET_CHECK_CALL(
+            CopyBuffer(input_tensors[2], gpu_data_->raw_anchors_buffer));
       }
       anchors_init_ = true;
     }
 
     // Run shaders.
-  MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
-      [this, &input_tensors]() -> ::mediapipe::Status {
     // Decode boxes.
-        decoded_boxes_buffer_->BindToIndex(0);
+    RET_CHECK_CALL(gpu_data_->decoded_boxes_buffer.BindToIndex(0));
-        raw_boxes_buffer_->BindToIndex(1);
+    RET_CHECK_CALL(gpu_data_->raw_boxes_buffer.BindToIndex(1));
-        raw_anchors_buffer_->BindToIndex(2);
+    RET_CHECK_CALL(gpu_data_->raw_anchors_buffer.BindToIndex(2));
     const tflite::gpu::uint3 decode_workgroups = {num_boxes_, 1, 1};
-        decode_program_->Dispatch(decode_workgroups);
+    RET_CHECK_CALL(gpu_data_->decode_program.Dispatch(decode_workgroups));
 
     // Score boxes.
-        scored_boxes_buffer_->BindToIndex(0);
+    RET_CHECK_CALL(gpu_data_->scored_boxes_buffer.BindToIndex(0));
-        raw_scores_buffer_->BindToIndex(1);
+    RET_CHECK_CALL(gpu_data_->raw_scores_buffer.BindToIndex(1));
     const tflite::gpu::uint3 score_workgroups = {num_boxes_, 1, 1};
-        score_program_->Dispatch(score_workgroups);
+    RET_CHECK_CALL(gpu_data_->score_program.Dispatch(score_workgroups));
-
-        return ::mediapipe::OkStatus();
-      }));
 
     // Copy decoded boxes from GPU to CPU.
     std::vector<float> boxes(num_boxes_ * num_coords_);
-  auto status = decoded_boxes_buffer_->Read(absl::MakeSpan(boxes));
+    RET_CHECK_CALL(gpu_data_->decoded_boxes_buffer.Read(absl::MakeSpan(boxes)));
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     std::vector<float> score_class_id_pairs(num_boxes_ * 2);
-  status = scored_boxes_buffer_->Read(absl::MakeSpan(score_class_id_pairs));
+    RET_CHECK_CALL(gpu_data_->scored_boxes_buffer.Read(
-  if (!status.ok()) {
+        absl::MakeSpan(score_class_id_pairs)));
-    return ::mediapipe::InternalError(status.error_message());
-  }
 
     // TODO: b/138851969. Is it possible to output a float vector
     // for score and an int vector for class so that we can avoid copying twice?
@@ -419,28 +457,116 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
       detection_scores[i] = score_class_id_pairs[i * 2];
       detection_classes[i] = static_cast<int>(score_class_id_pairs[i * 2 + 1]);
     }
+    MP_RETURN_IF_ERROR(
+        ConvertToDetections(boxes.data(), detection_scores.data(),
+                            detection_classes.data(), output_detections));
+
+    return ::mediapipe::OkStatus();
+  }));
+#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+
+  const auto& input_tensors =
+      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+  RET_CHECK_GE(input_tensors.size(), 2);
+
+  // Copy inputs.
+  [MPPMetalUtil blitMetalBufferTo:gpu_data_->raw_boxes_buffer
+                             from:input_tensors[0]
+                         blocking:true
+                    commandBuffer:[gpu_helper_ commandBuffer]];
+  [MPPMetalUtil blitMetalBufferTo:gpu_data_->raw_scores_buffer
+                             from:input_tensors[1]
+                         blocking:true
+                    commandBuffer:[gpu_helper_ commandBuffer]];
+  if (!anchors_init_) {
+    if (side_packet_anchors_) {
+      CHECK(!cc->InputSidePackets().Tag("ANCHORS").IsEmpty());
+      const auto& anchors =
+          cc->InputSidePackets().Tag("ANCHORS").Get<std::vector<Anchor>>();
+      std::vector<float> raw_anchors(num_boxes_ * kNumCoordsPerBox);
+      ConvertAnchorsToRawValues(anchors, num_boxes_, raw_anchors.data());
+      memcpy([gpu_data_->raw_anchors_buffer contents], raw_anchors.data(),
+             raw_anchors.size() * sizeof(float));
+    } else {
+      RET_CHECK_EQ(input_tensors.size(), kNumInputTensorsWithAnchors);
+      [MPPMetalUtil blitMetalBufferTo:gpu_data_->raw_anchors_buffer
+                                 from:input_tensors[2]
+                             blocking:true
+                        commandBuffer:[gpu_helper_ commandBuffer]];
+    }
+    anchors_init_ = true;
+  }
+
+  // Run shaders.
+  {
+    id<MTLCommandBuffer> command_buffer = [gpu_helper_ commandBuffer];
+    command_buffer.label = @"TfLiteDecodeBoxes";
+    id<MTLComputeCommandEncoder> decode_command =
+        [command_buffer computeCommandEncoder];
+    [decode_command setComputePipelineState:gpu_data_->decode_program];
+    [decode_command setBuffer:gpu_data_->decoded_boxes_buffer
+                       offset:0
+                      atIndex:0];
+    [decode_command setBuffer:gpu_data_->raw_boxes_buffer offset:0 atIndex:1];
+    [decode_command setBuffer:gpu_data_->raw_anchors_buffer offset:0 atIndex:2];
+    MTLSize decode_threads_per_group = MTLSizeMake(1, 1, 1);
+    MTLSize decode_threadgroups = MTLSizeMake(num_boxes_, 1, 1);
+    [decode_command dispatchThreadgroups:decode_threadgroups
+                   threadsPerThreadgroup:decode_threads_per_group];
+    [decode_command endEncoding];
+    [command_buffer commit];
+    [command_buffer waitUntilCompleted];
+  }
+  {
+    id<MTLCommandBuffer> command_buffer = [gpu_helper_ commandBuffer];
+    command_buffer.label = @"TfLiteScoreBoxes";
+    id<MTLComputeCommandEncoder> score_command =
+        [command_buffer computeCommandEncoder];
+    [score_command setComputePipelineState:gpu_data_->score_program];
+    [score_command setBuffer:gpu_data_->scored_boxes_buffer offset:0 atIndex:0];
+    [score_command setBuffer:gpu_data_->raw_scores_buffer offset:0 atIndex:1];
+    MTLSize score_threads_per_group = MTLSizeMake(1, num_classes_, 1);
+    MTLSize score_threadgroups = MTLSizeMake(num_boxes_, 1, 1);
+    [score_command dispatchThreadgroups:score_threadgroups
+                  threadsPerThreadgroup:score_threads_per_group];
+    [score_command endEncoding];
+    [command_buffer commit];
+    [command_buffer waitUntilCompleted];
+  }
+
+  // Copy decoded boxes from GPU to CPU.
+  std::vector<float> boxes(num_boxes_ * num_coords_);
+  memcpy(boxes.data(), [gpu_data_->decoded_boxes_buffer contents],
+         num_boxes_ * num_coords_ * sizeof(float));
+  std::vector<float> score_class_id_pairs(num_boxes_ * 2);
+  memcpy(score_class_id_pairs.data(), [gpu_data_->scored_boxes_buffer contents],
+         num_boxes_ * 2 * sizeof(float));
+
+  // Output detections.
+  // TODO Adjust shader to avoid copying shader output twice.
+  std::vector<float> detection_scores(num_boxes_);
+  std::vector<int> detection_classes(num_boxes_);
+  for (int i = 0; i < num_boxes_; ++i) {
+    detection_scores[i] = score_class_id_pairs[i * 2];
+    detection_classes[i] = static_cast<int>(score_class_id_pairs[i * 2 + 1]);
+  }
   MP_RETURN_IF_ERROR(ConvertToDetections(boxes.data(), detection_scores.data(),
                                          detection_classes.data(),
                                          output_detections));
+
 #else
   LOG(ERROR) << "GPU input on non-Android not supported yet.";
-#endif  // defined(__ANDROID__)
+#endif
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::Close(
     CalculatorContext* cc) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
-  gpu_helper_.RunInGlContext([this] {
+  gpu_helper_.RunInGlContext([this] { gpu_data_.reset(); });
-    decode_program_.reset();
+#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
-    score_program_.reset();
+  gpu_data_.reset();
-    decoded_boxes_buffer_.reset();
+#endif                                      //  !MEDIAPIPE_DISABLE_GPU
-    raw_boxes_buffer_.reset();
-    raw_anchors_buffer_.reset();
-    scored_boxes_buffer_.reset();
-    raw_scores_buffer_.reset();
-  });
-#endif  // __ANDROID__
 
   return ::mediapipe::OkStatus();
 }
@@ -530,6 +656,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
       }
     }
   }
+
   return ::mediapipe::OkStatus();
 }
 
@@ -586,9 +713,13 @@ Detection TfLiteTensorsToDetectionsCalculator::ConvertToDetection(
   return detection;
 }
 
-::mediapipe::Status TfLiteTensorsToDetectionsCalculator::GlSetup(
+::mediapipe::Status TfLiteTensorsToDetectionsCalculator::GpuInit(
     CalculatorContext* cc) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+  MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]()
+                                                    -> ::mediapipe::Status {
+    gpu_data_ = absl::make_unique<GPUData>();
+
     // A shader to decode detection boxes.
     const std::string decode_src = absl::Substitute(
         R"( #version 310 es
@@ -687,41 +818,23 @@ void main() {
 
     // Shader program
     GlShader decode_shader;
-  auto status =
+    RET_CHECK_CALL(
-      GlShader::CompileShader(GL_COMPUTE_SHADER, decode_src, &decode_shader);
+        GlShader::CompileShader(GL_COMPUTE_SHADER, decode_src, &decode_shader));
-  if (!status.ok()) {
+    RET_CHECK_CALL(GpuProgram::CreateWithShader(decode_shader,
-    return ::mediapipe::InternalError(status.error_message());
+                                                &gpu_data_->decode_program));
-  }
-  decode_program_ = absl::make_unique<GlProgram>();
-  status = GlProgram::CreateWithShader(decode_shader, decode_program_.get());
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     // Outputs
     size_t decoded_boxes_length = num_boxes_ * num_coords_;
-  decoded_boxes_buffer_ = absl::make_unique<GlBuffer>();
+    RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-  status = CreateReadWriteShaderStorageBuffer<float>(
+        decoded_boxes_length, &gpu_data_->decoded_boxes_buffer));
-      decoded_boxes_length, decoded_boxes_buffer_.get());
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     // Inputs
     size_t raw_boxes_length = num_boxes_ * num_coords_;
-  raw_boxes_buffer_ = absl::make_unique<GlBuffer>();
+    RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-  status = CreateReadWriteShaderStorageBuffer<float>(raw_boxes_length,
+        raw_boxes_length, &gpu_data_->raw_boxes_buffer));
-                                                     raw_boxes_buffer_.get());
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     size_t raw_anchors_length = num_boxes_ * kNumCoordsPerBox;
-  raw_anchors_buffer_ = absl::make_unique<GlBuffer>();
+    RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-  status = CreateReadWriteShaderStorageBuffer<float>(raw_anchors_length,
+        raw_anchors_length, &gpu_data_->raw_anchors_buffer));
-                                                     raw_anchors_buffer_.get());
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     // Parameters
-  glUseProgram(decode_program_->id());
+    glUseProgram(gpu_data_->decode_program.id());
     glUniform4f(0, options_.x_scale(), options_.y_scale(), options_.w_scale(),
                 options_.h_scale());
 
@@ -786,44 +899,268 @@ void main() {
         options_.has_score_clipping_thresh() ? 1 : 0,
         options_.has_score_clipping_thresh() ? options_.score_clipping_thresh()
                                              : 0,
-      ignore_classes_.size() ? 1 : 0);
+        !ignore_classes_.empty() ? 1 : 0);
 
     // # filter classes supported is hardware dependent.
     int max_wg_size;  //  typically <= 1024
-  glGetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_SIZE, 1, &max_wg_size);  // y-dim
+    glGetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_SIZE, 1,
-  CHECK_LT(num_classes_, max_wg_size) << "# classes must be < " << max_wg_size;
+                    &max_wg_size);  // y-dim
+    CHECK_LT(num_classes_, max_wg_size)
+        << "# classes must be < " << max_wg_size;
     // TODO support better filtering.
     CHECK_LE(ignore_classes_.size(), 1) << "Only ignore class 0 is allowed";
 
     // Shader program
     GlShader score_shader;
-  status = GlShader::CompileShader(GL_COMPUTE_SHADER, score_src, &score_shader);
+    RET_CHECK_CALL(
-  if (!status.ok()) {
+        GlShader::CompileShader(GL_COMPUTE_SHADER, score_src, &score_shader));
-    return ::mediapipe::InternalError(status.error_message());
+    RET_CHECK_CALL(
-  }
+        GpuProgram::CreateWithShader(score_shader, &gpu_data_->score_program));
-  score_program_ = absl::make_unique<GlProgram>();
-  status = GlProgram::CreateWithShader(score_shader, score_program_.get());
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     // Outputs
     size_t scored_boxes_length = num_boxes_ * 2;  // score, class
-  scored_boxes_buffer_ = absl::make_unique<GlBuffer>();
+    RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-  status = CreateReadWriteShaderStorageBuffer<float>(
+        scored_boxes_length, &gpu_data_->scored_boxes_buffer));
-      scored_boxes_length, scored_boxes_buffer_.get());
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     // Inputs
     size_t raw_scores_length = num_boxes_ * num_classes_;
-  raw_scores_buffer_ = absl::make_unique<GlBuffer>();
+    RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-  status = CreateReadWriteShaderStorageBuffer<float>(raw_scores_length,
+        raw_scores_length, &gpu_data_->raw_scores_buffer));
-                                                     raw_scores_buffer_.get());
+
-  if (!status.ok()) {
+    return ::mediapipe::OkStatus();
-    return ::mediapipe::InternalError(status.error_message());
+  }));
+
+#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+  // TODO consolidate Metal and OpenGL shaders via vulkan.
+
+  gpu_data_ = absl::make_unique<GPUData>();
+  id<MTLDevice> device = gpu_helper_.mtlDevice;
+
+  // A shader to decode detection boxes.
+  std::string decode_src = absl::Substitute(
+      R"(
+#include <metal_stdlib>
+
+using namespace metal;
+
+kernel void decodeKernel(
+    device float*                   boxes       [[ buffer(0) ]],
+    device float*                   raw_boxes   [[ buffer(1) ]],
+    device float*                   raw_anchors [[ buffer(2) ]],
+    uint2                           gid         [[ thread_position_in_grid ]]) {
+
+  uint num_coords = uint($0);
+  int reverse_output_order = int($1);
+  int apply_exponential = int($2);
+  int box_coord_offset = int($3);
+  int num_keypoints = int($4);
+  int keypt_coord_offset = int($5);
+  int num_values_per_keypt = int($6);
+)",
+      options_.num_coords(),  // box xywh
+      options_.reverse_output_order() ? 1 : 0,
+      options_.apply_exponential_on_box_size() ? 1 : 0,
+      options_.box_coord_offset(), options_.num_keypoints(),
+      options_.keypoint_coord_offset(), options_.num_values_per_keypoint());
+  decode_src += absl::Substitute(
+      R"(
+  float4 scale = float4(($0),($1),($2),($3));
+)",
+      options_.x_scale(), options_.y_scale(), options_.w_scale(),
+      options_.h_scale());
+  decode_src += R"(
+  uint g_idx = gid.x;
+  uint box_offset = g_idx * num_coords + uint(box_coord_offset);
+  uint anchor_offset = g_idx * uint(4);  // check kNumCoordsPerBox
+
+  float y_center, x_center, h, w;
+
+  if (reverse_output_order == int(0)) {
+    y_center = raw_boxes[box_offset + uint(0)];
+    x_center = raw_boxes[box_offset + uint(1)];
+    h = raw_boxes[box_offset + uint(2)];
+    w = raw_boxes[box_offset + uint(3)];
+  } else {
+    x_center = raw_boxes[box_offset + uint(0)];
+    y_center = raw_boxes[box_offset + uint(1)];
+    w = raw_boxes[box_offset + uint(2)];
+    h = raw_boxes[box_offset + uint(3)];
   }
 
-#endif  // defined(__ANDROID__)
+  float anchor_yc = raw_anchors[anchor_offset + uint(0)];
+  float anchor_xc = raw_anchors[anchor_offset + uint(1)];
+  float anchor_h  = raw_anchors[anchor_offset + uint(2)];
+  float anchor_w  = raw_anchors[anchor_offset + uint(3)];
+
+  x_center = x_center / scale.x * anchor_w + anchor_xc;
+  y_center = y_center / scale.y * anchor_h + anchor_yc;
+
+  if (apply_exponential == int(1)) {
+    h = exp(h / scale.w) * anchor_h;
+    w = exp(w / scale.z) * anchor_w;
+  } else {
+    h = (h / scale.w) * anchor_h;
+    w = (w / scale.z) * anchor_w;
+  }
+
+  float ymin = y_center - h / 2.0;
+  float xmin = x_center - w / 2.0;
+  float ymax = y_center + h / 2.0;
+  float xmax = x_center + w / 2.0;
+
+  boxes[box_offset + uint(0)] = ymin;
+  boxes[box_offset + uint(1)] = xmin;
+  boxes[box_offset + uint(2)] = ymax;
+  boxes[box_offset + uint(3)] = xmax;
+
+  if (num_keypoints > int(0)){
+    for (int k = 0; k < num_keypoints; ++k) {
+      int kp_offset =
+        int(g_idx * num_coords) + keypt_coord_offset + k * num_values_per_keypt;
+      float kp_y, kp_x;
+      if (reverse_output_order == int(0)) {
+        kp_y = raw_boxes[kp_offset + int(0)];
+        kp_x = raw_boxes[kp_offset + int(1)];
+      } else {
+        kp_x = raw_boxes[kp_offset + int(0)];
+        kp_y = raw_boxes[kp_offset + int(1)];
+      }
+      boxes[kp_offset + int(0)] = kp_x / scale.x * anchor_w + anchor_xc;
+      boxes[kp_offset + int(1)] = kp_y / scale.y * anchor_h + anchor_yc;
+    }
+  }
+})";
+
+  {
+    // Shader program
+    NSString* library_source =
+        [NSString stringWithUTF8String:decode_src.c_str()];
+    NSError* error = nil;
+    id<MTLLibrary> library = [device newLibraryWithSource:library_source
+                                                  options:nullptr
+                                                    error:&error];
+    RET_CHECK(library != nil) << "Couldn't create shader library "
+                              << [[error localizedDescription] UTF8String];
+    id<MTLFunction> kernel_func = nil;
+    kernel_func = [library newFunctionWithName:@"decodeKernel"];
+    RET_CHECK(kernel_func != nil) << "Couldn't create kernel function.";
+    gpu_data_->decode_program =
+        [device newComputePipelineStateWithFunction:kernel_func error:&error];
+    RET_CHECK(gpu_data_->decode_program != nil)
+        << "Couldn't create pipeline state "
+        << [[error localizedDescription] UTF8String];
+    // Outputs
+    size_t decoded_boxes_length = num_boxes_ * num_coords_ * sizeof(float);
+    gpu_data_->decoded_boxes_buffer =
+        [device newBufferWithLength:decoded_boxes_length
+                            options:MTLResourceStorageModeShared];
+    // Inputs
+    size_t raw_boxes_length = num_boxes_ * num_coords_ * sizeof(float);
+    gpu_data_->raw_boxes_buffer =
+        [device newBufferWithLength:raw_boxes_length
+                            options:MTLResourceStorageModeShared];
+    size_t raw_anchors_length = num_boxes_ * kNumCoordsPerBox * sizeof(float);
+    gpu_data_->raw_anchors_buffer =
+        [device newBufferWithLength:raw_anchors_length
+                            options:MTLResourceStorageModeShared];
+  }
+
+  // A shader to score detection boxes.
+  const std::string score_src = absl::Substitute(
+      R"(
+#include <metal_stdlib>
+
+using namespace metal;
+
+float optional_sigmoid(float x) {
+  int apply_sigmoid = int($1);
+  int apply_clipping_thresh = int($2);
+  float clipping_thresh = float($3);
+  if (apply_sigmoid == int(0)) return x;
+  if (apply_clipping_thresh == int(1)) {
+    x = clamp(x, -clipping_thresh, clipping_thresh);
+  }
+  x = 1.0 / (1.0 + exp(-x));
+  return x;
+}
+
+kernel void scoreKernel(
+    device float*             scored_boxes [[ buffer(0) ]],
+    device float*             raw_scores   [[ buffer(1) ]],
+    uint2                     tid          [[ thread_position_in_threadgroup ]],
+    uint2                     gid          [[ thread_position_in_grid ]]) {
+
+  uint num_classes = uint($0);
+  int apply_sigmoid = int($1);
+  int apply_clipping_thresh = int($2);
+  float clipping_thresh = float($3);
+  int ignore_class_0 = int($4);
+
+  uint g_idx = gid.x;   // box idx
+  uint s_idx = tid.y;   // score/class idx
+
+  // load all scores into shared memory
+  threadgroup float local_scores[$0];
+  float score = raw_scores[g_idx * num_classes + s_idx];
+  local_scores[s_idx] = optional_sigmoid(score);
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // find max score in shared memory
+  if (s_idx == uint(0)) {
+    float max_score = -FLT_MAX;
+    float max_class = -1.0;
+    for (int i=ignore_class_0; i<int(num_classes); ++i) {
+      if (local_scores[i] > max_score) {
+        max_score = local_scores[i];
+        max_class = float(i);
+      }
+    }
+    scored_boxes[g_idx * uint(2) + uint(0)] = max_score;
+    scored_boxes[g_idx * uint(2) + uint(1)] = max_class;
+  }
+})",
+      num_classes_, options_.sigmoid_score() ? 1 : 0,
+      options_.has_score_clipping_thresh() ? 1 : 0,
+      options_.has_score_clipping_thresh() ? options_.score_clipping_thresh()
+                                           : 0,
+      ignore_classes_.size() ? 1 : 0);
+
+  // TODO support better filtering.
+  CHECK_LE(ignore_classes_.size(), 1) << "Only ignore class 0 is allowed";
+
+  {
+    // Shader program
+    NSString* library_source =
+        [NSString stringWithUTF8String:score_src.c_str()];
+    NSError* error = nil;
+    id<MTLLibrary> library = [device newLibraryWithSource:library_source
+                                                  options:nullptr
+                                                    error:&error];
+    RET_CHECK(library != nil) << "Couldn't create shader library "
+                              << [[error localizedDescription] UTF8String];
+    id<MTLFunction> kernel_func = nil;
+    kernel_func = [library newFunctionWithName:@"scoreKernel"];
+    RET_CHECK(kernel_func != nil) << "Couldn't create kernel function.";
+    gpu_data_->score_program =
+        [device newComputePipelineStateWithFunction:kernel_func error:&error];
+    RET_CHECK(gpu_data_->score_program != nil)
+        << "Couldn't create pipeline state "
+        << [[error localizedDescription] UTF8String];
+    // Outputs
+    size_t scored_boxes_length = num_boxes_ * 2 * sizeof(float);  // score,class
+    gpu_data_->scored_boxes_buffer =
+        [device newBufferWithLength:scored_boxes_length
+                            options:MTLResourceStorageModeShared];
+    // Inputs
+    size_t raw_scores_length = num_boxes_ * num_classes_ * sizeof(float);
+    gpu_data_->raw_scores_buffer =
+        [device newBufferWithLength:raw_scores_length
+                            options:MTLResourceStorageModeShared];
+    // # filter classes supported is hardware dependent.
+    int max_wg_size = gpu_data_->score_program.maxTotalThreadsPerThreadgroup;
+    CHECK_LT(num_classes_, max_wg_size) << "# classes must be <" << max_wg_size;
+  }
+
+#endif  // __ANDROID__ or iOS
+
   return ::mediapipe::OkStatus();
 }
 

mediapipe/calculators/tflite/tflite_tensors_to_landmarks_calculator.cc

@@ -96,7 +96,8 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
               options_.has_input_image_width())
         << "Must provide input with/height for getting normalized landmarks.";
   }
-  if (cc->Outputs().HasTag("LANDMARKS") && options_.flip_vertically()) {
+  if (cc->Outputs().HasTag("LANDMARKS") &&
+      (options_.flip_vertically() || options_.flip_horizontally())) {
     RET_CHECK(options_.has_input_image_height() &&
               options_.has_input_image_width())
         << "Must provide input with/height for using flip_vertically option "
@@ -133,7 +134,12 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
   for (int ld = 0; ld < num_landmarks_; ++ld) {
     const int offset = ld * num_dimensions;
     Landmark landmark;
+
+    if (options_.flip_horizontally()) {
+      landmark.set_x(options_.input_image_width() - raw_landmarks[offset]);
+    } else {
       landmark.set_x(raw_landmarks[offset]);
+    }
     if (num_dimensions > 1) {
       if (options_.flip_vertically()) {
         landmark.set_y(options_.input_image_height() -

mediapipe/calculators/tflite/tflite_tensors_to_landmarks_calculator.proto

@@ -40,6 +40,12 @@ message TfLiteTensorsToLandmarksCalculatorOptions {
   // representation has a bottom-left origin (e.g., in OpenGL).
   optional bool flip_vertically = 4 [default = false];
 
+  // Whether the detection coordinates from the input tensors should be flipped
+  // horizontally (along the x-direction). This is useful, for example, when the
+  // input image is horizontally flipped in ImageTransformationCalculator
+  // beforehand.
+  optional bool flip_horizontally = 6 [default = false];
+
   // A value that z values should be divided by.
   optional float normalize_z = 5 [default = 1.0];
 }

mediapipe/calculators/tflite/tflite_tensors_to_segmentation_calculator.cc

@@ -17,6 +17,7 @@
 #include "absl/strings/str_format.h"
 #include "absl/types/span.h"
 #include "mediapipe/calculators/tflite/tflite_tensors_to_segmentation_calculator.pb.h"
+#include "mediapipe/calculators/tflite/util.h"
 #include "mediapipe/framework/calculator_context.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame.h"
@@ -27,7 +28,7 @@
 #include "mediapipe/util/resource_util.h"
 #include "tensorflow/lite/interpreter.h"
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
 #include "mediapipe/gpu/shader_util.h"
@@ -36,7 +37,7 @@
 #include "tensorflow/lite/delegates/gpu/gl/gl_shader.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_texture.h"
 #include "tensorflow/lite/delegates/gpu/gl_delegate.h"
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace {
 constexpr int kWorkgroupSize = 8;  // Block size for GPU shader.
@@ -52,12 +53,14 @@ float Clamp(float val, float min, float max) {
 
 namespace mediapipe {
 
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+using ::tflite::gpu::gl::CopyBuffer;
+using ::tflite::gpu::gl::CreateReadWriteRgbaImageTexture;
 using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
 using ::tflite::gpu::gl::GlBuffer;
 using ::tflite::gpu::gl::GlProgram;
 using ::tflite::gpu::gl::GlShader;
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 // Converts TFLite tensors from a tflite segmentation model to an image mask.
 //
@@ -126,13 +129,13 @@ class TfLiteTensorsToSegmentationCalculator : public CalculatorBase {
   int tensor_channels_ = 0;
 
   bool use_gpu_ = false;
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::unique_ptr<GlProgram> mask_program_with_prev_;
   std::unique_ptr<GlProgram> mask_program_no_prev_;
   std::unique_ptr<GlBuffer> tensor_buffer_;
   GLuint upsample_program_;
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 };
 REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 
@@ -142,6 +145,8 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   RET_CHECK(!cc->Inputs().GetTags().empty());
   RET_CHECK(!cc->Outputs().GetTags().empty());
 
+  bool use_gpu = false;
+
   // Inputs CPU.
   if (cc->Inputs().HasTag("TENSORS")) {
     cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
@@ -154,32 +159,37 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   }
 
   // Inputs GPU.
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
     cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GlBuffer>>();
+    use_gpu |= true;
   }
   if (cc->Inputs().HasTag("PREV_MASK_GPU")) {
     cc->Inputs().Tag("PREV_MASK_GPU").Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
   if (cc->Inputs().HasTag("REFERENCE_IMAGE_GPU")) {
     cc->Inputs().Tag("REFERENCE_IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   // Outputs.
   if (cc->Outputs().HasTag("MASK")) {
     cc->Outputs().Tag("MASK").Set<ImageFrame>();
   }
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   if (cc->Outputs().HasTag("MASK_GPU")) {
     cc->Outputs().Tag("MASK_GPU").Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
-#if defined(__ANDROID__)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
-
+  }
   return ::mediapipe::OkStatus();
 }
 
@@ -189,24 +199,23 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
     use_gpu_ = true;
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   MP_RETURN_IF_ERROR(LoadOptions(cc));
 
   if (use_gpu_) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
           MP_RETURN_IF_ERROR(InitGpu(cc));
           return ::mediapipe::OkStatus();
         }));
 #else
-    RET_CHECK_FAIL()
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
-        << "GPU processing on non-Android devices is not supported yet.";
+#endif  //  !MEDIAPIPE_DISABLE_GPU
-#endif  // __ANDROID__
   }
 
   return ::mediapipe::OkStatus();
@@ -215,13 +224,13 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::Process(
     CalculatorContext* cc) {
   if (use_gpu_) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
           MP_RETURN_IF_ERROR(ProcessGpu(cc));
           return ::mediapipe::OkStatus();
         }));
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     MP_RETURN_IF_ERROR(ProcessCpu(cc));
   }
@@ -231,7 +240,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::Close(
     CalculatorContext* cc) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   gpu_helper_.RunInGlContext([this] {
     if (upsample_program_) glDeleteProgram(upsample_program_);
     upsample_program_ = 0;
@@ -239,7 +248,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
     mask_program_no_prev_.reset();
     tensor_buffer_.reset();
   });
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
@@ -358,7 +367,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   if (cc->Inputs().Tag("TENSORS_GPU").IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   // Get input streams.
   const auto& input_tensors =
       cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GlBuffer>>();
@@ -379,9 +388,9 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 
   // Create initial working mask texture.
   ::tflite::gpu::gl::GlTexture small_mask_texture;
-  ::tflite::gpu::gl::CreateReadWriteRgbaImageTexture(
+  RET_CHECK_CALL(CreateReadWriteRgbaImageTexture(
       tflite::gpu::DataType::UINT8,  // GL_RGBA8
-      {tensor_width_, tensor_height_}, &small_mask_texture);
+      {tensor_width_, tensor_height_}, &small_mask_texture));
 
   // Get input previous mask.
   auto input_mask_texture = has_prev_mask
@@ -389,7 +398,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
                                 : mediapipe::GlTexture();
 
   // Copy input tensor.
-  tflite::gpu::gl::CopyBuffer(input_tensors[0], *tensor_buffer_);
+  RET_CHECK_CALL(CopyBuffer(input_tensors[0], *tensor_buffer_));
 
   // Run shader, process mask tensor.
   // Run softmax over tensor output and blend with previous mask.
@@ -397,18 +406,18 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
     const int output_index = 0;
     glBindImageTexture(output_index, small_mask_texture.id(), 0, GL_FALSE, 0,
                        GL_WRITE_ONLY, GL_RGBA8);
-    tensor_buffer_->BindToIndex(2);
+    RET_CHECK_CALL(tensor_buffer_->BindToIndex(2));
 
     const tflite::gpu::uint3 workgroups = {
         NumGroups(tensor_width_, kWorkgroupSize),
         NumGroups(tensor_height_, kWorkgroupSize), 1};
 
     if (!has_prev_mask) {
-      mask_program_no_prev_->Dispatch(workgroups);
+      RET_CHECK_CALL(mask_program_no_prev_->Dispatch(workgroups));
     } else {
       glActiveTexture(GL_TEXTURE1);
       glBindTexture(GL_TEXTURE_2D, input_mask_texture.name());
-      mask_program_with_prev_->Dispatch(workgroups);
+      RET_CHECK_CALL(mask_program_with_prev_->Dispatch(workgroups));
       glActiveTexture(GL_TEXTURE1);
       glBindTexture(GL_TEXTURE_2D, 0);
     }
@@ -438,13 +447,13 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   // Cleanup
   input_mask_texture.Release();
   output_texture.Release();
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 void TfLiteTensorsToSegmentationCalculator::GlRender() {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
   static const GLfloat square_vertices[] = {
       -1.0f, -1.0f,  // bottom left
       1.0f,  -1.0f,  // bottom right
@@ -492,7 +501,7 @@ void TfLiteTensorsToSegmentationCalculator::GlRender() {
   glBindVertexArray(0);
   glDeleteVertexArrays(1, &vao);
   glDeleteBuffers(2, vbo);
-#endif  // __ANDROID__
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 }
 
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::LoadOptions(
@@ -516,8 +525,9 @@ void TfLiteTensorsToSegmentationCalculator::GlRender() {
 
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::InitGpu(
     CalculatorContext* cc) {
-#if defined(__ANDROID__)
+#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
-
+  MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]()
+                                                    -> ::mediapipe::Status {
     // A shader to process a segmentation tensor into an output mask,
     // and use an optional previous mask as input.
     // Currently uses 4 channels for output,
@@ -598,49 +608,33 @@ void main() {
         options_.combine_with_previous_ratio(), "#define READ_PREVIOUS",
         options_.flip_vertically() ? "out_height - gid.y - 1" : "gid.y");
 
-  auto status = ::tflite::gpu::OkStatus();
-
     // Shader programs.
     GlShader shader_without_previous;
-  status = GlShader::CompileShader(GL_COMPUTE_SHADER, shader_src_no_previous,
+    RET_CHECK_CALL(GlShader::CompileShader(
-                                   &shader_without_previous);
+        GL_COMPUTE_SHADER, shader_src_no_previous, &shader_without_previous));
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     mask_program_no_prev_ = absl::make_unique<GlProgram>();
-  status = GlProgram::CreateWithShader(shader_without_previous,
+    RET_CHECK_CALL(GlProgram::CreateWithShader(shader_without_previous,
-                                       mask_program_no_prev_.get());
+                                               mask_program_no_prev_.get()));
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     GlShader shader_with_previous;
-  status = GlShader::CompileShader(GL_COMPUTE_SHADER, shader_src_with_previous,
+    RET_CHECK_CALL(GlShader::CompileShader(
-                                   &shader_with_previous);
+        GL_COMPUTE_SHADER, shader_src_with_previous, &shader_with_previous));
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
     mask_program_with_prev_ = absl::make_unique<GlProgram>();
-  status = GlProgram::CreateWithShader(shader_with_previous,
+    RET_CHECK_CALL(GlProgram::CreateWithShader(shader_with_previous,
-                                       mask_program_with_prev_.get());
+                                               mask_program_with_prev_.get()));
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
 
     // Buffer storage for input tensor.
     size_t tensor_length = tensor_width_ * tensor_height_ * tensor_channels_;
     tensor_buffer_ = absl::make_unique<GlBuffer>();
-  status = CreateReadWriteShaderStorageBuffer<float>(tensor_length,
+    RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
-                                                     tensor_buffer_.get());
+        tensor_length, tensor_buffer_.get()));
-  if (!status.ok()) {
-    return ::mediapipe::InternalError(status.error_message());
-  }
 
     // Parameters.
     glUseProgram(mask_program_with_prev_->id());
     glUniform2i(glGetUniformLocation(mask_program_with_prev_->id(), "out_size"),
                 tensor_width_, tensor_height_);
     glUniform1i(
-      glGetUniformLocation(mask_program_with_prev_->id(), "input_texture"), 1);
+        glGetUniformLocation(mask_program_with_prev_->id(), "input_texture"),
+        1);
     glUseProgram(mask_program_no_prev_->id());
     glUniform2i(glGetUniformLocation(mask_program_no_prev_->id(), "out_size"),
                 tensor_width_, tensor_height_);
@@ -684,15 +678,18 @@ void main() {
 )";
 
     // Program
-  mediapipe::GlhCreateProgram(mediapipe::kBasicVertexShader,
+    mediapipe::GlhCreateProgram(
-                              upsample_shader_base.c_str(), NUM_ATTRIBUTES,
+        mediapipe::kBasicVertexShader, upsample_shader_base.c_str(),
-                              &attr_name[0], attr_location, &upsample_program_);
+        NUM_ATTRIBUTES, &attr_name[0], attr_location, &upsample_program_);
     RET_CHECK(upsample_program_) << "Problem initializing the program.";
 
     // Parameters
     glUseProgram(upsample_program_);
     glUniform1i(glGetUniformLocation(upsample_program_, "input_data"), 1);
-#endif  // __ANDROID__
+
+    return ::mediapipe::OkStatus();
+  }));
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/tflite/util.h

@@ -0,0 +1,25 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_
+#define MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_
+
+#define RET_CHECK_CALL(call)                                     \
+  do {                                                           \
+    const auto status = (call);                                  \
+    if (ABSL_PREDICT_FALSE(!status.ok()))                        \
+      return ::mediapipe::InternalError(status.error_message()); \
+  } while (0);
+
+#endif  // MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_

mediapipe/calculators/util/BUILD

@@ -235,19 +235,13 @@ cc_library(
         "//mediapipe/framework/port:vector",
         "//mediapipe/util:annotation_renderer",
     ] + select({
-        "//mediapipe:android": [
+        "//mediapipe/gpu:disable_gpu": [],
+        "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
             "//mediapipe/gpu:gpu_buffer",
             "//mediapipe/gpu:shader_util",
         ],
-        "//mediapipe:ios": [
-            "//mediapipe/gpu:gl_calculator_helper",
-            "//mediapipe/gpu:gl_simple_shaders",
-            "//mediapipe/gpu:gpu_buffer",
-            "//mediapipe/gpu:shader_util",
-        ],
-        "//conditions:default": [],
     }),
     alwayslink = 1,
 )
@@ -694,3 +688,65 @@ cc_test(
         "//mediapipe/framework/tool:validate_type",
     ],
 )
+
+proto_library(
+    name = "top_k_scores_calculator_proto",
+    srcs = ["top_k_scores_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+    ],
+)
+
+mediapipe_cc_proto_library(
+    name = "top_k_scores_calculator_cc_proto",
+    srcs = ["top_k_scores_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":top_k_scores_calculator_proto"],
+)
+
+cc_library(
+    name = "top_k_scores_calculator",
+    srcs = ["top_k_scores_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":top_k_scores_calculator_cc_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/port:statusor",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/util:resource_util",
+    ] + select({
+        "//mediapipe:android": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:apple": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:macos": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+        "//conditions:default": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+    }),
+    alwayslink = 1,
+)
+
+cc_test(
+    name = "top_k_scores_calculator_test",
+    srcs = ["top_k_scores_calculator_test.cc"],
+    deps = [
+        ":top_k_scores_calculator",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework:packet",
+        "//mediapipe/framework/deps:message_matchers",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
+    ],
+)

mediapipe/calculators/util/annotation_overlay_calculator.cc

@@ -27,12 +27,12 @@
 #include "mediapipe/util/annotation_renderer.h"
 #include "mediapipe/util/color.pb.h"
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
 #include "mediapipe/gpu/gpu_buffer.h"
 #include "mediapipe/gpu/shader_util.h"
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
 namespace mediapipe {
 
@@ -146,13 +146,13 @@ class AnnotationOverlayCalculator : public CalculatorBase {
 
   bool use_gpu_ = false;
   bool gpu_initialized_ = false;
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   mediapipe::GlCalculatorHelper gpu_helper_;
   GLuint program_ = 0;
   GLuint image_mat_tex_ = 0;  // Overlay drawing image for GPU.
   int width_ = 0;
   int height_ = 0;
-#endif  // __ANDROID__ or iOS
+#endif  //  MEDIAPIPE_DISABLE_GPU
 };
 REGISTER_CALCULATOR(AnnotationOverlayCalculator);
 
@@ -160,6 +160,8 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
     CalculatorContract* cc) {
   CHECK_GE(cc->Inputs().NumEntries(), 1);
 
+  bool use_gpu = false;
+
   if (cc->Inputs().HasTag(kInputFrameTag) &&
       cc->Inputs().HasTag(kInputFrameTagGpu)) {
     return ::mediapipe::InternalError("Cannot have multiple input images.");
@@ -173,12 +175,13 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   int num_render_streams = cc->Inputs().NumEntries();
 
   // Input image to render onto copy of.
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag(kInputFrameTagGpu)) {
     cc->Inputs().Tag(kInputFrameTagGpu).Set<mediapipe::GpuBuffer>();
     num_render_streams = cc->Inputs().NumEntries() - 1;
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag(kInputFrameTag)) {
     cc->Inputs().Tag(kInputFrameTag).Set<ImageFrame>();
     num_render_streams = cc->Inputs().NumEntries() - 1;
@@ -190,18 +193,21 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   }
 
   // Rendered image.
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Outputs().HasTag(kOutputFrameTagGpu)) {
     cc->Outputs().Tag(kOutputFrameTagGpu).Set<mediapipe::GpuBuffer>();
+    use_gpu |= true;
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Outputs().HasTag(kOutputFrameTag)) {
     cc->Outputs().Tag(kOutputFrameTag).Set<ImageFrame>();
   }
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+  if (use_gpu) {
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -212,11 +218,11 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   options_ = cc->Options<AnnotationOverlayCalculatorOptions>();
   if (cc->Inputs().HasTag(kInputFrameTagGpu) &&
       cc->Outputs().HasTag(kOutputFrameTagGpu)) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     use_gpu_ = true;
 #else
-    RET_CHECK_FAIL() << "GPU processing is for Android and iOS only.";
+    RET_CHECK_FAIL() << "GPU processing not enabled.";
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   if (cc->Inputs().HasTag(kInputFrameTagGpu) ||
@@ -246,9 +252,9 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   }
 
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
   return ::mediapipe::OkStatus();
@@ -260,7 +266,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   std::unique_ptr<cv::Mat> image_mat;
   ImageFormat::Format target_format;
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     if (!gpu_initialized_) {
       MP_RETURN_IF_ERROR(
           gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
@@ -269,7 +275,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
           }));
       gpu_initialized_ = true;
     }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
     MP_RETURN_IF_ERROR(CreateRenderTargetGpu(cc, image_mat));
   } else {
     MP_RETURN_IF_ERROR(CreateRenderTargetCpu(cc, image_mat, &target_format));
@@ -288,7 +294,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   }
 
   if (use_gpu_) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
     // Overlay rendered image in OpenGL, onto a copy of input.
     uchar* image_mat_ptr = image_mat->data;
     MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
@@ -296,7 +302,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
           MP_RETURN_IF_ERROR(RenderToGpu(cc, image_mat_ptr));
           return ::mediapipe::OkStatus();
         }));
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
     // Copy the rendered image to output.
     uchar* image_mat_ptr = image_mat->data;
@@ -307,14 +313,14 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
 }
 
 ::mediapipe::Status AnnotationOverlayCalculator::Close(CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   gpu_helper_.RunInGlContext([this] {
     if (program_) glDeleteProgram(program_);
     program_ = 0;
     if (image_mat_tex_) glDeleteTextures(1, &image_mat_tex_);
     image_mat_tex_ = 0;
   });
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
@@ -325,7 +331,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   auto output_frame = absl::make_unique<ImageFrame>(
       target_format, renderer_->GetImageWidth(), renderer_->GetImageHeight());
 
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   output_frame->CopyPixelData(target_format, renderer_->GetImageWidth(),
                               renderer_->GetImageHeight(), data_image,
                               ImageFrame::kGlDefaultAlignmentBoundary);
@@ -333,7 +339,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   output_frame->CopyPixelData(target_format, renderer_->GetImageWidth(),
                               renderer_->GetImageHeight(), data_image,
                               ImageFrame::kDefaultAlignmentBoundary);
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   cc->Outputs()
       .Tag(kOutputFrameTag)
@@ -344,7 +350,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
 
 ::mediapipe::Status AnnotationOverlayCalculator::RenderToGpu(
     CalculatorContext* cc, uchar* overlay_image) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   // Source and destination textures.
   const auto& input_frame =
       cc->Inputs().Tag(kInputFrameTagGpu).Get<mediapipe::GpuBuffer>();
@@ -390,7 +396,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   // Cleanup
   input_texture.Release();
   output_texture.Release();
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
@@ -451,15 +457,16 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
 
 ::mediapipe::Status AnnotationOverlayCalculator::CreateRenderTargetGpu(
     CalculatorContext* cc, std::unique_ptr<cv::Mat>& image_mat) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   if (image_frame_available_) {
     const auto& input_frame =
         cc->Inputs().Tag(kInputFrameTagGpu).Get<mediapipe::GpuBuffer>();
 
     const mediapipe::ImageFormat::Format format =
         mediapipe::ImageFormatForGpuBufferFormat(input_frame.format());
-    if (format != mediapipe::ImageFormat::SRGBA)
+    if (format != mediapipe::ImageFormat::SRGBA &&
-      RET_CHECK_FAIL() << "Unsupported GPU input format.";
+        format != mediapipe::ImageFormat::SRGB)
+      RET_CHECK_FAIL() << "Unsupported GPU input format: " << format;
 
     image_mat = absl::make_unique<cv::Mat>(
         height_, width_, CV_8UC3,
@@ -471,14 +478,14 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
         cv::Scalar(options_.canvas_color().r(), options_.canvas_color().g(),
                    options_.canvas_color().b()));
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status AnnotationOverlayCalculator::GlRender(
     CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   static const GLfloat square_vertices[] = {
       -1.0f, -1.0f,  // bottom left
       1.0f,  -1.0f,  // bottom right
@@ -526,14 +533,14 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   glBindVertexArray(0);
   glDeleteVertexArrays(1, &vao);
   glDeleteBuffers(2, vbo);
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status AnnotationOverlayCalculator::GlSetup(
     CalculatorContext* cc) {
-#if defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
+#if !defined(MEDIAPIPE_DISABLE_GPU)
   const GLint attr_location[NUM_ATTRIBUTES] = {
       ATTRIB_VERTEX,
       ATTRIB_TEXTURE_POSITION,
@@ -609,7 +616,7 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
     glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
     glBindTexture(GL_TEXTURE_2D, 0);
   }
-#endif  // __ANDROID__ or iOS
+#endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/util/rect_to_render_data_calculator.cc

@@ -23,10 +23,25 @@ namespace mediapipe {
 
 namespace {
 
-constexpr char kNormalizedRectTag[] = "NORM_RECT";
+constexpr char kNormRectTag[] = "NORM_RECT";
 constexpr char kRectTag[] = "RECT";
+constexpr char kNormRectsTag[] = "NORM_RECTS";
+constexpr char kRectsTag[] = "RECTS";
 constexpr char kRenderDataTag[] = "RENDER_DATA";
 
+RenderAnnotation::Rectangle* NewRect(
+    const RectToRenderDataCalculatorOptions& options, RenderData* render_data) {
+  auto* annotation = render_data->add_render_annotations();
+  annotation->mutable_color()->set_r(options.color().r());
+  annotation->mutable_color()->set_g(options.color().g());
+  annotation->mutable_color()->set_b(options.color().b());
+  annotation->set_thickness(options.thickness());
+
+  return options.filled()
+             ? annotation->mutable_filled_rectangle()->mutable_rectangle()
+             : annotation->mutable_rectangle();
+}
+
 void SetRect(bool normalized, double xmin, double ymin, double width,
              double height, double rotation,
              RenderAnnotation::Rectangle* rect) {
@@ -51,6 +66,8 @@ void SetRect(bool normalized, double xmin, double ymin, double width,
 //   One of the following:
 //   NORM_RECT: A NormalizedRect
 //   RECT: A Rect
+//   NORM_RECTS: An std::vector<NormalizedRect>
+//   RECTS: An std::vector<Rect>
 //
 // Output:
 //   RENDER_DATA: A RenderData
@@ -83,16 +100,27 @@ REGISTER_CALCULATOR(RectToRenderDataCalculator);
 
 ::mediapipe::Status RectToRenderDataCalculator::GetContract(
     CalculatorContract* cc) {
-  RET_CHECK(cc->Inputs().HasTag(kNormalizedRectTag) ^
+  RET_CHECK_EQ((cc->Inputs().HasTag(kNormRectTag) ? 1 : 0) +
-            cc->Inputs().HasTag(kRectTag));
+                   (cc->Inputs().HasTag(kRectTag) ? 1 : 0) +
+                   (cc->Inputs().HasTag(kNormRectsTag) ? 1 : 0) +
+                   (cc->Inputs().HasTag(kRectsTag) ? 1 : 0),
+               1)
+      << "Exactly one of NORM_RECT, RECT, NORM_RECTS or RECTS input stream "
+         "should be provided.";
   RET_CHECK(cc->Outputs().HasTag(kRenderDataTag));
 
-  if (cc->Inputs().HasTag(kNormalizedRectTag)) {
+  if (cc->Inputs().HasTag(kNormRectTag)) {
-    cc->Inputs().Tag(kNormalizedRectTag).Set<NormalizedRect>();
+    cc->Inputs().Tag(kNormRectTag).Set<NormalizedRect>();
   }
   if (cc->Inputs().HasTag(kRectTag)) {
     cc->Inputs().Tag(kRectTag).Set<Rect>();
   }
+  if (cc->Inputs().HasTag(kNormRectsTag)) {
+    cc->Inputs().Tag(kNormRectsTag).Set<std::vector<NormalizedRect>>();
+  }
+  if (cc->Inputs().HasTag(kRectsTag)) {
+    cc->Inputs().Tag(kRectsTag).Set<std::vector<Rect>>();
+  }
   cc->Outputs().Tag(kRenderDataTag).Set<RenderData>();
 
   return ::mediapipe::OkStatus();
@@ -108,31 +136,43 @@ REGISTER_CALCULATOR(RectToRenderDataCalculator);
 
 ::mediapipe::Status RectToRenderDataCalculator::Process(CalculatorContext* cc) {
   auto render_data = absl::make_unique<RenderData>();
-  auto* annotation = render_data->add_render_annotations();
-  annotation->mutable_color()->set_r(options_.color().r());
-  annotation->mutable_color()->set_g(options_.color().g());
-  annotation->mutable_color()->set_b(options_.color().b());
-  annotation->set_thickness(options_.thickness());
 
-  auto* rectangle =
+  if (cc->Inputs().HasTag(kNormRectTag) &&
-      options_.filled()
+      !cc->Inputs().Tag(kNormRectTag).IsEmpty()) {
-          ? annotation->mutable_filled_rectangle()->mutable_rectangle()
+    const auto& rect = cc->Inputs().Tag(kNormRectTag).Get<NormalizedRect>();
-          : annotation->mutable_rectangle();
+    auto* rectangle = NewRect(options_, render_data.get());
-
-  if (cc->Inputs().HasTag(kNormalizedRectTag) &&
-      !cc->Inputs().Tag(kNormalizedRectTag).IsEmpty()) {
-    const auto& rect =
-        cc->Inputs().Tag(kNormalizedRectTag).Get<NormalizedRect>();
     SetRect(/*normalized=*/true, rect.x_center() - rect.width() / 2.f,
             rect.y_center() - rect.height() / 2.f, rect.width(), rect.height(),
             rect.rotation(), rectangle);
   }
   if (cc->Inputs().HasTag(kRectTag) && !cc->Inputs().Tag(kRectTag).IsEmpty()) {
     const auto& rect = cc->Inputs().Tag(kRectTag).Get<Rect>();
+    auto* rectangle = NewRect(options_, render_data.get());
     SetRect(/*normalized=*/false, rect.x_center() - rect.width() / 2.f,
             rect.y_center() - rect.height() / 2.f, rect.width(), rect.height(),
             rect.rotation(), rectangle);
   }
+  if (cc->Inputs().HasTag(kNormRectsTag) &&
+      !cc->Inputs().Tag(kNormRectsTag).IsEmpty()) {
+    const auto& rects =
+        cc->Inputs().Tag(kNormRectsTag).Get<std::vector<NormalizedRect>>();
+    for (auto& rect : rects) {
+      auto* rectangle = NewRect(options_, render_data.get());
+      SetRect(/*normalized=*/true, rect.x_center() - rect.width() / 2.f,
+              rect.y_center() - rect.height() / 2.f, rect.width(),
+              rect.height(), rect.rotation(), rectangle);
+    }
+  }
+  if (cc->Inputs().HasTag(kRectsTag) &&
+      !cc->Inputs().Tag(kRectsTag).IsEmpty()) {
+    const auto& rects = cc->Inputs().Tag(kRectsTag).Get<std::vector<Rect>>();
+    for (auto& rect : rects) {
+      auto* rectangle = NewRect(options_, render_data.get());
+      SetRect(/*normalized=*/false, rect.x_center() - rect.width() / 2.f,
+              rect.y_center() - rect.height() / 2.f, rect.width(),
+              rect.height(), rect.rotation(), rectangle);
+    }
+  }
 
   cc->Outputs()
       .Tag(kRenderDataTag)

mediapipe/calculators/util/top_k_scores_calculator.cc

@@ -0,0 +1,194 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <algorithm>
+#include <istream>
+#include <iterator>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "mediapipe/calculators/util/top_k_scores_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/statusor.h"
+#include "mediapipe/util/resource_util.h"
+
+#if defined(MEDIAPIPE_LITE) || defined(__ANDROID__) || \
+    (defined(__APPLE__) && !TARGET_OS_OSX)
+#include "mediapipe/util/android/file/base/file.h"
+#include "mediapipe/util/android/file/base/helpers.h"
+#else
+#include "mediapipe/framework/port/file_helpers.h"
+#endif
+
+namespace mediapipe {
+// A calculator that takes a vector of scores and returns the indexes, scores,
+// labels of the top k elements.
+//
+// Usage example:
+// node {
+//   calculator: "TopKScoresCalculator"
+//   input_stream: "SCORES:score_vector"
+//   output_stream: "TOP_K_INDEXES:top_k_indexes"
+//   output_stream: "TOP_K_SCORES:top_k_scores"
+//   output_stream: "TOP_K_LABELS:top_k_labels"
+//   options: {
+//     [mediapipe.TopKScoresCalculatorOptions.ext] {
+//       top_k: 5
+//       threshold: 0.1
+//       label_map_path: "/path/to/label/map"
+//     }
+//   }
+// }
+class TopKScoresCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  ::mediapipe::Status LoadLabelmap(std::string label_map_path);
+
+  int top_k_ = -1;
+  float threshold_ = 0.0;
+  std::unordered_map<int, std::string> label_map_;
+};
+REGISTER_CALCULATOR(TopKScoresCalculator);
+
+::mediapipe::Status TopKScoresCalculator::GetContract(CalculatorContract* cc) {
+  RET_CHECK(cc->Inputs().HasTag("SCORES"));
+  cc->Inputs().Tag("SCORES").Set<std::vector<float>>();
+  if (cc->Outputs().HasTag("TOP_K_INDEXES")) {
+    cc->Outputs().Tag("TOP_K_INDEXES").Set<std::vector<int>>();
+  }
+  if (cc->Outputs().HasTag("TOP_K_SCORES")) {
+    cc->Outputs().Tag("TOP_K_SCORES").Set<std::vector<float>>();
+  }
+  if (cc->Outputs().HasTag("TOP_K_LABELS")) {
+    cc->Outputs().Tag("TOP_K_LABELS").Set<std::vector<std::string>>();
+  }
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status TopKScoresCalculator::Open(CalculatorContext* cc) {
+  const auto& options = cc->Options<::mediapipe::TopKScoresCalculatorOptions>();
+  RET_CHECK(options.has_top_k() || options.has_threshold())
+      << "Must specify at least one of the top_k and threshold fields in "
+         "TopKScoresCalculatorOptions.";
+  if (options.has_top_k()) {
+    RET_CHECK(options.top_k() > 0) << "top_k must be greater than zero.";
+    top_k_ = options.top_k();
+  }
+  if (options.has_threshold()) {
+    threshold_ = options.threshold();
+  }
+  if (options.has_label_map_path()) {
+    MP_RETURN_IF_ERROR(LoadLabelmap(options.label_map_path()));
+  }
+  if (cc->Outputs().HasTag("TOP_K_LABELS")) {
+    RET_CHECK(!label_map_.empty());
+  }
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status TopKScoresCalculator::Process(CalculatorContext* cc) {
+  const std::vector<float>& input_vector =
+      cc->Inputs().Tag("SCORES").Get<std::vector<float>>();
+  std::vector<int> top_k_indexes;
+
+  std::vector<float> top_k_scores;
+
+  std::vector<std::string> top_k_labels;
+
+  if (top_k_ > 0) {
+    top_k_indexes.reserve(top_k_);
+    top_k_scores.reserve(top_k_);
+    top_k_labels.reserve(top_k_);
+  }
+  std::priority_queue<std::pair<float, int>, std::vector<std::pair<float, int>>,
+                      std::greater<std::pair<float, int>>>
+      pq;
+  for (int i = 0; i < input_vector.size(); ++i) {
+    if (input_vector[i] < threshold_) {
+      continue;
+    }
+    if (top_k_ > 0) {
+      if (pq.size() < top_k_) {
+        pq.push(std::pair<float, int>(input_vector[i], i));
+      } else if (pq.top().first < input_vector[i]) {
+        pq.pop();
+        pq.push(std::pair<float, int>(input_vector[i], i));
+      }
+    } else {
+      pq.push(std::pair<float, int>(input_vector[i], i));
+    }
+  }
+
+  while (!pq.empty()) {
+    top_k_indexes.push_back(pq.top().second);
+    top_k_scores.push_back(pq.top().first);
+    pq.pop();
+  }
+  reverse(top_k_indexes.begin(), top_k_indexes.end());
+  reverse(top_k_scores.begin(), top_k_scores.end());
+
+  if (cc->Outputs().HasTag("TOP_K_LABELS")) {
+    for (int index : top_k_indexes) {
+      top_k_labels.push_back(label_map_[index]);
+    }
+  }
+  if (cc->Outputs().HasTag("TOP_K_INDEXES")) {
+    cc->Outputs()
+        .Tag("TOP_K_INDEXES")
+        .AddPacket(MakePacket<std::vector<int>>(top_k_indexes)
+                       .At(cc->InputTimestamp()));
+  }
+  if (cc->Outputs().HasTag("TOP_K_SCORES")) {
+    cc->Outputs()
+        .Tag("TOP_K_SCORES")
+        .AddPacket(MakePacket<std::vector<float>>(top_k_scores)
+                       .At(cc->InputTimestamp()));
+  }
+  if (cc->Outputs().HasTag("TOP_K_LABELS")) {
+    cc->Outputs()
+        .Tag("TOP_K_LABELS")
+        .AddPacket(MakePacket<std::vector<std::string>>(top_k_labels)
+                       .At(cc->InputTimestamp()));
+  }
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status TopKScoresCalculator::LoadLabelmap(
+    std::string label_map_path) {
+  std::string string_path;
+  ASSIGN_OR_RETURN(string_path, PathToResourceAsFile(label_map_path));
+  std::string label_map_string;
+  MP_RETURN_IF_ERROR(file::GetContents(string_path, &label_map_string));
+
+  std::istringstream stream(label_map_string);
+  std::string line;
+  int i = 0;
+  while (std::getline(stream, line)) {
+    label_map_[i++] = line;
+  }
+  return ::mediapipe::OkStatus();
+}
+
+}  // namespace mediapipe

mediapipe/calculators/util/top_k_scores_calculator.proto

@@ -0,0 +1,33 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+
+message TopKScoresCalculatorOptions {
+  extend CalculatorOptions {
+    optional TopKScoresCalculatorOptions ext = 271211788;
+  }
+  // How many highest scoring packets to output.
+  optional int32 top_k = 1;
+
+  // If set, only keep the scores that are greater than the threshold.
+  optional float threshold = 2;
+
+  // Path to a label map file for getting the actual name of classes.
+  optional string label_map_path = 3;
+}

mediapipe/calculators/util/top_k_scores_calculator_test.cc

@@ -0,0 +1,150 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <vector>
+
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/status_matchers.h"
+
+namespace mediapipe {
+
+TEST(TopKScoresCalculatorTest, TestNodeConfig) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "TopKScoresCalculator"
+    input_stream: "SCORES:score_vector"
+    output_stream: "TOP_K_INDEXES:top_k_indexes"
+    output_stream: "TOP_K_SCORES:top_k_scores"
+    options: {
+      [mediapipe.TopKScoresCalculatorOptions.ext] {}
+    }
+  )"));
+
+  auto status = runner.Run();
+  ASSERT_TRUE(!status.ok());
+  EXPECT_THAT(
+      status.ToString(),
+      testing::HasSubstr(
+          "Must specify at least one of the top_k and threshold fields"));
+}
+
+TEST(TopKScoresCalculatorTest, TestTopKOnly) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "TopKScoresCalculator"
+    input_stream: "SCORES:score_vector"
+    output_stream: "TOP_K_INDEXES:top_k_indexes"
+    output_stream: "TOP_K_SCORES:top_k_scores"
+    options: {
+      [mediapipe.TopKScoresCalculatorOptions.ext] { top_k: 2 }
+    }
+  )"));
+
+  std::vector<float> score_vector{0.9, 0.2, 0.3, 1.0, 0.1};
+
+  runner.MutableInputs()->Tag("SCORES").packets.push_back(
+      MakePacket<std::vector<float>>(score_vector).At(Timestamp(0)));
+
+  MP_ASSERT_OK(runner.Run());
+  const std::vector<Packet>& indexes_outputs =
+      runner.Outputs().Tag("TOP_K_INDEXES").packets;
+  ASSERT_EQ(1, indexes_outputs.size());
+  const auto& indexes = indexes_outputs[0].Get<std::vector<int>>();
+  EXPECT_EQ(2, indexes.size());
+  EXPECT_EQ(3, indexes[0]);
+  EXPECT_EQ(0, indexes[1]);
+  const std::vector<Packet>& scores_outputs =
+      runner.Outputs().Tag("TOP_K_SCORES").packets;
+  ASSERT_EQ(1, scores_outputs.size());
+  const auto& scores = scores_outputs[0].Get<std::vector<float>>();
+  EXPECT_EQ(2, scores.size());
+  EXPECT_NEAR(1, scores[0], 1e-5);
+  EXPECT_NEAR(0.9, scores[1], 1e-5);
+}
+
+TEST(TopKScoresCalculatorTest, TestThresholdOnly) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "TopKScoresCalculator"
+    input_stream: "SCORES:score_vector"
+    output_stream: "TOP_K_INDEXES:top_k_indexes"
+    output_stream: "TOP_K_SCORES:top_k_scores"
+    options: {
+      [mediapipe.TopKScoresCalculatorOptions.ext] { threshold: 0.2 }
+    }
+  )"));
+
+  std::vector<float> score_vector{0.9, 0.2, 0.3, 1.0, 0.1};
+
+  runner.MutableInputs()->Tag("SCORES").packets.push_back(
+      MakePacket<std::vector<float>>(score_vector).At(Timestamp(0)));
+
+  MP_ASSERT_OK(runner.Run());
+  const std::vector<Packet>& indexes_outputs =
+      runner.Outputs().Tag("TOP_K_INDEXES").packets;
+  ASSERT_EQ(1, indexes_outputs.size());
+  const auto& indexes = indexes_outputs[0].Get<std::vector<int>>();
+  EXPECT_EQ(4, indexes.size());
+  EXPECT_EQ(3, indexes[0]);
+  EXPECT_EQ(0, indexes[1]);
+  EXPECT_EQ(2, indexes[2]);
+  EXPECT_EQ(1, indexes[3]);
+  const std::vector<Packet>& scores_outputs =
+      runner.Outputs().Tag("TOP_K_SCORES").packets;
+  ASSERT_EQ(1, scores_outputs.size());
+  const auto& scores = scores_outputs[0].Get<std::vector<float>>();
+  EXPECT_EQ(4, scores.size());
+  EXPECT_NEAR(1.0, scores[0], 1e-5);
+  EXPECT_NEAR(0.9, scores[1], 1e-5);
+  EXPECT_NEAR(0.3, scores[2], 1e-5);
+  EXPECT_NEAR(0.2, scores[3], 1e-5);
+}
+
+TEST(TopKScoresCalculatorTest, TestBothTopKAndThreshold) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "TopKScoresCalculator"
+    input_stream: "SCORES:score_vector"
+    output_stream: "TOP_K_INDEXES:top_k_indexes"
+    output_stream: "TOP_K_SCORES:top_k_scores"
+    options: {
+      [mediapipe.TopKScoresCalculatorOptions.ext] { top_k: 4 threshold: 0.3 }
+    }
+  )"));
+
+  std::vector<float> score_vector{0.9, 0.2, 0.3, 1.0, 0.1};
+
+  runner.MutableInputs()->Tag("SCORES").packets.push_back(
+      MakePacket<std::vector<float>>(score_vector).At(Timestamp(0)));
+
+  MP_ASSERT_OK(runner.Run());
+  const std::vector<Packet>& indexes_outputs =
+      runner.Outputs().Tag("TOP_K_INDEXES").packets;
+  ASSERT_EQ(1, indexes_outputs.size());
+  const auto& indexes = indexes_outputs[0].Get<std::vector<int>>();
+  EXPECT_EQ(3, indexes.size());
+  EXPECT_EQ(3, indexes[0]);
+  EXPECT_EQ(0, indexes[1]);
+  EXPECT_EQ(2, indexes[2]);
+  const std::vector<Packet>& scores_outputs =
+      runner.Outputs().Tag("TOP_K_SCORES").packets;
+  ASSERT_EQ(1, scores_outputs.size());
+  const auto& scores = scores_outputs[0].Get<std::vector<float>>();
+  EXPECT_EQ(3, scores.size());
+  EXPECT_NEAR(1.0, scores[0], 1e-5);
+  EXPECT_NEAR(0.9, scores[1], 1e-5);
+  EXPECT_NEAR(0.3, scores[2], 1e-5);
+}
+
+}  // namespace mediapipe

mediapipe/docs/examples.md

@@ -51,6 +51,15 @@ and model details are described in the
 *   [Android](./face_detection_mobile_gpu.md)
 *   [iOS](./face_detection_mobile_gpu.md)
 
+### Face Detection with CPU
+
+[Face Detection with CPU](./face_detection_mobile_cpu.md) illustrates using the
+same TFLite model in a CPU-based pipeline. This example highlights how graphs
+can be easily adapted to run on CPU v.s. GPU.
+
+*   [Android](./face_detection_mobile_cpu.md)
+*   [iOS](./face_detection_mobile_cpu.md)
+
 ### Hand Detection with GPU
 
 [Hand Detection with GPU](./hand_detection_mobile_gpu.md) illustrates how to use
@@ -103,3 +112,29 @@ object detection models (TensorFlow and TFLite) using the MediaPipe C++ APIs.
 
 [Sobel edge detection]:https://en.wikipedia.org/wiki/Sobel_operator
 [CameraX]:https://developer.android.com/training/camerax
+
+### Face Detection on Desktop with Webcam
+
+[Face Detection on Desktop with Webcam](./face_detection_desktop.md) shows how
+to use MediaPipe with a TFLite model for face detection on desktop using CPU or
+GPU with live video from a webcam.
+
+*   [Desktop GPU](./face_detection_desktop.md)
+*   [Desktop CPU](./face_detection_desktop.md)
+
+### Hand Tracking on Desktop with Webcam
+
+[Hand Tracking on Desktop with Webcam](./hand_tracking_desktop.md) shows how to
+use MediaPipe with a TFLite model for hand tracking on desktop using CPU or GPU
+with live video from a webcam.
+
+*   [Desktop GPU](./hand_tracking_desktop.md)
+*   [Desktop CPU](./hand_tracking_desktop.md)
+
+### Hair Segmentation on Desktop with Webcam
+
+[Hair Segmentation on Desktop with Webcam](./hair_segmentation_desktop.md) shows
+how to use MediaPipe with a TFLite model for hair segmentation on desktop using
+GPU with live video from a webcam.
+
+*   [Desktop GPU](./hair_segmentation_desktop.md)

mediapipe/docs/face_detection_desktop.md

@@ -0,0 +1,265 @@
+## Face Detection on Desktop
+
+This is an example of using MediaPipe to run face detection models (TensorFlow
+Lite) and render bounding boxes on the detected faces. To know more about the
+face detection models, please refer to the model [`README file`]. Moreover, if
+you are interested in running the same TensorfFlow Lite model on Android/iOS,
+please see the
+[Face Detection on GPU on Android/iOS](face_detection_mobile_gpu.md) and
+[Face Detection on CPU on Android/iOS](face_detection_mobile_cpu.md) examples.
+
+We show the face detection demos with TensorFlow Lite model using the Webcam:
+
+-   [TensorFlow Lite Face Detection Demo with Webcam (CPU)](#tensorflow-lite-face-detection-demo-with-webcam-cpu)
+
+-   [TensorFlow Lite Face Detection Demo with Webcam (GPU)](#tensorflow-lite-face-detection-demo-with-webcam-gpu)
+
+Note: Desktop GPU works only on Linux. Mesa drivers need to be installed. Please
+see
+[step 4 of "Installing on Debian and Ubuntu" in the installation guide](./install.md).
+
+Note: If MediaPipe depends on OpenCV 2, please see the [known issues with OpenCV 2](#known-issues-with-opencv-2) section.
+
+### TensorFlow Lite Face Detection Demo with Webcam (CPU)
+
+To build and run the TensorFlow Lite example on desktop (CPU) with Webcam, run:
+
+```bash
+# Video from webcam running on desktop CPU
+$ bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 \
+    mediapipe/examples/desktop/face_detection:face_detection_cpu
+
+# It should print:
+# Target //mediapipe/examples/desktop/face_detection:face_detection_cpu up-to-date:
+#   bazel-bin/mediapipe/examples/desktop/face_detection/face_detection_cpu
+# INFO: Elapsed time: 36.417s, Critical Path: 23.22s
+# INFO: 711 processes: 710 linux-sandbox, 1 local.
+# INFO: Build completed successfully, 734 total actions
+
+$ export GLOG_logtostderr=1
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible
+$ bazel-bin/mediapipe/examples/desktop/face_detection/face_detection_cpu \
+    --calculator_graph_config_file=mediapipe/graphs/face_detection/face_detection_desktop_live.pbtxt
+```
+
+### TensorFlow Lite Face Detection Demo with Webcam (GPU)
+
+To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
+
+```bash
+# Video from webcam running on desktop GPU
+# This works only for linux currently
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+    mediapipe/examples/desktop/face_detection:face_detection_gpu
+
+# It should print:
+# Target //mediapipe/examples/desktop/face_detection:face_detection_gpu up-to-date:
+#   bazel-bin/mediapipe/examples/desktop/face_detection/face_detection_gpu
+# INFO: Elapsed time: 36.417s, Critical Path: 23.22s
+# INFO: 711 processes: 710 linux-sandbox, 1 local.
+# INFO: Build completed successfully, 734 total actions
+
+$ export GLOG_logtostderr=1
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible,
+# or GPU drivers not setup properly.
+$ bazel-bin/mediapipe/examples/desktop/face_detection/face_detection_gpu \
+    --calculator_graph_config_file=mediapipe/graphs/face_detection/face_detection_mobile_gpu.pbtxt
+```
+
+#### Graph
+
+![graph visualization](images/face_detection_desktop.png)
+
+To visualize the graph as shown above, copy the text specification of the graph
+below and paste it into
+[MediaPipe Visualizer](https://viz.mediapipe.dev).
+
+```bash
+# MediaPipe graph that performs face detection with TensorFlow Lite on CPU & GPU.
+# Used in the examples in
+# mediapipie/examples/desktop/face_detection:face_detection_cpu.
+
+# Images on CPU coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Throttles the images flowing downstream for flow control. It passes through
+# the very first incoming image unaltered, and waits for
+# TfLiteTensorsToDetectionsCalculator downstream in the graph to finish
+# generating the corresponding detections before it passes through another
+# image. All images that come in while waiting are dropped, limiting the number
+# of in-flight images between this calculator and
+# TfLiteTensorsToDetectionsCalculator to 1. This prevents the nodes in between
+# from queuing up incoming images and data excessively, which leads to increased
+# latency and memory usage, unwanted in real-time mobile applications. It also
+# eliminates unnecessarily computation, e.g., a transformed image produced by
+# ImageTransformationCalculator may get dropped downstream if the subsequent
+# TfLiteConverterCalculator or TfLiteInferenceCalculator is still busy
+# processing previous inputs.
+node {
+  calculator: "FlowLimiterCalculator"
+  input_stream: "input_video"
+  input_stream: "FINISHED:detections"
+  input_stream_info: {
+    tag_index: "FINISHED"
+    back_edge: true
+  }
+  output_stream: "throttled_input_video"
+}
+
+# Transforms the input image on CPU to a 128x128 image. To scale the input
+# image, the scale_mode option is set to FIT to preserve the aspect ratio,
+# resulting in potential letterboxing in the transformed image.
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE:throttled_input_video"
+  output_stream: "IMAGE:transformed_input_video_cpu"
+  output_stream: "LETTERBOX_PADDING:letterbox_padding"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 128
+      output_height: 128
+      scale_mode: FIT
+    }
+  }
+}
+
+# Converts the transformed input image on CPU into an image tensor stored as a
+# TfLiteTensor.
+node {
+  calculator: "TfLiteConverterCalculator"
+  input_stream: "IMAGE:transformed_input_video_cpu"
+  output_stream: "TENSORS:image_tensor"
+}
+
+# Runs a TensorFlow Lite model on CPU that takes an image tensor and outputs a
+# vector of tensors representing, for instance, detection boxes/keypoints and
+# scores.
+node {
+  calculator: "TfLiteInferenceCalculator"
+  input_stream: "TENSORS:image_tensor"
+  output_stream: "TENSORS:detection_tensors"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "mediapipe/models/face_detection_front.tflite"
+    }
+  }
+}
+
+# Generates a single side packet containing a vector of SSD anchors based on
+# the specification in the options.
+node {
+  calculator: "SsdAnchorsCalculator"
+  output_side_packet: "anchors"
+  node_options: {
+    [type.googleapis.com/mediapipe.SsdAnchorsCalculatorOptions] {
+      num_layers: 4
+      min_scale: 0.1484375
+      max_scale: 0.75
+      input_size_height: 128
+      input_size_width: 128
+      anchor_offset_x: 0.5
+      anchor_offset_y: 0.5
+      strides: 8
+      strides: 16
+      strides: 16
+      strides: 16
+      aspect_ratios: 1.0
+      fixed_anchor_size: true
+    }
+  }
+}
+
+# Decodes the detection tensors generated by the TensorFlow Lite model, based on
+# the SSD anchors and the specification in the options, into a vector of
+# detections. Each detection describes a detected object.
+node {
+  calculator: "TfLiteTensorsToDetectionsCalculator"
+  input_stream: "TENSORS:detection_tensors"
+  input_side_packet: "ANCHORS:anchors"
+  output_stream: "DETECTIONS:detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToDetectionsCalculatorOptions] {
+      num_classes: 1
+      num_boxes: 896
+      num_coords: 16
+      box_coord_offset: 0
+      keypoint_coord_offset: 4
+      num_keypoints: 6
+      num_values_per_keypoint: 2
+      sigmoid_score: true
+      score_clipping_thresh: 100.0
+      reverse_output_order: true
+      x_scale: 128.0
+      y_scale: 128.0
+      h_scale: 128.0
+      w_scale: 128.0
+      min_score_thresh: 0.75
+    }
+  }
+}
+
+# Performs non-max suppression to remove excessive detections.
+node {
+  calculator: "NonMaxSuppressionCalculator"
+  input_stream: "detections"
+  output_stream: "filtered_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.NonMaxSuppressionCalculatorOptions] {
+      min_suppression_threshold: 0.3
+      overlap_type: INTERSECTION_OVER_UNION
+      algorithm: WEIGHTED
+      return_empty_detections: true
+    }
+  }
+}
+
+# Maps detection label IDs to the corresponding label text ("Face"). The label
+# map is provided in the label_map_path option.
+node {
+  calculator: "DetectionLabelIdToTextCalculator"
+  input_stream: "filtered_detections"
+  output_stream: "labeled_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
+      label_map_path: "mediapipe/models/face_detection_front_labelmap.txt"
+    }
+  }
+}
+
+# Adjusts detection locations (already normalized to [0.f, 1.f]) on the
+# letterboxed image (after image transformation with the FIT scale mode) to the
+# corresponding locations on the same image with the letterbox removed (the
+# input image to the graph before image transformation).
+node {
+  calculator: "DetectionLetterboxRemovalCalculator"
+  input_stream: "DETECTIONS:labeled_detections"
+  input_stream: "LETTERBOX_PADDING:letterbox_padding"
+  output_stream: "DETECTIONS:output_detections"
+}
+
+# Converts the detections to drawing primitives for annotation overlay.
+node {
+  calculator: "DetectionsToRenderDataCalculator"
+  input_stream: "DETECTIONS:output_detections"
+  output_stream: "RENDER_DATA:render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
+      thickness: 4.0
+      color { r: 255 g: 0 b: 0 }
+    }
+  }
+}
+
+# Draws annotations and overlays them on top of the input images.
+node {
+  calculator: "AnnotationOverlayCalculator"
+  input_stream: "INPUT_FRAME:throttled_input_video"
+  input_stream: "render_data"
+  output_stream: "OUTPUT_FRAME:output_video"
+}
+```
+
+[`README file`]:https://github.com/google/mediapipe/tree/master/mediapipe/models/object_detection_saved_model/README.md

mediapipe/docs/face_detection_mobile_cpu.md

@@ -0,0 +1,244 @@
+# Face Detection (CPU)
+
+This doc focuses on the
+[example graph](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/face_detection/face_detection_mobile_cpu.pbtxt)
+that performs face detection with TensorFlow Lite on CPU.
+
+## Android
+
+[Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/android/src/java/com/google/mediapipe/apps/facedetectioncpu)
+
+To build and install the app:
+
+```bash
+bazel build -c opt --config=android_arm64 mediapipe/examples/android/src/java/com/google/mediapipe/apps/facedetectioncpu
+adb install bazel-bin/mediapipe/examples/android/src/java/com/google/mediapipe/apps/facedetectioncpu/facedetectioncpu.apk
+```
+
+## iOS
+
+[Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/ios/facedetectioncpu).
+
+See the general [instructions](./mediapipe_ios_setup.md) for building iOS
+examples and generating an Xcode project. This will be the FaceDetectionCpuApp
+target.
+
+To build on the command line:
+
+```bash
+bazel build -c opt --config=ios_arm64 mediapipe/examples/ios/facedetectioncpu:FaceDetectionCpuApp
+```
+
+## Graph
+
+![face_detection_mobile_cpu_graph](images/mobile/face_detection_mobile_cpu.png)
+
+To visualize the graph as shown above, copy the text specification of the graph
+below and paste it into [MediaPipe Visualizer](https://viz.mediapipe.dev/).
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/face_detection/face_detection_mobile_cpu.pbtxt)
+
+```bash
+# MediaPipe graph that performs face detection with TensorFlow Lite on CPU.
+# Used in the examples in
+# mediapipie/examples/android/src/java/com/mediapipe/apps/facedetectioncpu and
+# mediapipie/examples/ios/facedetectioncpu.
+
+# Images on GPU coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Throttles the images flowing downstream for flow control. It passes through
+# the very first incoming image unaltered, and waits for
+# TfLiteTensorsToDetectionsCalculator downstream in the graph to finish
+# generating the corresponding detections before it passes through another
+# image. All images that come in while waiting are dropped, limiting the number
+# of in-flight images between this calculator and
+# TfLiteTensorsToDetectionsCalculator to 1. This prevents the nodes in between
+# from queuing up incoming images and data excessively, which leads to increased
+# latency and memory usage, unwanted in real-time mobile applications. It also
+# eliminates unnecessarily computation, e.g., a transformed image produced by
+# ImageTransformationCalculator may get dropped downstream if the subsequent
+# TfLiteConverterCalculator or TfLiteInferenceCalculator is still busy
+# processing previous inputs.
+node {
+  calculator: "FlowLimiterCalculator"
+  input_stream: "input_video"
+  input_stream: "FINISHED:detections"
+  input_stream_info: {
+    tag_index: "FINISHED"
+    back_edge: true
+  }
+  output_stream: "throttled_input_video"
+}
+
+# Transfers the input image from GPU to CPU memory for the purpose of
+# demonstrating a CPU-based pipeline. Note that the input image on GPU has the
+# origin defined at the bottom-left corner (OpenGL convention). As a result,
+# the transferred image on CPU also shares the same representation.
+node: {
+  calculator: "GpuBufferToImageFrameCalculator"
+  input_stream: "throttled_input_video"
+  output_stream: "input_video_cpu"
+}
+
+# Transforms the input image on CPU to a 128x128 image. To scale the input
+# image, the scale_mode option is set to FIT to preserve the aspect ratio,
+# resulting in potential letterboxing in the transformed image.
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE:input_video_cpu"
+  output_stream: "IMAGE:transformed_input_video_cpu"
+  output_stream: "LETTERBOX_PADDING:letterbox_padding"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 128
+      output_height: 128
+      scale_mode: FIT
+    }
+  }
+}
+
+# Converts the transformed input image on CPU into an image tensor stored as a
+# TfLiteTensor.
+node {
+  calculator: "TfLiteConverterCalculator"
+  input_stream: "IMAGE:transformed_input_video_cpu"
+  output_stream: "TENSORS:image_tensor"
+}
+
+# Runs a TensorFlow Lite model on CPU that takes an image tensor and outputs a
+# vector of tensors representing, for instance, detection boxes/keypoints and
+# scores.
+node {
+  calculator: "TfLiteInferenceCalculator"
+  input_stream: "TENSORS:image_tensor"
+  output_stream: "TENSORS:detection_tensors"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "mediapipe/models/face_detection_front.tflite"
+    }
+  }
+}
+
+# Generates a single side packet containing a vector of SSD anchors based on
+# the specification in the options.
+node {
+  calculator: "SsdAnchorsCalculator"
+  output_side_packet: "anchors"
+  node_options: {
+    [type.googleapis.com/mediapipe.SsdAnchorsCalculatorOptions] {
+      num_layers: 4
+      min_scale: 0.1484375
+      max_scale: 0.75
+      input_size_height: 128
+      input_size_width: 128
+      anchor_offset_x: 0.5
+      anchor_offset_y: 0.5
+      strides: 8
+      strides: 16
+      strides: 16
+      strides: 16
+      aspect_ratios: 1.0
+      fixed_anchor_size: true
+    }
+  }
+}
+
+# Decodes the detection tensors generated by the TensorFlow Lite model, based on
+# the SSD anchors and the specification in the options, into a vector of
+# detections. Each detection describes a detected object.
+node {
+  calculator: "TfLiteTensorsToDetectionsCalculator"
+  input_stream: "TENSORS:detection_tensors"
+  input_side_packet: "ANCHORS:anchors"
+  output_stream: "DETECTIONS:detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToDetectionsCalculatorOptions] {
+      num_classes: 1
+      num_boxes: 896
+      num_coords: 16
+      box_coord_offset: 0
+      keypoint_coord_offset: 4
+      num_keypoints: 6
+      num_values_per_keypoint: 2
+      sigmoid_score: true
+      score_clipping_thresh: 100.0
+      reverse_output_order: true
+      x_scale: 128.0
+      y_scale: 128.0
+      h_scale: 128.0
+      w_scale: 128.0
+      min_score_thresh: 0.75
+    }
+  }
+}
+
+# Performs non-max suppression to remove excessive detections.
+node {
+  calculator: "NonMaxSuppressionCalculator"
+  input_stream: "detections"
+  output_stream: "filtered_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.NonMaxSuppressionCalculatorOptions] {
+      min_suppression_threshold: 0.3
+      overlap_type: INTERSECTION_OVER_UNION
+      algorithm: WEIGHTED
+      return_empty_detections: true
+    }
+  }
+}
+
+# Maps detection label IDs to the corresponding label text ("Face"). The label
+# map is provided in the label_map_path option.
+node {
+  calculator: "DetectionLabelIdToTextCalculator"
+  input_stream: "filtered_detections"
+  output_stream: "labeled_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
+      label_map_path: "mediapipe/models/face_detection_front_labelmap.txt"
+    }
+  }
+}
+
+# Adjusts detection locations (already normalized to [0.f, 1.f]) on the
+# letterboxed image (after image transformation with the FIT scale mode) to the
+# corresponding locations on the same image with the letterbox removed (the
+# input image to the graph before image transformation).
+node {
+  calculator: "DetectionLetterboxRemovalCalculator"
+  input_stream: "DETECTIONS:labeled_detections"
+  input_stream: "LETTERBOX_PADDING:letterbox_padding"
+  output_stream: "DETECTIONS:output_detections"
+}
+
+# Converts the detections to drawing primitives for annotation overlay.
+node {
+  calculator: "DetectionsToRenderDataCalculator"
+  input_stream: "DETECTIONS:output_detections"
+  output_stream: "RENDER_DATA:render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
+      thickness: 4.0
+      color { r: 255 g: 0 b: 0 }
+    }
+  }
+}
+
+# Draws annotations and overlays them on top of the input images.
+node {
+  calculator: "AnnotationOverlayCalculator"
+  input_stream: "INPUT_FRAME:input_video_cpu"
+  input_stream: "render_data"
+  output_stream: "OUTPUT_FRAME:output_video_cpu"
+}
+
+# Transfers the annotated image from CPU back to GPU memory, to be sent out of
+# the graph.
+node: {
+  calculator: "ImageFrameToGpuBufferCalculator"
+  input_stream: "output_video_cpu"
+  output_stream: "output_video"
+}
+```

mediapipe/docs/hair_segmentation_desktop.md

@@ -0,0 +1,209 @@
+## Hair Segmentation on Desktop
+
+This is an example of using MediaPipe to run hair segmentation models
+(TensorFlow Lite) and render a color to the detected hair. To know more about
+the hair segmentation models, please refer to the model [`README file`].
+Moreover, if you are interested in running the same TensorfFlow Lite model on
+Android/iOS, please see the
+[Hair Segmentation on GPU on Android/iOS](hair_segmentation_mobile_gpu.md) and
+
+We show the hair segmentation demos with TensorFlow Lite model using the Webcam:
+
+-   [TensorFlow Lite Hair Segmentation Demo with Webcam (GPU)](#tensorflow-lite-hair-segmentation-demo-with-webcam-gpu)
+
+Note: Desktop GPU works only on Linux. Mesa drivers need to be installed. Please
+see
+[step 4 of "Installing on Debian and Ubuntu" in the installation guide](./install.md).
+
+Note: If MediaPipe depends on OpenCV 2, please see the [known issues with OpenCV 2](#known-issues-with-opencv-2) section.
+
+### TensorFlow Lite Hair Segmentation Demo with Webcam (GPU)
+
+To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
+
+```bash
+# Video from webcam running on desktop GPU
+# This works only for linux currently
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+    mediapipe/examples/desktop/hair_segmentation:hair_segmentation_gpu
+
+# It should print:
+#INFO: Found 1 target...
+#Target //mediapipe/examples/desktop/hair_segmentation:hair_segmentation_gpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/hair_segmentation/hair_segmentation_gpu
+#INFO: Elapsed time: 18.209s, Forge stats: 13026/13057 actions cached, 20.8s CPU used, 0.0s queue time, 89.3 MB ObjFS output (novel bytes: 87.4 MB), 0.0 MB local output, Critical Path: 11.88s, Remote (86.01% of the time): [queue: 0.00%, network: 16.83%, setup: 4.59%, process: 38.92%]
+#INFO: Streaming build results to: http://sponge2/37d5a184-293b-4e98-a43e-b22084db3142
+#INFO: Build completed successfully, 12210 total actions
+
+$ export GLOG_logtostderr=1
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible,
+# or GPU drivers not setup properly.
+$ bazel-bin/mediapipe/examples/desktop/hair_segmentation/hair_segmentation_gpu \
+    --calculator_graph_config_file=mediapipe/graphs/hair_segmentation/hair_segmentation_mobile_gpu.pbtxt
+```
+
+#### Graph
+
+![hair_segmentation_mobile_gpu_graph](images/mobile/hair_segmentation_mobile_gpu.png)
+
+To visualize the graph as shown above, copy the text specification of the graph
+below and paste it into
+[MediaPipe Visualizer](https://viz.mediapipe.dev).
+
+```bash
+# MediaPipe graph that performs hair segmentation with TensorFlow Lite on GPU.
+# Used in the example in
+# mediapipie/examples/android/src/java/com/mediapipe/apps/hairsegmentationgpu.
+
+# Images on GPU coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Throttles the images flowing downstream for flow control. It passes through
+# the very first incoming image unaltered, and waits for
+# TfLiteTensorsToSegmentationCalculator downstream in the graph to finish
+# generating the corresponding hair mask before it passes through another
+# image. All images that come in while waiting are dropped, limiting the number
+# of in-flight images between this calculator and
+# TfLiteTensorsToSegmentationCalculator to 1. This prevents the nodes in between
+# from queuing up incoming images and data excessively, which leads to increased
+# latency and memory usage, unwanted in real-time mobile applications. It also
+# eliminates unnecessarily computation, e.g., a transformed image produced by
+# ImageTransformationCalculator may get dropped downstream if the subsequent
+# TfLiteConverterCalculator or TfLiteInferenceCalculator is still busy
+# processing previous inputs.
+node {
+  calculator: "FlowLimiterCalculator"
+  input_stream: "input_video"
+  input_stream: "FINISHED:hair_mask"
+  input_stream_info: {
+    tag_index: "FINISHED"
+    back_edge: true
+  }
+  output_stream: "throttled_input_video"
+}
+
+# Transforms the input image on GPU to a 512x512 image. To scale the image, by
+# default it uses the STRETCH scale mode that maps the entire input image to the
+# entire transformed image. As a result, image aspect ratio may be changed and
+# objects in the image may be deformed (stretched or squeezed), but the hair
+# segmentation model used in this graph is agnostic to that deformation.
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE_GPU:throttled_input_video"
+  output_stream: "IMAGE_GPU:transformed_input_video"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 512
+      output_height: 512
+    }
+  }
+}
+
+# Caches a mask fed back from the previous round of hair segmentation, and upon
+# the arrival of the next input image sends out the cached mask with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous mask. Note that upon the arrival of the very first
+# input image, an empty packet is sent out to jump start the feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:throttled_input_video"
+  input_stream: "LOOP:hair_mask"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:previous_hair_mask"
+}
+
+# Embeds the hair mask generated from the previous round of hair segmentation
+# as the alpha channel of the current input image.
+node {
+  calculator: "SetAlphaCalculator"
+  input_stream: "IMAGE_GPU:transformed_input_video"
+  input_stream: "ALPHA_GPU:previous_hair_mask"
+  output_stream: "IMAGE_GPU:mask_embedded_input_video"
+}
+
+# Converts the transformed input image on GPU into an image tensor stored in
+# tflite::gpu::GlBuffer. The zero_center option is set to false to normalize the
+# pixel values to [0.f, 1.f] as opposed to [-1.f, 1.f]. With the
+# max_num_channels option set to 4, all 4 RGBA channels are contained in the
+# image tensor.
+node {
+  calculator: "TfLiteConverterCalculator"
+  input_stream: "IMAGE_GPU:mask_embedded_input_video"
+  output_stream: "TENSORS_GPU:image_tensor"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteConverterCalculatorOptions] {
+      zero_center: false
+      max_num_channels: 4
+    }
+  }
+}
+
+# Generates a single side packet containing a TensorFlow Lite op resolver that
+# supports custom ops needed by the model used in this graph.
+node {
+  calculator: "TfLiteCustomOpResolverCalculator"
+  output_side_packet: "op_resolver"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteCustomOpResolverCalculatorOptions] {
+      use_gpu: true
+    }
+  }
+}
+
+# Runs a TensorFlow Lite model on GPU that takes an image tensor and outputs a
+# tensor representing the hair segmentation, which has the same width and height
+# as the input image tensor.
+node {
+  calculator: "TfLiteInferenceCalculator"
+  input_stream: "TENSORS_GPU:image_tensor"
+  output_stream: "TENSORS_GPU:segmentation_tensor"
+  input_side_packet: "CUSTOM_OP_RESOLVER:op_resolver"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "mediapipe/models/hair_segmentation.tflite"
+      use_gpu: true
+    }
+  }
+}
+
+# Decodes the segmentation tensor generated by the TensorFlow Lite model into a
+# mask of values in [0.f, 1.f], stored in the R channel of a GPU buffer. It also
+# takes the mask generated previously as another input to improve the temporal
+# consistency.
+node {
+  calculator: "TfLiteTensorsToSegmentationCalculator"
+  input_stream: "TENSORS_GPU:segmentation_tensor"
+  input_stream: "PREV_MASK_GPU:previous_hair_mask"
+  output_stream: "MASK_GPU:hair_mask"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToSegmentationCalculatorOptions] {
+      tensor_width: 512
+      tensor_height: 512
+      tensor_channels: 2
+      combine_with_previous_ratio: 0.9
+      output_layer_index: 1
+    }
+  }
+}
+
+# Colors the hair segmentation with the color specified in the option.
+node {
+  calculator: "RecolorCalculator"
+  input_stream: "IMAGE_GPU:throttled_input_video"
+  input_stream: "MASK_GPU:hair_mask"
+  output_stream: "IMAGE_GPU:output_video"
+  node_options: {
+    [type.googleapis.com/mediapipe.RecolorCalculatorOptions] {
+      color { r: 0 g: 0 b: 255 }
+      mask_channel: RED
+    }
+  }
+}
+```
+
+[`README file`]:https://github.com/google/mediapipe/tree/master/mediapipe/README.md

mediapipe/docs/hand_detection_mobile_gpu.md

@@ -1,7 +1,7 @@
 # Hand Detection (GPU)
 
 This doc focuses on the
-[example graph](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/hand_detection_gpu.pbtxt)
+[example graph](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/hand_detection_mobile.pbtxt)
 that performs hand detection with TensorFlow Lite on GPU. It is related to the
 [hand tracking example](./hand_tracking_mobile_gpu.md).
 
@@ -147,7 +147,7 @@ node {
 
 ![hand_detection_gpu_subgraph](images/mobile/hand_detection_gpu_subgraph.png)
 
-[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/hand_detection_gpu.pbtxt)
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/hand_detection_gpu.pbtxt)
 
 ```bash
 # MediaPipe hand detection subgraph.

mediapipe/docs/hand_tracking_desktop.md

@@ -0,0 +1,184 @@
+## Hand Tracking on Desktop
+
+This is an example of using MediaPipe to run hand tracking models (TensorFlow
+Lite) and render bounding boxes on the detected hand (one hand only). To know
+more about the hand tracking models, please refer to the model [`README file`].
+Moreover, if you are interested in running the same TensorfFlow Lite model on
+Android/iOS, please see the
+[Hand Tracking on GPU on Android/iOS](hand_tracking_mobile_gpu.md) and
+
+We show the hand tracking demos with TensorFlow Lite model using the Webcam:
+
+-   [TensorFlow Lite Hand Tracking Demo with Webcam (CPU)](#tensorflow-lite-hand-tracking-demo-with-webcam-cpu)
+
+-   [TensorFlow Lite Hand Tracking Demo with Webcam (GPU)](#tensorflow-lite-hand-tracking-demo-with-webcam-gpu)
+
+Note: Desktop GPU works only on Linux. Mesa drivers need to be installed. Please
+see
+[step 4 of "Installing on Debian and Ubuntu" in the installation guide](./install.md).
+
+Note: If MediaPipe depends on OpenCV 2, please see the [known issues with OpenCV 2](#known-issues-with-opencv-2) section.
+
+### TensorFlow Lite Hand Tracking Demo with Webcam (CPU)
+
+To build and run the TensorFlow Lite example on desktop (CPU) with Webcam, run:
+
+```bash
+# Video from webcam running on desktop CPU
+$ bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 \
+    mediapipe/examples/desktop/hand_tracking:hand_tracking_cpu
+
+# It should print:
+#Target //mediapipe/examples/desktop/hand_tracking:hand_tracking_cpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/hand_tracking/hand_tracking_cpu
+#INFO: Elapsed time: 22.645s, Forge stats: 13356/13463 actions cached, 1.5m CPU used, 0.0s queue time, 819.8 MB ObjFS output (novel bytes: 85.6 MB), 0.0 MB local output, Critical Path: 14.43s, Remote (87.25% of the time): [queue: 0.00%, network: 14.88%, setup: 4.80%, process: 39.80%, fetch: 18.15%]
+#INFO: Streaming build results to: http://sponge2/360196b9-33ab-44b1-84a7-1022b5043307
+#INFO: Build completed successfully, 12517 total actions
+
+$ export GLOG_logtostderr=1
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible
+$ bazel-bin/mediapipe/examples/desktop/hand_tracking/hand_tracking_cpu \
+    --calculator_graph_config_file=mediapipe/graphs/hand_tracking/hand_tracking_desktop_live.pbtxt
+```
+
+### TensorFlow Lite Hand Tracking Demo with Webcam (GPU)
+
+To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
+
+```bash
+# Video from webcam running on desktop GPU
+# This works only for linux currently
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+    mediapipe/examples/desktop/hand_tracking:hand_tracking_gpu
+
+# It should print:
+# Target //mediapipe/examples/desktop/hand_tracking:hand_tracking_gpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/hand_tracking/hand_tracking_gpu
+#INFO: Elapsed time: 84.055s, Forge stats: 6858/19343 actions cached, 1.6h CPU used, 0.9s queue time, 1.68 GB ObjFS output (novel bytes: 485.1 MB), 0.0 MB local output, Critical Path: 48.14s, Remote (99.40% of the time): [queue: 0.00%, setup: 5.59%, process: 74.44%]
+#INFO: Streaming build results to: http://sponge2/00c7f95f-6fbc-432d-8978-f5d361efca3b
+#INFO: Build completed successfully, 22455 total actions
+
+$ export GLOG_logtostderr=1
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible,
+# or GPU drivers not setup properly.
+$ bazel-bin/mediapipe/examples/desktop/hand_tracking/hand_tracking_gpu \
+    --calculator_graph_config_file=mediapipe/graphs/hand_tracking/hand_tracking_mobile.pbtxt
+```
+
+#### Graph
+
+![graph visualization](images/hand_tracking_desktop.png)
+
+To visualize the graph as shown above, copy the text specification of the graph
+below and paste it into
+[MediaPipe Visualizer](https://viz.mediapipe.dev).
+
+```bash
+# MediaPipe graph that performs hand tracking on desktop with TensorFlow Lite
+# on CPU & GPU.
+# Used in the example in
+# mediapipie/examples/desktop/hand_tracking:hand_tracking_cpu.
+
+# Images coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Caches a hand-presence decision fed back from HandLandmarkSubgraph, and upon
+# the arrival of the next input image sends out the cached decision with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand-presence decision. Note that upon the arrival
+# of the very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:input_video"
+  input_stream: "LOOP:hand_presence"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_hand_presence"
+}
+
+# Drops the incoming image if HandLandmarkSubgraph was able to identify hand
+# presence in the previous image. Otherwise, passes the incoming image through
+# to trigger a new round of hand detection in HandDetectionSubgraph.
+node {
+  calculator: "GateCalculator"
+  input_stream: "input_video"
+  input_stream: "DISALLOW:prev_hand_presence"
+  output_stream: "hand_detection_input_video"
+
+  node_options: {
+    [type.googleapis.com/mediapipe.GateCalculatorOptions] {
+      empty_packets_as_allow: true
+    }
+  }
+}
+
+# Subgraph that detections hands (see hand_detection_cpu.pbtxt).
+node {
+  calculator: "HandDetectionSubgraph"
+  input_stream: "hand_detection_input_video"
+  output_stream: "DETECTIONS:palm_detections"
+  output_stream: "NORM_RECT:hand_rect_from_palm_detections"
+}
+
+# Subgraph that localizes hand landmarks (see hand_landmark_cpu.pbtxt).
+node {
+  calculator: "HandLandmarkSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "NORM_RECT:hand_rect"
+  output_stream: "LANDMARKS:hand_landmarks"
+  output_stream: "NORM_RECT:hand_rect_from_landmarks"
+  output_stream: "PRESENCE:hand_presence"
+}
+
+# Caches a hand rectangle fed back from HandLandmarkSubgraph, and upon the
+# arrival of the next input image sends out the cached rectangle with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand rectangle. Note that upon the arrival of the
+# very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:input_video"
+  input_stream: "LOOP:hand_rect_from_landmarks"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_hand_rect_from_landmarks"
+}
+
+# Merges a stream of hand rectangles generated by HandDetectionSubgraph and that
+# generated by HandLandmarkSubgraph into a single output stream by selecting
+# between one of the two streams. The former is selected if the incoming packet
+# is not empty, i.e., hand detection is performed on the current image by
+# HandDetectionSubgraph (because HandLandmarkSubgraph could not identify hand
+# presence in the previous image). Otherwise, the latter is selected, which is
+# never empty because HandLandmarkSubgraphs processes all images (that went
+# through FlowLimiterCaculator).
+node {
+  calculator: "MergeCalculator"
+  input_stream: "hand_rect_from_palm_detections"
+  input_stream: "prev_hand_rect_from_landmarks"
+  output_stream: "hand_rect"
+}
+
+# Subgraph that renders annotations and overlays them on top of the input
+# images (see renderer_cpu.pbtxt).
+node {
+  calculator: "RendererSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "LANDMARKS:hand_landmarks"
+  input_stream: "NORM_RECT:hand_rect"
+  input_stream: "DETECTIONS:palm_detections"
+  output_stream: "IMAGE:output_video"
+}
+
+```
+
+[`README file`]:https://github.com/google/mediapipe/tree/master/mediapipe/README.md

mediapipe/docs/hand_tracking_mobile_gpu.md

@@ -227,7 +227,7 @@ node {
 
 ![hand_detection_gpu_subgraph](images/mobile/hand_detection_gpu_subgraph.png)
 
-[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/hand_detection_gpu.pbtxt)
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/hand_detection_gpu.pbtxt)
 
 ```bash
 # MediaPipe hand detection subgraph.
@@ -433,7 +433,7 @@ node {
 
 ![hand_landmark_gpu_subgraph.pbtxt](images/mobile/hand_landmark_gpu_subgraph.png)
 
-[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/hand_landmark_gpu.pbtxt)
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/hand_landmark_gpu.pbtxt)
 
 ```bash
 # MediaPipe hand landmark localization subgraph.
@@ -617,7 +617,7 @@ node {
 
 ![hand_renderer_gpu_subgraph.pbtxt](images/mobile/hand_renderer_gpu_subgraph.png)
 
-[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/renderer_gpu.pbtxt)
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/renderer_gpu.pbtxt)
 
 ```bash
 # MediaPipe hand tracking rendering subgraph.

mediapipe/docs/install.md

@@ -107,14 +107,34 @@ To build and run iOS apps:
     )
     ```
 
-4.  Run the [Hello World desktop example](./hello_world_desktop.md).
+4.  For running desktop examples on Linux only (not on OS X) with GPU
+    acceleration.
+
+    ```bash
+    # Requires a GPU with EGL driver support.
+    # Can use mesa GPU libraries for desktop, (or Nvidia/AMD equivalent).
+    sudo apt-get install mesa-common-dev libegl1-mesa-dev libgles2-mesa-dev
+
+    # To compile with GPU support, replace
+    --define MEDIAPIPE_DISABLE_GPU=1
+    # with
+    --copt -DMESA_EGL_NO_X11_HEADERS
+    # when building GPU examples.
+    ```
+
+5.  Run the [Hello World desktop example](./hello_world_desktop.md).
 
     ```bash
     $ export GLOG_logtostderr=1
-    # Need bazel flag 'MEDIAPIPE_DISABLE_GPU=1' as desktop GPU is currently not supported
+
+    # if you are running on Linux desktop with CPU only
     $ bazel run --define MEDIAPIPE_DISABLE_GPU=1 \
         mediapipe/examples/desktop/hello_world:hello_world
 
+    # If you are running on Linux desktop with GPU support enabled (via mesa drivers)
+    $ bazel run --copt -DMESA_EGL_NO_X11_HEADERS \
+        mediapipe/examples/desktop/hello_world:hello_world
+
     # Should print:
     # Hello World!
     # Hello World!
@@ -194,7 +214,7 @@ To build and run iOS apps:
 
     ```bash
     $ export GLOG_logtostderr=1
-    # Need bazel flag 'MEDIAPIPE_DISABLE_GPU=1' as desktop GPU is currently not supported
+    # Need bazel flag 'MEDIAPIPE_DISABLE_GPU=1' if you are running on Linux desktop with CPU only
     $ bazel run --define MEDIAPIPE_DISABLE_GPU=1 \
         mediapipe/examples/desktop/hello_world:hello_world
 

mediapipe/docs/object_detection_desktop.md

@@ -12,6 +12,7 @@ We show the object detection demo with both TensorFlow model and TensorFlow Lite
 
 -   [TensorFlow Object Detection Demo](#tensorflow-object-detection-demo)
 -   [TensorFlow Lite Object Detection Demo](#tensorflow-lite-object-detection-demo)
+-   [TensorFlow Lite Object Detection Demo with Webcam (CPU)](#tensorflow-lite-object-detection-demo)
 
 Note: If MediaPipe depends on OpenCV 2, please see the [known issues with OpenCV 2](#known-issues-with-opencv-2) section.
 
@@ -207,6 +208,29 @@ $ bazel-bin/mediapipe/examples/desktop/object_detection/object_detection_tflite
     --input_side_packets=input_video_path=<input video path>,output_video_path=<output video path>
 ```
 
+### TensorFlow Lite Object Detection Demo with Webcam (CPU)
+
+To build and run the TensorFlow Lite example on desktop (CPU) with Webcam, run:
+
+```bash
+# Video from webcam running on desktop CPU
+$ bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 \
+    mediapipe/examples/desktop/object_detection:object_detection_cpu
+
+# It should print:
+#Target //mediapipe/examples/desktop/object_detection:object_detection_cpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/object_detection/object_detection_cpu
+#INFO: Elapsed time: 16.020s, Forge stats: 13001/13003 actions cached, 2.1s CPU used, 0.0s queue time, 89.0 MB ObjFS output (novel bytes: 88.0 MB), 0.0 MB local output, Critical Path: 10.01s, Remote (41.42% of the time): [queue: 0.00%, setup: 4.21%, process: 12.48%]
+#INFO: Streaming build results to: http://sponge2/1824d4cc-ba63-4350-bdc0-aacbd45b902b
+#INFO: Build completed successfully, 12154 total actions
+
+$ export GLOG_logtostderr=1
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible
+$ bazel-bin/mediapipe/examples/desktop/object_detection/object_detection_cpu \
+    --calculator_graph_config_file=mediapipe/graphs/object_detection/object_detection_desktop_live.pbtxt
+```
+
 #### Graph
 
 ![graph visualization](images/object_detection_desktop_tflite.png)

mediapipe/docs/visualizer.md

@@ -77,7 +77,7 @@ For instance, there are two graphs involved in the
 [hand detection example](./hand_detection_mobile_gpu.md): the main graph
 ([source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/hand_detection_mobile.pbtxt))
 and its associated subgraph
-([source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/hand_detection_gpu.pbtxt)).
+([source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/hand_detection_gpu.pbtxt)).
 To visualize them:
 
 *   In the MediaPipe visualizer, click on the upload graph button and select the

mediapipe/examples/desktop/BUILD

@@ -14,7 +14,9 @@
 
 licenses(["notice"])  # Apache 2.0
 
-package(default_visibility = ["//mediapipe/examples:__subpackages__"])
+package(default_visibility = [
+    "//visibility:public",
+])
 
 cc_library(
     name = "simple_run_graph_main",
@@ -29,3 +31,44 @@ cc_library(
         "@com_google_absl//absl/strings",
     ],
 )
+
+cc_library(
+    name = "demo_run_graph_main",
+    srcs = ["demo_run_graph_main.cc"],
+    deps = [
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
+        "//mediapipe/framework/port:commandlineflags",
+        "//mediapipe/framework/port:file_helpers",
+        "//mediapipe/framework/port:opencv_highgui",
+        "//mediapipe/framework/port:opencv_imgproc",
+        "//mediapipe/framework/port:opencv_video",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
+    ],
+)
+
+# Linux only.
+# Must have a GPU with EGL support:
+# ex: sudo aptitude install mesa-common-dev libegl1-mesa-dev libgles2-mesa-dev
+# (or similar nvidia/amd equivalent)
+cc_library(
+    name = "demo_run_graph_main_gpu",
+    srcs = ["demo_run_graph_main_gpu.cc"],
+    deps = [
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
+        "//mediapipe/framework/port:commandlineflags",
+        "//mediapipe/framework/port:file_helpers",
+        "//mediapipe/framework/port:opencv_highgui",
+        "//mediapipe/framework/port:opencv_imgproc",
+        "//mediapipe/framework/port:opencv_video",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/gpu:gl_calculator_helper",
+        "//mediapipe/gpu:gpu_buffer",
+        "//mediapipe/gpu:gpu_shared_data_internal",
+    ],
+)

mediapipe/examples/desktop/demo_run_graph_main.cc

@@ -0,0 +1,146 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// An example of sending OpenCV webcam frames into a MediaPipe graph.
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/port/commandlineflags.h"
+#include "mediapipe/framework/port/file_helpers.h"
+#include "mediapipe/framework/port/opencv_highgui_inc.h"
+#include "mediapipe/framework/port/opencv_imgproc_inc.h"
+#include "mediapipe/framework/port/opencv_video_inc.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status.h"
+
+constexpr char kInputStream[] = "input_video";
+constexpr char kOutputStream[] = "output_video";
+constexpr char kWindowName[] = "MediaPipe";
+
+DEFINE_string(
+    calculator_graph_config_file, "",
+    "Name of file containing text format CalculatorGraphConfig proto.");
+DEFINE_string(input_video_path, "",
+              "Full path of video to load. "
+              "If not provided, attempt to use a webcam.");
+DEFINE_string(output_video_path, "",
+              "Full path of where to save result (.mp4 only). "
+              "If not provided, show result in a window.");
+
+::mediapipe::Status RunMPPGraph() {
+  std::string calculator_graph_config_contents;
+  MP_RETURN_IF_ERROR(mediapipe::file::GetContents(
+      FLAGS_calculator_graph_config_file, &calculator_graph_config_contents));
+  LOG(INFO) << "Get calculator graph config contents: "
+            << calculator_graph_config_contents;
+  mediapipe::CalculatorGraphConfig config =
+      mediapipe::ParseTextProtoOrDie<mediapipe::CalculatorGraphConfig>(
+          calculator_graph_config_contents);
+
+  LOG(INFO) << "Initialize the calculator graph.";
+  mediapipe::CalculatorGraph graph;
+  MP_RETURN_IF_ERROR(graph.Initialize(config));
+
+  LOG(INFO) << "Initialize the camera or load the video.";
+  cv::VideoCapture capture;
+  const bool load_video = !FLAGS_input_video_path.empty();
+  if (load_video) {
+    capture.open(FLAGS_input_video_path);
+  } else {
+    capture.open(0);
+  }
+  RET_CHECK(capture.isOpened());
+
+  cv::VideoWriter writer;
+  const bool save_video = !FLAGS_output_video_path.empty();
+  if (save_video) {
+    LOG(INFO) << "Prepare video writer.";
+    cv::Mat test_frame;
+    capture.read(test_frame);                    // Consume first frame.
+    capture.set(cv::CAP_PROP_POS_AVI_RATIO, 0);  // Rewind to beginning.
+    writer.open(FLAGS_output_video_path,
+                mediapipe::fourcc('a', 'v', 'c', '1'),  // .mp4
+                capture.get(cv::CAP_PROP_FPS), test_frame.size());
+    RET_CHECK(writer.isOpened());
+  } else {
+    cv::namedWindow(kWindowName, /*flags=WINDOW_AUTOSIZE*/ 1);
+  }
+
+  LOG(INFO) << "Start running the calculator graph.";
+  ASSIGN_OR_RETURN(mediapipe::OutputStreamPoller poller,
+                   graph.AddOutputStreamPoller(kOutputStream));
+  MP_RETURN_IF_ERROR(graph.StartRun({}));
+
+  LOG(INFO) << "Start grabbing and processing frames.";
+  size_t frame_timestamp = 0;
+  bool grab_frames = true;
+  while (grab_frames) {
+    // Capture opencv camera or video frame.
+    cv::Mat camera_frame_raw;
+    capture >> camera_frame_raw;
+    if (camera_frame_raw.empty()) break;  // End of video.
+    cv::Mat camera_frame;
+    cv::cvtColor(camera_frame_raw, camera_frame, cv::COLOR_BGR2RGB);
+    if (!load_video) {
+      cv::flip(camera_frame, camera_frame, /*flipcode=HORIZONTAL*/ 1);
+    }
+
+    // Wrap Mat into an ImageFrame.
+    auto input_frame = absl::make_unique<mediapipe::ImageFrame>(
+        mediapipe::ImageFormat::SRGB, camera_frame.cols, camera_frame.rows,
+        mediapipe::ImageFrame::kDefaultAlignmentBoundary);
+    cv::Mat input_frame_mat = mediapipe::formats::MatView(input_frame.get());
+    camera_frame.copyTo(input_frame_mat);
+
+    // Send image packet into the graph.
+    MP_RETURN_IF_ERROR(graph.AddPacketToInputStream(
+        kInputStream, mediapipe::Adopt(input_frame.release())
+                          .At(mediapipe::Timestamp(frame_timestamp++))));
+
+    // Get the graph result packet, or stop if that fails.
+    mediapipe::Packet packet;
+    if (!poller.Next(&packet)) break;
+    auto& output_frame = packet.Get<mediapipe::ImageFrame>();
+
+    // Convert back to opencv for display or saving.
+    cv::Mat output_frame_mat = mediapipe::formats::MatView(&output_frame);
+    cv::cvtColor(output_frame_mat, output_frame_mat, cv::COLOR_RGB2BGR);
+    if (save_video) {
+      writer.write(output_frame_mat);
+    } else {
+      cv::imshow(kWindowName, output_frame_mat);
+      // Press any key to exit.
+      const int pressed_key = cv::waitKey(5);
+      if (pressed_key >= 0 && pressed_key != 255) grab_frames = false;
+    }
+  }
+
+  LOG(INFO) << "Shutting down.";
+  if (writer.isOpened()) writer.release();
+  MP_RETURN_IF_ERROR(graph.CloseInputStream(kInputStream));
+  return graph.WaitUntilDone();
+}
+
+int main(int argc, char** argv) {
+  google::InitGoogleLogging(argv[0]);
+  gflags::ParseCommandLineFlags(&argc, &argv, true);
+  ::mediapipe::Status run_status = RunMPPGraph();
+  if (!run_status.ok()) {
+    LOG(ERROR) << "Failed to run the graph: " << run_status.message();
+  } else {
+    LOG(INFO) << "Success!";
+  }
+  return 0;
+}

mediapipe/examples/desktop/demo_run_graph_main_gpu.cc

@@ -0,0 +1,186 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// An example of sending OpenCV webcam frames into a MediaPipe graph.
+// This example requires a linux computer and a GPU with EGL support drivers.
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/port/commandlineflags.h"
+#include "mediapipe/framework/port/file_helpers.h"
+#include "mediapipe/framework/port/opencv_highgui_inc.h"
+#include "mediapipe/framework/port/opencv_imgproc_inc.h"
+#include "mediapipe/framework/port/opencv_video_inc.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/gpu/gl_calculator_helper.h"
+#include "mediapipe/gpu/gpu_buffer.h"
+#include "mediapipe/gpu/gpu_shared_data_internal.h"
+
+constexpr char kInputStream[] = "input_video";
+constexpr char kOutputStream[] = "output_video";
+constexpr char kWindowName[] = "MediaPipe";
+
+DEFINE_string(
+    calculator_graph_config_file, "",
+    "Name of file containing text format CalculatorGraphConfig proto.");
+DEFINE_string(input_video_path, "",
+              "Full path of video to load. "
+              "If not provided, attempt to use a webcam.");
+DEFINE_string(output_video_path, "",
+              "Full path of where to save result (.mp4 only). "
+              "If not provided, show result in a window.");
+
+::mediapipe::Status RunMPPGraph() {
+  std::string calculator_graph_config_contents;
+  MP_RETURN_IF_ERROR(mediapipe::file::GetContents(
+      FLAGS_calculator_graph_config_file, &calculator_graph_config_contents));
+  LOG(INFO) << "Get calculator graph config contents: "
+            << calculator_graph_config_contents;
+  mediapipe::CalculatorGraphConfig config =
+      mediapipe::ParseTextProtoOrDie<mediapipe::CalculatorGraphConfig>(
+          calculator_graph_config_contents);
+
+  LOG(INFO) << "Initialize the calculator graph.";
+  mediapipe::CalculatorGraph graph;
+  MP_RETURN_IF_ERROR(graph.Initialize(config));
+
+  LOG(INFO) << "Initialize the GPU.";
+  ASSIGN_OR_RETURN(auto gpu_resources, mediapipe::GpuResources::Create());
+  MP_RETURN_IF_ERROR(graph.SetGpuResources(std::move(gpu_resources)));
+  mediapipe::GlCalculatorHelper gpu_helper;
+  gpu_helper.InitializeForTest(graph.GetGpuResources().get());
+
+  LOG(INFO) << "Initialize the camera or load the video.";
+  cv::VideoCapture capture;
+  const bool load_video = !FLAGS_input_video_path.empty();
+  if (load_video) {
+    capture.open(FLAGS_input_video_path);
+  } else {
+    capture.open(0);
+  }
+  RET_CHECK(capture.isOpened());
+
+  cv::VideoWriter writer;
+  const bool save_video = !FLAGS_output_video_path.empty();
+  if (save_video) {
+    LOG(INFO) << "Prepare video writer.";
+    cv::Mat test_frame;
+    capture.read(test_frame);                    // Consume first frame.
+    capture.set(cv::CAP_PROP_POS_AVI_RATIO, 0);  // Rewind to beginning.
+    writer.open(FLAGS_output_video_path,
+                mediapipe::fourcc('a', 'v', 'c', '1'),  // .mp4
+                capture.get(cv::CAP_PROP_FPS), test_frame.size());
+    RET_CHECK(writer.isOpened());
+  } else {
+    cv::namedWindow(kWindowName, /*flags=WINDOW_AUTOSIZE*/ 1);
+  }
+
+  LOG(INFO) << "Start running the calculator graph.";
+  ASSIGN_OR_RETURN(mediapipe::OutputStreamPoller poller,
+                   graph.AddOutputStreamPoller(kOutputStream));
+  MP_RETURN_IF_ERROR(graph.StartRun({}));
+
+  LOG(INFO) << "Start grabbing and processing frames.";
+  size_t frame_timestamp = 0;
+  bool grab_frames = true;
+  while (grab_frames) {
+    // Capture opencv camera or video frame.
+    cv::Mat camera_frame_raw;
+    capture >> camera_frame_raw;
+    if (camera_frame_raw.empty()) break;  // End of video.
+    cv::Mat camera_frame;
+    cv::cvtColor(camera_frame_raw, camera_frame, cv::COLOR_BGR2RGB);
+    if (!load_video) {
+      cv::flip(camera_frame, camera_frame, /*flipcode=HORIZONTAL*/ 1);
+    }
+
+    // Wrap Mat into an ImageFrame.
+    auto input_frame = absl::make_unique<mediapipe::ImageFrame>(
+        mediapipe::ImageFormat::SRGB, camera_frame.cols, camera_frame.rows,
+        mediapipe::ImageFrame::kGlDefaultAlignmentBoundary);
+    cv::Mat input_frame_mat = mediapipe::formats::MatView(input_frame.get());
+    camera_frame.copyTo(input_frame_mat);
+
+    // Prepare and add graph input packet.
+    MP_RETURN_IF_ERROR(
+        gpu_helper.RunInGlContext([&input_frame, &frame_timestamp, &graph,
+                                   &gpu_helper]() -> ::mediapipe::Status {
+          // Convert ImageFrame to GpuBuffer.
+          auto texture = gpu_helper.CreateSourceTexture(*input_frame.get());
+          auto gpu_frame = texture.GetFrame<mediapipe::GpuBuffer>();
+          glFlush();
+          texture.Release();
+          // Send GPU image packet into the graph.
+          MP_RETURN_IF_ERROR(graph.AddPacketToInputStream(
+              kInputStream, mediapipe::Adopt(gpu_frame.release())
+                                .At(mediapipe::Timestamp(frame_timestamp++))));
+          return ::mediapipe::OkStatus();
+        }));
+
+    // Get the graph result packet, or stop if that fails.
+    mediapipe::Packet packet;
+    if (!poller.Next(&packet)) break;
+    std::unique_ptr<mediapipe::ImageFrame> output_frame;
+
+    // Convert GpuBuffer to ImageFrame.
+    MP_RETURN_IF_ERROR(gpu_helper.RunInGlContext(
+        [&packet, &output_frame, &gpu_helper]() -> ::mediapipe::Status {
+          auto& gpu_frame = packet.Get<mediapipe::GpuBuffer>();
+          auto texture = gpu_helper.CreateSourceTexture(gpu_frame);
+          output_frame = absl::make_unique<mediapipe::ImageFrame>(
+              mediapipe::ImageFormatForGpuBufferFormat(gpu_frame.format()),
+              gpu_frame.width(), gpu_frame.height(),
+              mediapipe::ImageFrame::kGlDefaultAlignmentBoundary);
+          gpu_helper.BindFramebuffer(texture);
+          const auto info =
+              mediapipe::GlTextureInfoForGpuBufferFormat(gpu_frame.format(), 0);
+          glReadPixels(0, 0, texture.width(), texture.height(), info.gl_format,
+                       info.gl_type, output_frame->MutablePixelData());
+          glFlush();
+          texture.Release();
+          return ::mediapipe::OkStatus();
+        }));
+
+    // Convert back to opencv for display or saving.
+    cv::Mat output_frame_mat = mediapipe::formats::MatView(output_frame.get());
+    cv::cvtColor(output_frame_mat, output_frame_mat, cv::COLOR_RGB2BGR);
+    if (save_video) {
+      writer.write(output_frame_mat);
+    } else {
+      cv::imshow(kWindowName, output_frame_mat);
+      // Press any key to exit.
+      const int pressed_key = cv::waitKey(5);
+      if (pressed_key >= 0 && pressed_key != 255) grab_frames = false;
+    }
+  }
+
+  LOG(INFO) << "Shutting down.";
+  if (writer.isOpened()) writer.release();
+  MP_RETURN_IF_ERROR(graph.CloseInputStream(kInputStream));
+  return graph.WaitUntilDone();
+}
+
+int main(int argc, char** argv) {
+  google::InitGoogleLogging(argv[0]);
+  gflags::ParseCommandLineFlags(&argc, &argv, true);
+  ::mediapipe::Status run_status = RunMPPGraph();
+  if (!run_status.ok()) {
+    LOG(ERROR) << "Failed to run the graph: " << run_status.message();
+  } else {
+    LOG(INFO) << "Success!";
+  }
+  return 0;
+}

mediapipe/examples/desktop/face_detection/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2019 The MediaPipe Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+licenses(["notice"])  # Apache 2.0
+
+package(default_visibility = ["//mediapipe/examples:__subpackages__"])
+
+cc_binary(
+    name = "face_detection_cpu",
+    deps = [
+        "//mediapipe/examples/desktop:demo_run_graph_main",
+        "//mediapipe/graphs/face_detection:desktop_tflite_calculators",
+    ],
+)
+
+# Linux only
+cc_binary(
+    name = "face_detection_gpu",
+    deps = [
+        "//mediapipe/examples/desktop:demo_run_graph_main_gpu",
+        "//mediapipe/graphs/face_detection:mobile_calculators",
+    ],
+)

mediapipe/examples/desktop/hair_segmentation/BUILD

@@ -0,0 +1,26 @@
+# Copyright 2019 The MediaPipe Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+licenses(["notice"])  # Apache 2.0
+
+package(default_visibility = ["//mediapipe/examples:__subpackages__"])
+
+# Linux only
+cc_binary(
+    name = "hair_segmentation_gpu",
+    deps = [
+        "//mediapipe/examples/desktop:demo_run_graph_main_gpu",
+        "//mediapipe/graphs/hair_segmentation:mobile_calculators",
+    ],
+)

mediapipe/examples/desktop/hand_tracking/BUILD

@@ -0,0 +1,42 @@
+# Copyright 2019 The MediaPipe Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+licenses(["notice"])  # Apache 2.0
+
+package(default_visibility = ["//mediapipe/examples:__subpackages__"])
+
+cc_binary(
+    name = "hand_tracking_tflite",
+    deps = [
+        "//mediapipe/examples/desktop:simple_run_graph_main",
+        "//mediapipe/graphs/hand_tracking:desktop_tflite_calculators",
+    ],
+)
+
+cc_binary(
+    name = "hand_tracking_cpu",
+    deps = [
+        "//mediapipe/examples/desktop:demo_run_graph_main",
+        "//mediapipe/graphs/hand_tracking:desktop_tflite_calculators",
+    ],
+)
+
+# Linux only
+cc_binary(
+    name = "hand_tracking_gpu",
+    deps = [
+        "//mediapipe/examples/desktop:demo_run_graph_main_gpu",
+        "//mediapipe/graphs/hand_tracking:mobile_calculators",
+    ],
+)

mediapipe/examples/desktop/object_detection/BUILD

@@ -72,3 +72,11 @@ cc_binary(
         "//mediapipe/graphs/object_detection:desktop_tflite_calculators",
     ],
 )
+
+cc_binary(
+    name = "object_detection_cpu",
+    deps = [
+        "//mediapipe/examples/desktop:demo_run_graph_main",
+        "//mediapipe/graphs/object_detection:desktop_tflite_calculators",
+    ],
+)

mediapipe/examples/desktop/youtube8m/BUILD

@@ -27,7 +27,7 @@ cc_binary(
         "//mediapipe/framework/port:map_util",
         "//mediapipe/framework/port:parse_text_proto",
         "//mediapipe/framework/port:status",
-        "//mediapipe/graphs/youtube8m:yt8m_calculators_deps",
+        "//mediapipe/graphs/youtube8m:yt8m_feature_extraction_calculators",
         # TODO: Figure out the minimum set of the kernels needed by this example.
         "@org_tensorflow//tensorflow/core:all_kernels",
         "@org_tensorflow//tensorflow/core:direct_session",

mediapipe/examples/desktop/youtube8m/README.md

@@ -37,7 +37,9 @@
 
     ```bash
     python -m mediapipe.examples.desktop.youtube8m.generate_input_sequence_example \
-      --path_to_input_video=/absolute/path/to/the/local/video/file
+      --path_to_input_video=/absolute/path/to/the/local/video/file \
+      --clip_start_time_sec=0 \
+      --clip_end_time_sec=10
     ```
 
 5.  Run the MediaPipe binary to extract the features

mediapipe/examples/desktop/youtube8m/generate_input_sequence_example.py

@@ -37,20 +37,29 @@ def bytes23(string):
 
 
 def main(argv):
-  if len(argv) > 1:
+  if len(argv) > 3:
     raise app.UsageError('Too many command-line arguments.')
 
   if not flags.FLAGS.path_to_input_video:
     raise ValueError('You must specify the path to the input video.')
+  if not flags.FLAGS.clip_end_time_sec:
+    raise ValueError('You must specify the clip end timestamp in seconds.')
+  if flags.FLAGS.clip_start_time_sec >= flags.FLAGS.clip_end_time_sec:
+    raise ValueError(
+        'The clip start time must be greater than the clip end time.')
   metadata = tf.train.SequenceExample()
   ms.set_clip_data_path(bytes23(flags.FLAGS.path_to_input_video), metadata)
-  ms.set_clip_start_timestamp(0, metadata)
+  ms.set_clip_start_timestamp(
+      flags.FLAGS.clip_start_time_sec * SECONDS_TO_MICROSECONDS, metadata)
   ms.set_clip_end_timestamp(
-      int(float(300 * SECONDS_TO_MICROSECONDS)), metadata)
+      flags.FLAGS.clip_end_time_sec * SECONDS_TO_MICROSECONDS, metadata)
   with open('/tmp/mediapipe/metadata.tfrecord', 'wb') as writer:
     writer.write(metadata.SerializeToString())
 
 
 if __name__ == '__main__':
   flags.DEFINE_string('path_to_input_video', '', 'Path to the input video.')
+  flags.DEFINE_integer('clip_start_time_sec', 0,
+                       'Clip start timestamp in seconds')
+  flags.DEFINE_integer('clip_end_time_sec', 10, 'Clip end timestamp in seconds')
   app.run(main)

mediapipe/framework/BUILD

@@ -1134,6 +1134,7 @@ cc_library(
         "//mediapipe/framework/port:status",
         "//mediapipe/framework/port:statusor",
         "//mediapipe/framework/tool:calculator_graph_template_cc_proto",
+        "//mediapipe/framework/tool:options_util",
         "//mediapipe/framework/tool:template_expander",
         "@com_google_absl//absl/base:core_headers",
         "@com_google_absl//absl/memory",
@@ -1499,13 +1500,47 @@ cc_test(
     deps = [
         ":calculator_context",
         ":calculator_framework",
+        ":test_calculators",
+        ":thread_pool_executor",
         ":timestamp",
+        ":type_map",
+        "//mediapipe/calculators/core:counting_source_calculator",
+        "//mediapipe/calculators/core:mux_calculator",
         "//mediapipe/calculators/core:pass_through_calculator",
         "//mediapipe/framework/port:gtest_main",
         "//mediapipe/framework/port:logging",
         "//mediapipe/framework/port:parse_text_proto",
         "//mediapipe/framework/port:status",
+        "//mediapipe/framework/stream_handler:barrier_input_stream_handler",
+        "//mediapipe/framework/stream_handler:early_close_input_stream_handler",
+        "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
         "//mediapipe/framework/stream_handler:immediate_input_stream_handler",
+        "//mediapipe/framework/stream_handler:mux_input_stream_handler",
+        "//mediapipe/framework/tool:sink",
+    ],
+)
+
+cc_test(
+    name = "calculator_graph_side_packet_test",
+    size = "small",
+    srcs = [
+        "calculator_graph_side_packet_test.cc",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":calculator_framework",
+        ":test_calculators",
+        "//mediapipe/calculators/core:counting_source_calculator",
+        "//mediapipe/calculators/core:mux_calculator",
+        "//mediapipe/calculators/core:pass_through_calculator",
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:logging",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/tool:sink",
+        "@com_google_absl//absl/time",
     ],
 )
 

mediapipe/framework/calculator_graph_bounds_test.cc

@@ -21,11 +21,258 @@
 #include "mediapipe/framework/port/parse_text_proto.h"
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/framework/port/status_matchers.h"
+#include "mediapipe/framework/thread_pool_executor.h"
 #include "mediapipe/framework/timestamp.h"
 
 namespace mediapipe {
 namespace {
 
+typedef std::function<::mediapipe::Status(CalculatorContext* cc)>
+    CalculatorContextFunction;
+
+// A simple Semaphore for synchronizing test threads.
+class AtomicSemaphore {
+ public:
+  AtomicSemaphore(int64_t supply) : supply_(supply) {}
+  void Acquire(int64_t amount) {
+    while (supply_.fetch_sub(amount) - amount < 0) {
+      Release(amount);
+    }
+  }
+  void Release(int64_t amount) { supply_ += amount; }
+
+ private:
+  std::atomic<int64_t> supply_;
+};
+
+// A mediapipe::Executor that signals the start and finish of each task.
+class CountingExecutor : public Executor {
+ public:
+  CountingExecutor(int num_threads, std::function<void()> start_callback,
+                   std::function<void()> finish_callback)
+      : thread_pool_(num_threads),
+        start_callback_(std::move(start_callback)),
+        finish_callback_(std::move(finish_callback)) {
+    thread_pool_.StartWorkers();
+  }
+  void Schedule(std::function<void()> task) override {
+    start_callback_();
+    thread_pool_.Schedule([this, task] {
+      task();
+      finish_callback_();
+    });
+  }
+
+ private:
+  ThreadPool thread_pool_;
+  std::function<void()> start_callback_;
+  std::function<void()> finish_callback_;
+};
+
+// A Calculator that adds the integer values in the packets in all the input
+// streams and outputs the sum to the output stream.
+class IntAdderCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    for (int i = 0; i < cc->Inputs().NumEntries(); ++i) {
+      cc->Inputs().Index(i).Set<int>();
+    }
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    cc->SetOffset(TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    int sum = 0;
+    for (int i = 0; i < cc->Inputs().NumEntries(); ++i) {
+      sum += cc->Inputs().Index(i).Get<int>();
+    }
+    cc->Outputs().Index(0).Add(new int(sum), cc->InputTimestamp());
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(IntAdderCalculator);
+
+template <typename InputType>
+class TypedSinkCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).Set<InputType>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    return ::mediapipe::OkStatus();
+  }
+};
+typedef TypedSinkCalculator<std::string> StringSinkCalculator;
+typedef TypedSinkCalculator<int> IntSinkCalculator;
+REGISTER_CALCULATOR(StringSinkCalculator);
+REGISTER_CALCULATOR(IntSinkCalculator);
+
+// A Calculator that passes an input packet through if it contains an even
+// integer.
+class EvenIntFilterCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).Set<int>();
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    int value = cc->Inputs().Index(0).Get<int>();
+    if (value % 2 == 0) {
+      cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+    } else {
+      cc->Outputs().Index(0).SetNextTimestampBound(
+          cc->InputTimestamp().NextAllowedInStream());
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(EvenIntFilterCalculator);
+
+// A Calculator that passes packets through or not, depending on a second
+// input. The first input stream's packets are only propagated if the second
+// input stream carries the value true.
+class ValveCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->Inputs().Index(1).Set<bool>();
+    cc->Outputs().Index(0).SetSameAs(&cc->Inputs().Index(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    cc->Outputs().Index(0).SetHeader(cc->Inputs().Index(0).Header());
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    if (cc->Inputs().Index(1).Get<bool>()) {
+      cc->GetCounter("PassThrough")->Increment();
+      cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+    } else {
+      cc->GetCounter("Block")->Increment();
+      // The next timestamp bound is the minimum timestamp that the next packet
+      // can have, so, if we want to inform the downstream that no packet at
+      // InputTimestamp() is coming, we need to set it to the next value.
+      // We could also just call SetOffset(TimestampDiff(0)) in Open, and then
+      // we would not have to call this manually.
+      cc->Outputs().Index(0).SetNextTimestampBound(
+          cc->InputTimestamp().NextAllowedInStream());
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(ValveCalculator);
+
+// A Calculator that simply passes its input Packets and header through,
+// but shifts the timestamp.
+class TimeShiftCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->Outputs().Index(0).SetSameAs(&cc->Inputs().Index(0));
+    cc->InputSidePackets().Index(0).Set<TimestampDiff>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    // Input: arbitrary Packets.
+    // Output: copy of the input.
+    cc->Outputs().Index(0).SetHeader(cc->Inputs().Index(0).Header());
+    shift_ = cc->InputSidePackets().Index(0).Get<TimestampDiff>();
+    cc->SetOffset(shift_);
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    cc->GetCounter("PassThrough")->Increment();
+    cc->Outputs().Index(0).AddPacket(
+        cc->Inputs().Index(0).Value().At(cc->InputTimestamp() + shift_));
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  TimestampDiff shift_;
+};
+REGISTER_CALCULATOR(TimeShiftCalculator);
+
+// A source calculator that alternates between outputting an integer (0, 1, 2,
+// ..., 100) and setting the next timestamp bound. The timestamps of the output
+// packets and next timestamp bounds are 0, 10, 20, 30, ...
+//
+//   T=0    Output 0
+//   T=10   Set timestamp bound
+//   T=20   Output 1
+//   T=30   Set timestamp bound
+//   ...
+//   T=2000 Output 100
+class OutputAndBoundSourceCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    counter_ = 0;
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    Timestamp timestamp(counter_);
+    if (counter_ % 20 == 0) {
+      cc->Outputs().Index(0).AddPacket(
+          MakePacket<int>(counter_ / 20).At(timestamp));
+    } else {
+      cc->Outputs().Index(0).SetNextTimestampBound(timestamp);
+    }
+    if (counter_ == 2000) {
+      return tool::StatusStop();
+    }
+    counter_ += 10;
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  int counter_;
+};
+REGISTER_CALCULATOR(OutputAndBoundSourceCalculator);
+
+// A calculator that outputs an initial packet of value 0 at time 0 in the
+// Open() method, and then delays each input packet by 20 time units in the
+// Process() method. The input stream and output stream have the integer type.
+class Delay20Calculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).Set<int>();
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    cc->SetOffset(TimestampDiff(20));
+    cc->Outputs().Index(0).AddPacket(MakePacket<int>(0).At(Timestamp(0)));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    const Packet& packet = cc->Inputs().Index(0).Value();
+    Timestamp timestamp = packet.Timestamp() + 20;
+    cc->Outputs().Index(0).AddPacket(packet.At(timestamp));
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(Delay20Calculator);
+
 class CustomBoundCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
@@ -45,8 +292,280 @@ class CustomBoundCalculator : public CalculatorBase {
 };
 REGISTER_CALCULATOR(CustomBoundCalculator);
 
+// Test that SetNextTimestampBound propagates.
+TEST(CalculatorGraph, SetNextTimestampBoundPropagation) {
+  CalculatorGraph graph;
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: 'in'
+        input_stream: 'gate'
+        node {
+          calculator: 'ValveCalculator'
+          input_stream: 'in'
+          input_stream: 'gate'
+          output_stream: 'gated'
+        }
+        node {
+          calculator: 'PassThroughCalculator'
+          input_stream: 'gated'
+          output_stream: 'passed'
+        }
+        node {
+          calculator: 'TimeShiftCalculator'
+          input_stream: 'passed'
+          output_stream: 'shifted'
+          input_side_packet: 'shift'
+        }
+        node {
+          calculator: 'MergeCalculator'
+          input_stream: 'in'
+          input_stream: 'shifted'
+          output_stream: 'merged'
+        }
+        node {
+          name: 'merged_output'
+          calculator: 'PassThroughCalculator'
+          input_stream: 'merged'
+          output_stream: 'out'
+        }
+      )");
+
+  Timestamp timestamp = Timestamp(0);
+  auto send_inputs = [&graph, &timestamp](int input, bool pass) {
+    ++timestamp;
+    MP_EXPECT_OK(graph.AddPacketToInputStream(
+        "in", MakePacket<int>(input).At(timestamp)));
+    MP_EXPECT_OK(graph.AddPacketToInputStream(
+        "gate", MakePacket<bool>(pass).At(timestamp)));
+  };
+
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.StartRun({{"shift", MakePacket<TimestampDiff>(0)}}));
+
+  auto pass_counter =
+      graph.GetCounterFactory()->GetCounter("ValveCalculator-PassThrough");
+  auto block_counter =
+      graph.GetCounterFactory()->GetCounter("ValveCalculator-Block");
+  auto merged_counter =
+      graph.GetCounterFactory()->GetCounter("merged_output-PassThrough");
+
+  send_inputs(1, true);
+  send_inputs(2, true);
+  send_inputs(3, false);
+  send_inputs(4, false);
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  // Verify that MergeCalculator was able to run even when the gated branch
+  // was blocked.
+  EXPECT_EQ(2, pass_counter->Get());
+  EXPECT_EQ(2, block_counter->Get());
+  EXPECT_EQ(4, merged_counter->Get());
+
+  send_inputs(5, true);
+  send_inputs(6, false);
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  EXPECT_EQ(3, pass_counter->Get());
+  EXPECT_EQ(3, block_counter->Get());
+  EXPECT_EQ(6, merged_counter->Get());
+
+  MP_ASSERT_OK(graph.CloseAllInputStreams());
+  MP_ASSERT_OK(graph.WaitUntilDone());
+
+  // Now test with time shift
+  MP_ASSERT_OK(graph.StartRun({{"shift", MakePacket<TimestampDiff>(-1)}}));
+
+  send_inputs(7, true);
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  // The merger should have run only once now, at timestamp 6, with inputs
+  // <null, 7>. If we do not respect the offset and unblock the merger for
+  // timestamp 7 too, then it will have run twice, with 6: <null,7> and
+  // 7: <7, null>.
+  EXPECT_EQ(4, pass_counter->Get());
+  EXPECT_EQ(3, block_counter->Get());
+  EXPECT_EQ(7, merged_counter->Get());
+
+  MP_ASSERT_OK(graph.CloseAllInputStreams());
+  MP_ASSERT_OK(graph.WaitUntilDone());
+
+  EXPECT_EQ(4, pass_counter->Get());
+  EXPECT_EQ(3, block_counter->Get());
+  EXPECT_EQ(8, merged_counter->Get());
+}
+
+// Both input streams of the calculator node have the same next timestamp
+// bound. One input stream has a packet at that timestamp. The other input
+// stream is empty. We should not run the Process() method of the node in this
+// case.
+TEST(CalculatorGraph, NotAllInputPacketsAtNextTimestampBoundAvailable) {
+  //
+  // in0_unfiltered  in1_to_be_filtered
+  //        |             |
+  //        |             V
+  //        |  +-----------------------+
+  //        |  |EvenIntFilterCalculator|
+  //        |  +-----------------------+
+  //        |             |
+  //        \             /
+  //         \           / in1_filtered
+  //          \         /
+  //           |       |
+  //           V       V
+  //      +------------------+
+  //      |IntAdderCalculator|
+  //      +------------------+
+  //               |
+  //               V
+  //              out
+  //
+  CalculatorGraph graph;
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: 'in0_unfiltered'
+        input_stream: 'in1_to_be_filtered'
+        node {
+          calculator: 'EvenIntFilterCalculator'
+          input_stream: 'in1_to_be_filtered'
+          output_stream: 'in1_filtered'
+        }
+        node {
+          calculator: 'IntAdderCalculator'
+          input_stream: 'in0_unfiltered'
+          input_stream: 'in1_filtered'
+          output_stream: 'out'
+        }
+      )");
+  std::vector<Packet> packet_dump;
+  tool::AddVectorSink("out", &config, &packet_dump);
+
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.StartRun({}));
+
+  Timestamp timestamp = Timestamp(0);
+
+  // We send an integer with timestamp 1 to the in0_unfiltered input stream of
+  // the IntAdderCalculator. We then send an even integer with timestamp 1 to
+  // the EvenIntFilterCalculator. This packet will go through and
+  // the IntAdderCalculator will run. The next timestamp bounds of both the
+  // input streams of the IntAdderCalculator will become 2.
+
+  ++timestamp;  // Timestamp 1.
+  MP_EXPECT_OK(graph.AddPacketToInputStream("in0_unfiltered",
+                                            MakePacket<int>(1).At(timestamp)));
+  MP_EXPECT_OK(graph.AddPacketToInputStream("in1_to_be_filtered",
+                                            MakePacket<int>(2).At(timestamp)));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  ASSERT_EQ(1, packet_dump.size());
+  EXPECT_EQ(3, packet_dump[0].Get<int>());
+
+  // We send an odd integer with timestamp 2 to the EvenIntFilterCalculator.
+  // This packet will be filtered out and the next timestamp bound of the
+  // in1_filtered input stream of the IntAdderCalculator will become 3.
+
+  ++timestamp;  // Timestamp 2.
+  MP_EXPECT_OK(graph.AddPacketToInputStream("in1_to_be_filtered",
+                                            MakePacket<int>(3).At(timestamp)));
+
+  // We send an integer with timestamp 3 to the in0_unfiltered input stream of
+  // the IntAdderCalculator. MediaPipe should propagate the next timestamp bound
+  // across the IntAdderCalculator but should not run its Process() method.
+
+  ++timestamp;  // Timestamp 3.
+  MP_EXPECT_OK(graph.AddPacketToInputStream("in0_unfiltered",
+                                            MakePacket<int>(3).At(timestamp)));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  ASSERT_EQ(1, packet_dump.size());
+
+  // We send an even integer with timestamp 3 to the IntAdderCalculator. This
+  // packet will go through and the IntAdderCalculator will run.
+
+  MP_EXPECT_OK(graph.AddPacketToInputStream("in1_to_be_filtered",
+                                            MakePacket<int>(4).At(timestamp)));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  ASSERT_EQ(2, packet_dump.size());
+  EXPECT_EQ(7, packet_dump[1].Get<int>());
+
+  MP_ASSERT_OK(graph.CloseAllInputStreams());
+  MP_ASSERT_OK(graph.WaitUntilDone());
+  EXPECT_EQ(2, packet_dump.size());
+}
+
+TEST(CalculatorGraph, PropagateBoundLoop) {
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        node {
+          calculator: 'OutputAndBoundSourceCalculator'
+          output_stream: 'integers'
+        }
+        node {
+          calculator: 'IntAdderCalculator'
+          input_stream: 'integers'
+          input_stream: 'old_sum'
+          input_stream_info: {
+            tag_index: ':1'  # 'old_sum'
+            back_edge: true
+          }
+          output_stream: 'sum'
+          input_stream_handler {
+            input_stream_handler: 'EarlyCloseInputStreamHandler'
+          }
+        }
+        node {
+          calculator: 'Delay20Calculator'
+          input_stream: 'sum'
+          output_stream: 'old_sum'
+        }
+      )");
+  std::vector<Packet> packet_dump;
+  tool::AddVectorSink("sum", &config, &packet_dump);
+
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.Run());
+  ASSERT_EQ(101, packet_dump.size());
+  int sum = 0;
+  for (int i = 0; i < 101; ++i) {
+    sum += i;
+    EXPECT_EQ(sum, packet_dump[i].Get<int>());
+    EXPECT_EQ(Timestamp(i * 20), packet_dump[i].Timestamp());
+  }
+}
+
+TEST(CalculatorGraph, CheckBatchProcessingBoundPropagation) {
+  // The timestamp bound sent by OutputAndBoundSourceCalculator shouldn't be
+  // directly propagated to the output stream when PassThroughCalculator has
+  // anything in its default calculator context for batch processing. Otherwise,
+  // the sink calculator's input stream should report packet timestamp
+  // mismatches.
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        node {
+          calculator: 'OutputAndBoundSourceCalculator'
+          output_stream: 'integers'
+        }
+        node {
+          calculator: 'PassThroughCalculator'
+          input_stream: 'integers'
+          output_stream: 'output'
+          input_stream_handler {
+            input_stream_handler: "DefaultInputStreamHandler"
+            options: {
+              [mediapipe.DefaultInputStreamHandlerOptions.ext]: {
+                batch_size: 10
+              }
+            }
+          }
+        }
+        node { calculator: 'IntSinkCalculator' input_stream: 'output' }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.Run());
+}
+
 // Shows that ImmediateInputStreamHandler allows bounds propagation.
-TEST(CalculatorGraphBounds, ImmediateHandlerBounds) {
+TEST(CalculatorGraphBoundsTest, ImmediateHandlerBounds) {
   // CustomBoundCalculator produces only timestamp bounds.
   // The first PassThroughCalculator propagates bounds using SetOffset(0).
   // The second PassthroughCalculator delivers an output packet whenever the
@@ -101,5 +620,261 @@ TEST(CalculatorGraphBounds, ImmediateHandlerBounds) {
   EXPECT_EQ(output_packets.size(), 4);
 }
 
+// A Calculator that only sets timestamp bound by SetOffset().
+class OffsetBoundCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).Set<int>();
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    cc->SetOffset(0);
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(OffsetBoundCalculator);
+
+// A Calculator that produces a packet for each call to Process.
+class BoundToPacketCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).Set<int>();
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    cc->Outputs().Index(0).AddPacket(Adopt(new int(33)));
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(BoundToPacketCalculator);
+
+// Verifies that SetOffset still propagates when Process is called and
+// produces no output packets.
+TEST(CalculatorGraphBoundsTest, OffsetBoundPropagation) {
+  // OffsetBoundCalculator produces only timestamp bounds.
+  // The PassthroughCalculator delivers an output packet whenever the
+  // OffsetBoundCalculator delivers a timestamp bound.
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: 'input'
+        node {
+          calculator: 'OffsetBoundCalculator'
+          input_stream: 'input'
+          output_stream: 'bounds'
+        }
+        node {
+          calculator: 'PassThroughCalculator'
+          input_stream: 'bounds'
+          input_stream: 'input'
+          output_stream: 'bounds_output'
+          output_stream: 'output'
+        }
+      )");
+  CalculatorGraph graph;
+  std::vector<Packet> output_packets;
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.ObserveOutputStream("output", [&](const Packet& p) {
+    output_packets.push_back(p);
+    return ::mediapipe::OkStatus();
+  }));
+  MP_ASSERT_OK(graph.StartRun({}));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  // Add four packets into the graph.
+  constexpr int kNumInputs = 4;
+  for (int i = 0; i < kNumInputs; ++i) {
+    Packet p = MakePacket<int>(33).At(Timestamp(i));
+    MP_ASSERT_OK(graph.AddPacketToInputStream("input", p));
+  }
+
+  // Four packets arrive at the output only if timestamp bounds are propagated.
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  EXPECT_EQ(output_packets.size(), kNumInputs);
+
+  // Shutdown the graph.
+  MP_ASSERT_OK(graph.CloseAllPacketSources());
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
+// Shows that bounds changes alone do not invoke Process.
+// Note: Bounds changes alone will invoke Process eventually
+// when SetOffset is cleared, see: go/mediapipe-realtime-graph.
+TEST(CalculatorGraphBoundsTest, BoundWithoutInputPackets) {
+  // OffsetBoundCalculator produces only timestamp bounds.
+  // The BoundToPacketCalculator delivers an output packet whenever the
+  // OffsetBoundCalculator delivers a timestamp bound.
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: 'input'
+        node {
+          calculator: 'OffsetBoundCalculator'
+          input_stream: 'input'
+          output_stream: 'bounds'
+        }
+        node {
+          calculator: 'BoundToPacketCalculator'
+          input_stream: 'bounds'
+          output_stream: 'output'
+        }
+      )");
+  CalculatorGraph graph;
+  std::vector<Packet> output_packets;
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.ObserveOutputStream("output", [&](const Packet& p) {
+    output_packets.push_back(p);
+    return ::mediapipe::OkStatus();
+  }));
+  MP_ASSERT_OK(graph.StartRun({}));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  // Add four packets into the graph.
+  constexpr int kNumInputs = 4;
+  for (int i = 0; i < kNumInputs; ++i) {
+    Packet p = MakePacket<int>(33).At(Timestamp(i));
+    MP_ASSERT_OK(graph.AddPacketToInputStream("input", p));
+  }
+
+  // No packets arrive, because updated timestamp bounds do not invoke
+  // BoundToPacketCalculator::Process.
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  EXPECT_EQ(output_packets.size(), 0);
+
+  // Shutdown the graph.
+  MP_ASSERT_OK(graph.CloseAllPacketSources());
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
+// Shows that when fixed-size-input-stream-hanlder drops packets,
+// no timetamp bounds are announced.
+TEST(CalculatorGraphBoundsTest, FixedSizeHandlerBounds) {
+  // LambdaCalculator with FixedSizeInputStreamHandler will drop packets
+  // while it is busy.  Timetamps for the dropped packets are only relevant
+  // when SetOffset is active on the LambdaCalculator.
+  // The PassthroughCalculator delivers an output packet whenever the
+  // LambdaCalculator delivers a timestamp bound.
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: 'input'
+        input_side_packet: 'open_function'
+        input_side_packet: 'process_function'
+        node {
+          calculator: 'LambdaCalculator'
+          input_stream: 'input'
+          output_stream: 'thinned'
+          input_side_packet: 'OPEN:open_fn'
+          input_side_packet: 'PROCESS:process_fn'
+          input_stream_handler {
+            input_stream_handler: "FixedSizeInputStreamHandler"
+          }
+        }
+        node {
+          calculator: 'PassThroughCalculator'
+          input_stream: 'thinned'
+          input_stream: 'input'
+          output_stream: 'thinned_output'
+          output_stream: 'output'
+        }
+      )");
+  CalculatorGraph graph;
+
+  // The task_semaphore counts the number of running tasks.
+  constexpr int kTaskSupply = 10;
+  AtomicSemaphore task_semaphore(/*supply=*/kTaskSupply);
+
+  // This executor invokes a callback at the start and finish of each task.
+  auto executor = std::make_shared<CountingExecutor>(
+      4, /*start_callback=*/[&]() { task_semaphore.Acquire(1); },
+      /*finish_callback=*/[&]() { task_semaphore.Release(1); });
+  MP_ASSERT_OK(graph.SetExecutor(/*name=*/"", executor));
+
+  // Monitor output from the graph.
+  MP_ASSERT_OK(graph.Initialize(config));
+  std::vector<Packet> outputs;
+  MP_ASSERT_OK(graph.ObserveOutputStream("output", [&](const Packet& p) {
+    outputs.push_back(p);
+    return ::mediapipe::OkStatus();
+  }));
+  std::vector<Packet> thinned_outputs;
+  MP_ASSERT_OK(
+      graph.ObserveOutputStream("thinned_output", [&](const Packet& p) {
+        thinned_outputs.push_back(p);
+        return ::mediapipe::OkStatus();
+      }));
+
+  // The enter_semaphore is used to wait for LambdaCalculator::Process.
+  // The exit_semaphore blocks and unblocks LambdaCalculator::Process.
+  AtomicSemaphore enter_semaphore(0);
+  AtomicSemaphore exit_semaphore(0);
+  CalculatorContextFunction open_fn = [&](CalculatorContext* cc) {
+    cc->SetOffset(0);
+    return ::mediapipe::OkStatus();
+  };
+  CalculatorContextFunction process_fn = [&](CalculatorContext* cc) {
+    enter_semaphore.Release(1);
+    exit_semaphore.Acquire(1);
+    cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+    return ::mediapipe::OkStatus();
+  };
+  MP_ASSERT_OK(graph.StartRun({
+      {"open_fn", Adopt(new auto(open_fn))},
+      {"process_fn", Adopt(new auto(process_fn))},
+  }));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  // Add four packets into the graph.
+  constexpr int kNumInputs = 4;
+  for (int i = 0; i < kNumInputs; ++i) {
+    Packet p = MakePacket<int>(33).At(Timestamp(i));
+    MP_ASSERT_OK(graph.AddPacketToInputStream("input", p));
+  }
+
+  // Wait until only the LambdaCalculator is running,
+  // by wating until the task_semaphore has only one token occupied.
+  // At this point 2 packets were dropped by the FixedSizeInputStreamHandler.
+  task_semaphore.Acquire(kTaskSupply - 1);
+  task_semaphore.Release(kTaskSupply - 1);
+
+  // No timestamp bounds and no packets are emitted yet.
+  EXPECT_EQ(outputs.size(), 0);
+  EXPECT_EQ(thinned_outputs.size(), 0);
+
+  // Allow the first LambdaCalculator::Process call to complete.
+  // Wait for the second LambdaCalculator::Process call to begin.
+  // Wait until only the LambdaCalculator is running.
+  enter_semaphore.Acquire(1);
+  exit_semaphore.Release(1);
+  enter_semaphore.Acquire(1);
+  task_semaphore.Acquire(kTaskSupply - 1);
+  task_semaphore.Release(kTaskSupply - 1);
+
+  // Only one timestamp bound and one packet are emitted.
+  EXPECT_EQ(outputs.size(), 1);
+  EXPECT_EQ(thinned_outputs.size(), 1);
+
+  // Allow the second LambdaCalculator::Process call to complete.
+  exit_semaphore.Release(1);
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  // Packets 1 and 2 were dropped by the FixedSizeInputStreamHandler.
+  EXPECT_EQ(thinned_outputs.size(), 2);
+  EXPECT_EQ(thinned_outputs[0].Timestamp(), Timestamp(0));
+  EXPECT_EQ(thinned_outputs[1].Timestamp(), Timestamp(kNumInputs - 1));
+  EXPECT_EQ(outputs.size(), kNumInputs);
+  MP_ASSERT_OK(graph.CloseAllPacketSources());
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
 }  // namespace
 }  // namespace mediapipe

mediapipe/framework/calculator_graph_side_packet_test.cc

@@ -0,0 +1,747 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+#include "mediapipe/framework/calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/canonical_errors.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/logging.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/status_matchers.h"
+
+namespace mediapipe {
+
+namespace {
+
+// Takes an input stream packet and passes it (with timestamp removed) as an
+// output side packet.
+class OutputSidePacketInProcessCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->OutputSidePackets().Index(0).SetSameAs(&cc->Inputs().Index(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    cc->OutputSidePackets().Index(0).Set(
+        cc->Inputs().Index(0).Value().At(Timestamp::Unset()));
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(OutputSidePacketInProcessCalculator);
+
+// Takes an input stream packet and counts the number of the packets it
+// receives. Outputs the total number of packets as a side packet in Close.
+class CountAndOutputSummarySidePacketInCloseCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->OutputSidePackets().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    ++count_;
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Close(CalculatorContext* cc) final {
+    cc->OutputSidePackets().Index(0).Set(
+        MakePacket<int>(count_).At(Timestamp::Unset()));
+    return ::mediapipe::OkStatus();
+  }
+
+  int count_ = 0;
+};
+REGISTER_CALCULATOR(CountAndOutputSummarySidePacketInCloseCalculator);
+
+// Takes an input stream packet and passes it (with timestamp intact) as an
+// output side packet. This triggers an error in the graph.
+class OutputSidePacketWithTimestampCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->OutputSidePackets().Index(0).SetSameAs(&cc->Inputs().Index(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    cc->OutputSidePackets().Index(0).Set(cc->Inputs().Index(0).Value());
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(OutputSidePacketWithTimestampCalculator);
+
+// Generates an output side packet containing the integer 1.
+class IntegerOutputSidePacketCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->OutputSidePackets().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    cc->OutputSidePackets().Index(0).Set(MakePacket<int>(1));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    LOG(FATAL) << "Not reached.";
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(IntegerOutputSidePacketCalculator);
+
+// Generates an output side packet containing the sum of the two integer input
+// side packets.
+class SidePacketAdderCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->InputSidePackets().Index(0).Set<int>();
+    cc->InputSidePackets().Index(1).Set<int>();
+    cc->OutputSidePackets().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    cc->OutputSidePackets().Index(0).Set(
+        MakePacket<int>(cc->InputSidePackets().Index(1).Get<int>() +
+                        cc->InputSidePackets().Index(0).Get<int>()));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    LOG(FATAL) << "Not reached.";
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(SidePacketAdderCalculator);
+
+// Produces an output packet with the PostStream timestamp containing the
+// input side packet.
+class SidePacketToStreamPacketCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->InputSidePackets().Index(0).SetAny();
+    cc->Outputs().Index(0).SetSameAs(&cc->InputSidePackets().Index(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) final {
+    cc->Outputs().Index(0).AddPacket(
+        cc->InputSidePackets().Index(0).At(Timestamp::PostStream()));
+    cc->Outputs().Index(0).Close();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    return ::mediapipe::tool::StatusStop();
+  }
+};
+REGISTER_CALCULATOR(SidePacketToStreamPacketCalculator);
+
+// Packet generator for an arbitrary unit64 packet.
+class Uint64PacketGenerator : public PacketGenerator {
+ public:
+  static ::mediapipe::Status FillExpectations(
+      const PacketGeneratorOptions& extendable_options,
+      PacketTypeSet* input_side_packets, PacketTypeSet* output_side_packets) {
+    output_side_packets->Index(0).Set<uint64>();
+    return ::mediapipe::OkStatus();
+  }
+
+  static ::mediapipe::Status Generate(
+      const PacketGeneratorOptions& extendable_options,
+      const PacketSet& input_side_packets, PacketSet* output_side_packets) {
+    output_side_packets->Index(0) = Adopt(new uint64(15LL << 32 | 5));
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_PACKET_GENERATOR(Uint64PacketGenerator);
+
+TEST(CalculatorGraph, OutputSidePacketInProcess) {
+  const int64 offset = 100;
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "offset"
+        node {
+          calculator: "OutputSidePacketInProcessCalculator"
+          input_stream: "offset"
+          output_side_packet: "offset"
+        }
+        node {
+          calculator: "SidePacketToStreamPacketCalculator"
+          output_stream: "output"
+          input_side_packet: "offset"
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  std::vector<Packet> output_packets;
+  MP_ASSERT_OK(graph.ObserveOutputStream(
+      "output", [&output_packets](const Packet& packet) {
+        output_packets.push_back(packet);
+        return ::mediapipe::OkStatus();
+      }));
+
+  // Run the graph twice.
+  for (int run = 0; run < 2; ++run) {
+    output_packets.clear();
+    MP_ASSERT_OK(graph.StartRun({}));
+    MP_ASSERT_OK(graph.AddPacketToInputStream(
+        "offset", MakePacket<TimestampDiff>(offset).At(Timestamp(0))));
+    MP_ASSERT_OK(graph.CloseInputStream("offset"));
+    MP_ASSERT_OK(graph.WaitUntilDone());
+    ASSERT_EQ(1, output_packets.size());
+    EXPECT_EQ(offset, output_packets[0].Get<TimestampDiff>().Value());
+  }
+}
+
+// A PacketGenerator that simply passes its input Packets through
+// unchanged.  The inputs may be specified by tag or index.  The outputs
+// must match the inputs exactly.  Any options may be specified and will
+// also be ignored.
+class PassThroughGenerator : public PacketGenerator {
+ public:
+  static ::mediapipe::Status FillExpectations(
+      const PacketGeneratorOptions& extendable_options, PacketTypeSet* inputs,
+      PacketTypeSet* outputs) {
+    if (!inputs->TagMap()->SameAs(*outputs->TagMap())) {
+      return ::mediapipe::InvalidArgumentError(
+          "Input and outputs to PassThroughGenerator must use the same tags "
+          "and indexes.");
+    }
+    for (CollectionItemId id = inputs->BeginId(); id < inputs->EndId(); ++id) {
+      inputs->Get(id).SetAny();
+      outputs->Get(id).SetSameAs(&inputs->Get(id));
+    }
+    return ::mediapipe::OkStatus();
+  }
+
+  static ::mediapipe::Status Generate(
+      const PacketGeneratorOptions& extendable_options,
+      const PacketSet& input_side_packets, PacketSet* output_side_packets) {
+    for (CollectionItemId id = input_side_packets.BeginId();
+         id < input_side_packets.EndId(); ++id) {
+      output_side_packets->Get(id) = input_side_packets.Get(id);
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_PACKET_GENERATOR(PassThroughGenerator);
+
+TEST(CalculatorGraph, SharePacketGeneratorGraph) {
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        node {
+          calculator: 'CountingSourceCalculator'
+          output_stream: 'count1'
+          input_side_packet: 'MAX_COUNT:max_count1'
+        }
+        node {
+          calculator: 'CountingSourceCalculator'
+          output_stream: 'count2'
+          input_side_packet: 'MAX_COUNT:max_count2'
+        }
+        node {
+          calculator: 'CountingSourceCalculator'
+          output_stream: 'count3'
+          input_side_packet: 'MAX_COUNT:max_count3'
+        }
+        node {
+          calculator: 'CountingSourceCalculator'
+          output_stream: 'count4'
+          input_side_packet: 'MAX_COUNT:max_count4'
+        }
+        node {
+          calculator: 'PassThroughCalculator'
+          input_side_packet: 'MAX_COUNT:max_count5'
+          output_side_packet: 'MAX_COUNT:max_count6'
+        }
+        node {
+          calculator: 'CountingSourceCalculator'
+          output_stream: 'count5'
+          input_side_packet: 'MAX_COUNT:max_count6'
+        }
+        packet_generator {
+          packet_generator: 'PassThroughGenerator'
+          input_side_packet: 'max_count1'
+          output_side_packet: 'max_count2'
+        }
+        packet_generator {
+          packet_generator: 'PassThroughGenerator'
+          input_side_packet: 'max_count4'
+          output_side_packet: 'max_count5'
+        }
+      )");
+
+  // At this point config is a standard config which specifies both
+  // calculators and packet_factories/packet_genators.  The following
+  // code is an example of reusing side packets across a number of
+  // CalculatorGraphs.  It is particularly informative to note how each
+  // side packet is created.
+  //
+  // max_count1 is set for all graphs by a PacketFactory in the config.
+  // The side packet is created by generator_graph.InitializeGraph().
+  //
+  // max_count2 is set for all graphs by a PacketGenerator in the config.
+  // The side packet is created by generator_graph.InitializeGraph()
+  // because max_count1 is available at that time.
+  //
+  // max_count3 is set for all graphs by directly being specified as an
+  // argument to generator_graph.InitializeGraph().
+  //
+  // max_count4 is set per graph because it is directly specified as an
+  // argument to generator_graph.ProcessGraph().
+  //
+  // max_count5 is set per graph by a PacketGenerator which is run when
+  // generator_graph.ProcessGraph() is run (because max_count4 isn't
+  // available until then).
+
+  // Before anything else, split the graph config into two parts, one
+  // with the PacketFactory and PacketGenerator config and the other
+  // with the Calculator config.
+  CalculatorGraphConfig calculator_config = config;
+  calculator_config.clear_packet_factory();
+  calculator_config.clear_packet_generator();
+  CalculatorGraphConfig generator_config = config;
+  generator_config.clear_node();
+
+  // Next, create a ValidatedGraphConfig for both configs.
+  ValidatedGraphConfig validated_calculator_config;
+  MP_ASSERT_OK(validated_calculator_config.Initialize(calculator_config));
+  ValidatedGraphConfig validated_generator_config;
+  MP_ASSERT_OK(validated_generator_config.Initialize(generator_config));
+
+  // Create a PacketGeneratorGraph.  Side packets max_count1, max_count2,
+  // and max_count3 are created upon initialization.
+  // Note that validated_generator_config must outlive generator_graph.
+  PacketGeneratorGraph generator_graph;
+  MP_ASSERT_OK(
+      generator_graph.Initialize(&validated_generator_config, nullptr,
+                                 {{"max_count1", MakePacket<int>(10)},
+                                  {"max_count3", MakePacket<int>(20)}}));
+  ASSERT_THAT(generator_graph.BasePackets(),
+              testing::ElementsAre(testing::Key("max_count1"),
+                                   testing::Key("max_count2"),
+                                   testing::Key("max_count3")));
+
+  // Create a bunch of graphs.
+  std::vector<std::unique_ptr<CalculatorGraph>> graphs;
+  for (int i = 0; i < 100; ++i) {
+    graphs.emplace_back(absl::make_unique<CalculatorGraph>());
+    // Do not pass extra side packets here.
+    // Note that validated_calculator_config must outlive the graph.
+    MP_ASSERT_OK(graphs.back()->Initialize(calculator_config, {}));
+  }
+  // Run a bunch of graphs, reusing side packets max_count1, max_count2,
+  // and max_count3.  The side packet max_count4 is added per run,
+  // and triggers the execution of a packet generator which generates
+  // max_count5.
+  for (int i = 0; i < 100; ++i) {
+    std::map<std::string, Packet> all_side_packets;
+    // Creates max_count4 and max_count5.
+    MP_ASSERT_OK(generator_graph.RunGraphSetup(
+        {{"max_count4", MakePacket<int>(30 + i)}}, &all_side_packets));
+    ASSERT_THAT(all_side_packets,
+                testing::ElementsAre(
+                    testing::Key("max_count1"), testing::Key("max_count2"),
+                    testing::Key("max_count3"), testing::Key("max_count4"),
+                    testing::Key("max_count5")));
+    // Pass all the side packets prepared by generator_graph here.
+    MP_ASSERT_OK(graphs[i]->Run(all_side_packets));
+    // TODO Verify the actual output.
+  }
+
+  // Destroy all the graphs.
+  graphs.clear();
+}
+
+TEST(CalculatorGraph, OutputSidePacketAlreadySet) {
+  const int64 offset = 100;
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "offset"
+        node {
+          calculator: "OutputSidePacketInProcessCalculator"
+          input_stream: "offset"
+          output_side_packet: "offset"
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.StartRun({}));
+  // Send two input packets to cause OutputSidePacketInProcessCalculator to
+  // set the output side packet twice.
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "offset", MakePacket<TimestampDiff>(offset).At(Timestamp(0))));
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "offset", MakePacket<TimestampDiff>(offset).At(Timestamp(1))));
+  MP_ASSERT_OK(graph.CloseInputStream("offset"));
+
+  ::mediapipe::Status status = graph.WaitUntilDone();
+  EXPECT_EQ(status.code(), ::mediapipe::StatusCode::kAlreadyExists);
+  EXPECT_THAT(status.message(), testing::HasSubstr("was already set."));
+}
+
+TEST(CalculatorGraph, OutputSidePacketWithTimestamp) {
+  const int64 offset = 100;
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "offset"
+        node {
+          calculator: "OutputSidePacketWithTimestampCalculator"
+          input_stream: "offset"
+          output_side_packet: "offset"
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  MP_ASSERT_OK(graph.StartRun({}));
+  // The OutputSidePacketWithTimestampCalculator neglects to clear the
+  // timestamp in the input packet when it copies the input packet to the
+  // output side packet. The timestamp value should appear in the error
+  // message.
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "offset", MakePacket<TimestampDiff>(offset).At(Timestamp(237))));
+  MP_ASSERT_OK(graph.CloseInputStream("offset"));
+  ::mediapipe::Status status = graph.WaitUntilDone();
+  EXPECT_EQ(status.code(), ::mediapipe::StatusCode::kInvalidArgument);
+  EXPECT_THAT(status.message(), testing::HasSubstr("has a timestamp 237."));
+}
+
+TEST(CalculatorGraph, OutputSidePacketConsumedBySourceNode) {
+  const int max_count = 10;
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "max_count"
+        node {
+          calculator: "OutputSidePacketInProcessCalculator"
+          input_stream: "max_count"
+          output_side_packet: "max_count"
+        }
+        node {
+          calculator: "CountingSourceCalculator"
+          output_stream: "count"
+          input_side_packet: "MAX_COUNT:max_count"
+        }
+        node {
+          calculator: "PassThroughCalculator"
+          input_stream: "count"
+          output_stream: "output"
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  std::vector<Packet> output_packets;
+  MP_ASSERT_OK(graph.ObserveOutputStream(
+      "output", [&output_packets](const Packet& packet) {
+        output_packets.push_back(packet);
+        return ::mediapipe::OkStatus();
+      }));
+  MP_ASSERT_OK(graph.StartRun({}));
+  // Wait until the graph is idle so that
+  // Scheduler::TryToScheduleNextSourceLayer() gets called.
+  // Scheduler::TryToScheduleNextSourceLayer() should not activate source
+  // nodes that haven't been opened. We can't call graph.WaitUntilIdle()
+  // because the graph has a source node.
+  absl::SleepFor(absl::Milliseconds(10));
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "max_count", MakePacket<int>(max_count).At(Timestamp(0))));
+  MP_ASSERT_OK(graph.CloseInputStream("max_count"));
+  MP_ASSERT_OK(graph.WaitUntilDone());
+  ASSERT_EQ(max_count, output_packets.size());
+  for (int i = 0; i < output_packets.size(); ++i) {
+    EXPECT_EQ(i, output_packets[i].Get<int>());
+    EXPECT_EQ(Timestamp(i), output_packets[i].Timestamp());
+  }
+}
+
+// Returns the first packet of the input stream.
+class FirstPacketFilterCalculator : public CalculatorBase {
+ public:
+  FirstPacketFilterCalculator() {}
+  ~FirstPacketFilterCalculator() override {}
+
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->Outputs().Index(0).SetSameAs(&cc->Inputs().Index(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    if (!seen_first_packet_) {
+      cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+      cc->Outputs().Index(0).Close();
+      seen_first_packet_ = true;
+    }
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  bool seen_first_packet_ = false;
+};
+REGISTER_CALCULATOR(FirstPacketFilterCalculator);
+
+TEST(CalculatorGraph, SourceLayerInversion) {
+  // There are three CountingSourceCalculators, indexed 0, 1, and 2. Each of
+  // them outputs 10 packets.
+  //
+  // CountingSourceCalculator 0 should output 0, 1, 2, 3, ..., 9.
+  // CountingSourceCalculator 1 should output 100, 101, 102, 103, ..., 109.
+  // CountingSourceCalculator 2 should output 0, 100, 200, 300, ..., 900.
+  // However, there is a source layer inversion.
+  // CountingSourceCalculator 0 is in source layer 0.
+  // CountingSourceCalculator 1 is in source layer 1.
+  // CountingSourceCalculator 2 is in source layer 0, but consumes an output
+  // side packet generated by a downstream calculator of
+  // CountingSourceCalculator 1.
+  //
+  // This graph will deadlock when CountingSourceCalculator 0 runs to
+  // completion and CountingSourceCalculator 1 cannot be activated because
+  // CountingSourceCalculator 2 cannot be opened.
+
+  const int max_count = 10;
+  const int initial_value1 = 100;
+  // Set num_threads to 1 to force sequential execution for deterministic
+  // outputs.
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        num_threads: 1
+        node {
+          calculator: "CountingSourceCalculator"
+          output_stream: "count0"
+          input_side_packet: "MAX_COUNT:max_count"
+          source_layer: 0
+        }
+
+        node {
+          calculator: "CountingSourceCalculator"
+          output_stream: "count1"
+          input_side_packet: "MAX_COUNT:max_count"
+          input_side_packet: "INITIAL_VALUE:initial_value1"
+          source_layer: 1
+        }
+        node {
+          calculator: "FirstPacketFilterCalculator"
+          input_stream: "count1"
+          output_stream: "first_count1"
+        }
+        node {
+          calculator: "OutputSidePacketInProcessCalculator"
+          input_stream: "first_count1"
+          output_side_packet: "increment2"
+        }
+
+        node {
+          calculator: "CountingSourceCalculator"
+          output_stream: "count2"
+          input_side_packet: "MAX_COUNT:max_count"
+          input_side_packet: "INCREMENT:increment2"
+          source_layer: 0
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(
+      config, {{"max_count", MakePacket<int>(max_count)},
+               {"initial_value1", MakePacket<int>(initial_value1)}}));
+  ::mediapipe::Status status = graph.Run();
+  EXPECT_EQ(status.code(), ::mediapipe::StatusCode::kUnknown);
+  EXPECT_THAT(status.message(), testing::HasSubstr("deadlock"));
+}
+
+// Tests a graph of packet-generator-like calculators, which have no input
+// streams and no output streams.
+TEST(CalculatorGraph, PacketGeneratorLikeCalculators) {
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        node {
+          calculator: "IntegerOutputSidePacketCalculator"
+          output_side_packet: "one"
+        }
+        node {
+          calculator: "IntegerOutputSidePacketCalculator"
+          output_side_packet: "another_one"
+        }
+        node {
+          calculator: "SidePacketAdderCalculator"
+          input_side_packet: "one"
+          input_side_packet: "another_one"
+          output_side_packet: "two"
+        }
+        node {
+          calculator: "IntegerOutputSidePacketCalculator"
+          output_side_packet: "yet_another_one"
+        }
+        node {
+          calculator: "SidePacketAdderCalculator"
+          input_side_packet: "two"
+          input_side_packet: "yet_another_one"
+          output_side_packet: "three"
+        }
+        node {
+          calculator: "SidePacketToStreamPacketCalculator"
+          input_side_packet: "three"
+          output_stream: "output"
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  std::vector<Packet> output_packets;
+  MP_ASSERT_OK(graph.ObserveOutputStream(
+      "output", [&output_packets](const Packet& packet) {
+        output_packets.push_back(packet);
+        return ::mediapipe::OkStatus();
+      }));
+  MP_ASSERT_OK(graph.Run());
+  ASSERT_EQ(1, output_packets.size());
+  EXPECT_EQ(3, output_packets[0].Get<int>());
+  EXPECT_EQ(Timestamp::PostStream(), output_packets[0].Timestamp());
+}
+
+TEST(CalculatorGraph, OutputSummarySidePacketInClose) {
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "input_packets"
+        node {
+          calculator: "CountAndOutputSummarySidePacketInCloseCalculator"
+          input_stream: "input_packets"
+          output_side_packet: "num_of_packets"
+        }
+        node {
+          calculator: "SidePacketToStreamPacketCalculator"
+          input_side_packet: "num_of_packets"
+          output_stream: "output"
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  std::vector<Packet> output_packets;
+  MP_ASSERT_OK(graph.ObserveOutputStream(
+      "output", [&output_packets](const Packet& packet) {
+        output_packets.push_back(packet);
+        return ::mediapipe::OkStatus();
+      }));
+
+  // Run the graph twice.
+  int max_count = 100;
+  for (int run = 0; run < 1; ++run) {
+    output_packets.clear();
+    MP_ASSERT_OK(graph.StartRun({}));
+    for (int i = 0; i < max_count; ++i) {
+      MP_ASSERT_OK(graph.AddPacketToInputStream(
+          "input_packets", MakePacket<int>(i).At(Timestamp(i))));
+    }
+    MP_ASSERT_OK(graph.CloseInputStream("input_packets"));
+    MP_ASSERT_OK(graph.WaitUntilDone());
+    ASSERT_EQ(1, output_packets.size());
+    EXPECT_EQ(max_count, output_packets[0].Get<int>());
+    EXPECT_EQ(Timestamp::PostStream(), output_packets[0].Timestamp());
+  }
+}
+
+TEST(CalculatorGraph, GetOutputSidePacket) {
+  CalculatorGraphConfig config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "input_packets"
+        node {
+          calculator: "CountAndOutputSummarySidePacketInCloseCalculator"
+          input_stream: "input_packets"
+          output_side_packet: "num_of_packets"
+        }
+        packet_generator {
+          packet_generator: "Uint64PacketGenerator"
+          output_side_packet: "output_uint64"
+        }
+        packet_generator {
+          packet_generator: "IntSplitterPacketGenerator"
+          input_side_packet: "input_uint64"
+          output_side_packet: "output_uint32_pair"
+        }
+      )");
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(config));
+  // Check a packet generated by the PacketGenerator, which is available after
+  // graph initialization, can be fetched before graph starts.
+  ::mediapipe::StatusOr<Packet> status_or_packet =
+      graph.GetOutputSidePacket("output_uint64");
+  MP_ASSERT_OK(status_or_packet);
+  EXPECT_EQ(Timestamp::Unset(), status_or_packet.ValueOrDie().Timestamp());
+  // IntSplitterPacketGenerator is missing its input side packet and we
+  // won't be able to get its output side packet now.
+  status_or_packet = graph.GetOutputSidePacket("output_uint32_pair");
+  EXPECT_EQ(::mediapipe::StatusCode::kUnavailable,
+            status_or_packet.status().code());
+  // Run the graph twice.
+  int max_count = 100;
+  std::map<std::string, Packet> extra_side_packets;
+  extra_side_packets.insert({"input_uint64", MakePacket<uint64>(1123)});
+  for (int run = 0; run < 1; ++run) {
+    MP_ASSERT_OK(graph.StartRun(extra_side_packets));
+    status_or_packet = graph.GetOutputSidePacket("output_uint32_pair");
+    MP_ASSERT_OK(status_or_packet);
+    EXPECT_EQ(Timestamp::Unset(), status_or_packet.ValueOrDie().Timestamp());
+    for (int i = 0; i < max_count; ++i) {
+      MP_ASSERT_OK(graph.AddPacketToInputStream(
+          "input_packets", MakePacket<int>(i).At(Timestamp(i))));
+    }
+    MP_ASSERT_OK(graph.CloseInputStream("input_packets"));
+
+    // Should return NOT_FOUND for invalid side packets.
+    status_or_packet = graph.GetOutputSidePacket("unknown");
+    EXPECT_FALSE(status_or_packet.ok());
+    EXPECT_EQ(::mediapipe::StatusCode::kNotFound,
+              status_or_packet.status().code());
+    // Should return UNAVAILABLE before graph is done for valid non-base
+    // packets.
+    status_or_packet = graph.GetOutputSidePacket("num_of_packets");
+    EXPECT_FALSE(status_or_packet.ok());
+    EXPECT_EQ(::mediapipe::StatusCode::kUnavailable,
+              status_or_packet.status().code());
+    // Should stil return a base even before graph is done.
+    status_or_packet = graph.GetOutputSidePacket("output_uint64");
+    MP_ASSERT_OK(status_or_packet);
+    EXPECT_EQ(Timestamp::Unset(), status_or_packet.ValueOrDie().Timestamp());
+
+    MP_ASSERT_OK(graph.WaitUntilDone());
+
+    // Check packets are available after graph is done.
+    status_or_packet = graph.GetOutputSidePacket("num_of_packets");
+    MP_ASSERT_OK(status_or_packet);
+    EXPECT_EQ(max_count, status_or_packet.ValueOrDie().Get<int>());
+    EXPECT_EQ(Timestamp::Unset(), status_or_packet.ValueOrDie().Timestamp());
+    // Should still return a base packet after graph is done.
+    status_or_packet = graph.GetOutputSidePacket("output_uint64");
+    MP_ASSERT_OK(status_or_packet);
+    EXPECT_EQ(Timestamp::Unset(), status_or_packet.ValueOrDie().Timestamp());
+    // Should still return a non-base packet after graph is done.
+    status_or_packet = graph.GetOutputSidePacket("output_uint32_pair");
+    MP_ASSERT_OK(status_or_packet);
+    EXPECT_EQ(Timestamp::Unset(), status_or_packet.ValueOrDie().Timestamp());
+  }
+}
+
+}  // namespace
+}  // namespace mediapipe

mediapipe/framework/formats/matrix_data.proto

@@ -34,7 +34,7 @@ message MatrixData {
     ROW_MAJOR = 1;
   }
 
-  // Order in which the data are stored. Implicitly defaults to COLUMN_MAJOR,
+  // Order in which the data are stored. Defaults to COLUMN_MAJOR, which matches
-  // which matches the default for mediapipe::Matrix and Eigen::Matrix*.
+  // the default for mediapipe::Matrix and Eigen::Matrix*.
-  optional Layout layout = 4;
+  optional Layout layout = 4 [default = COLUMN_MAJOR];
 }

mediapipe/framework/graph_validation_test.cc

@@ -154,6 +154,7 @@ TEST(ValidatedGraphConfigTest, InitializeTemplateFromProtos) {
     }
   )");
   auto options = ParseTextProtoOrDie<Subgraph::SubgraphOptions>(R"(
+    options: {
       [mediapipe.TemplateSubgraphOptions.ext]: {
         dict: {
           arg: {
@@ -161,6 +162,7 @@ TEST(ValidatedGraphConfigTest, InitializeTemplateFromProtos) {
             value: { str: "stream_9" }
           }
         }
+      }
     })");
 
   GraphValidation validation_1;

mediapipe/framework/subgraph.cc

@@ -44,8 +44,8 @@ TemplateSubgraph::~TemplateSubgraph() {}
 
 ::mediapipe::StatusOr<CalculatorGraphConfig> TemplateSubgraph::GetConfig(
     const Subgraph::SubgraphOptions& options) {
-  const TemplateDict& arguments =
+  TemplateDict arguments =
-      options.GetExtension(TemplateSubgraphOptions::ext).dict();
+      Subgraph::GetOptions<mediapipe::TemplateSubgraphOptions>(options).dict();
   tool::TemplateExpander expander;
   CalculatorGraphConfig config;
   MP_RETURN_IF_ERROR(expander.ExpandTemplates(arguments, templ_, &config));

mediapipe/framework/subgraph.h

@@ -24,6 +24,7 @@
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/framework/port/statusor.h"
 #include "mediapipe/framework/tool/calculator_graph_template.pb.h"
+#include "mediapipe/framework/tool/options_util.h"
 
 namespace mediapipe {
 
@@ -32,7 +33,7 @@ namespace mediapipe {
 // the graph is running.
 class Subgraph {
  public:
-  using SubgraphOptions = CalculatorOptions;
+  using SubgraphOptions = CalculatorGraphConfig::Node;
   Subgraph();
   virtual ~Subgraph();
   // Returns the config to use for one instantiation of the subgraph. The
@@ -42,6 +43,12 @@ class Subgraph {
   // TODO: make this static?
   virtual ::mediapipe::StatusOr<CalculatorGraphConfig> GetConfig(
       const SubgraphOptions& options) = 0;
+
+  // Returns options of a specific type.
+  template <typename T>
+  static T GetOptions(Subgraph::SubgraphOptions supgraph_options) {
+    return tool::OptionsMap().Initialize(supgraph_options).Get<T>();
+  }
 };
 
 using SubgraphRegistry = GlobalFactoryRegistry<std::unique_ptr<Subgraph>>;

mediapipe/framework/test_calculators.cc

@@ -548,8 +548,12 @@ typedef std::function<::mediapipe::Status(const InputStreamShardSet&,
                                           OutputStreamShardSet*)>
     ProcessFunction;
 
+// A callback function for Calculator::Open, Process, or Close.
+typedef std::function<::mediapipe::Status(CalculatorContext* cc)>
+    CalculatorContextFunction;
+
 // A Calculator that runs a testing callback function in Process,
-// which is specified as an input side packet.
+// Open, or Close, which is specified as an input side packet.
 class LambdaCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
@@ -561,21 +565,49 @@ class LambdaCalculator : public CalculatorBase {
          id < cc->Outputs().EndId(); ++id) {
       cc->Outputs().Get(id).SetAny();
     }
-    cc->InputSidePackets().Index(0).Set<ProcessFunction>();
+    if (cc->InputSidePackets().HasTag("") > 0) {
+      cc->InputSidePackets().Tag("").Set<ProcessFunction>();
+    }
+    for (std::string tag : {"OPEN", "PROCESS", "CLOSE"}) {
+      if (cc->InputSidePackets().HasTag(tag)) {
+        cc->InputSidePackets().Tag(tag).Set<CalculatorContextFunction>();
+      }
+    }
     return ::mediapipe::OkStatus();
   }
 
   ::mediapipe::Status Open(CalculatorContext* cc) final {
-    callback_ = cc->InputSidePackets().Index(0).Get<ProcessFunction>();
+    if (cc->InputSidePackets().HasTag("OPEN")) {
+      return GetContextFn(cc, "OPEN")(cc);
+    }
     return ::mediapipe::OkStatus();
   }
 
   ::mediapipe::Status Process(CalculatorContext* cc) final {
-    return callback_(cc->Inputs(), &(cc->Outputs()));
+    if (cc->InputSidePackets().HasTag("PROCESS")) {
+      return GetContextFn(cc, "PROCESS")(cc);
+    }
+    if (cc->InputSidePackets().HasTag("") > 0) {
+      return GetProcessFn(cc, "")(cc->Inputs(), &cc->Outputs());
+    }
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Close(CalculatorContext* cc) final {
+    if (cc->InputSidePackets().HasTag("CLOSE")) {
+      return GetContextFn(cc, "CLOSE")(cc);
+    }
+    return ::mediapipe::OkStatus();
   }
 
  private:
-  ProcessFunction callback_;
+  ProcessFunction GetProcessFn(CalculatorContext* cc, std::string tag) {
+    return cc->InputSidePackets().Tag(tag).Get<ProcessFunction>();
+  }
+  CalculatorContextFunction GetContextFn(CalculatorContext* cc,
+                                         std::string tag) {
+    return cc->InputSidePackets().Tag(tag).Get<CalculatorContextFunction>();
+  }
 };
 REGISTER_CALCULATOR(LambdaCalculator);
 

mediapipe/framework/testdata/BUILD

@@ -55,6 +55,14 @@ proto_library(
     deps = ["@com_google_protobuf//:any_proto"],
 )
 
+mediapipe_cc_proto_library(
+    name = "zoo_mutator_cc_proto",
+    srcs = ["zoo_mutator.proto"],
+    cc_deps = ["@com_google_protobuf//:cc_wkt_protos"],
+    visibility = ["//mediapipe:__subpackages__"],
+    deps = [":zoo_mutator_proto"],
+)
+
 proto_library(
     name = "zoo_mutation_calculator_proto",
     srcs = ["zoo_mutation_calculator.proto"],

mediapipe/framework/tool/subgraph_expansion.cc

@@ -237,9 +237,9 @@ static ::mediapipe::Status PrefixNames(int subgraph_index,
     for (auto it = subgraph_nodes_start; it != nodes->end(); ++it) {
       const auto& node = *it;
       MP_RETURN_IF_ERROR(ValidateSubgraphFields(node));
-      ASSIGN_OR_RETURN(auto subgraph, graph_registry->CreateByName(
+      ASSIGN_OR_RETURN(auto subgraph,
-                                          config->package(), node.calculator(),
+                       graph_registry->CreateByName(config->package(),
-                                          &node.options()));
+                                                    node.calculator(), &node));
       MP_RETURN_IF_ERROR(PrefixNames(subgraph_counter++, &subgraph));
       MP_RETURN_IF_ERROR(ConnectSubgraphStreams(node, &subgraph));
       subgraphs.push_back(subgraph);

mediapipe/framework/tool/subgraph_expansion_test.cc

@@ -128,8 +128,8 @@ class NodeChainSubgraph : public Subgraph {
  public:
   ::mediapipe::StatusOr<CalculatorGraphConfig> GetConfig(
       const SubgraphOptions& options) override {
-    const mediapipe::NodeChainSubgraphOptions& opts =
+    auto opts =
-        options.GetExtension(mediapipe::NodeChainSubgraphOptions::ext);
+        Subgraph::GetOptions<mediapipe::NodeChainSubgraphOptions>(options);
     const ProtoString& node_type = opts.node_type();
     int chain_length = opts.chain_length();
     RET_CHECK(!node_type.empty());

mediapipe/gpu/BUILD

@@ -80,8 +80,7 @@ GL_BASE_LINK_OPTS_OSS = GL_BASE_LINK_OPTS + select({
         "-lEGL",
     ],
     "//mediapipe:android": [],
-    "//mediapipe:apple": [],
+    "//mediapipe:ios": [],
-    "//mediapipe:macos": [],
     ":disable_gpu": [],
 })
 
@@ -289,6 +288,25 @@ objc_library(
     ],
 )
 
+objc_library(
+    name = "MPPMetalUtil",
+    srcs = ["MPPMetalUtil.mm"],
+    hdrs = ["MPPMetalUtil.h"],
+    copts = [
+        "-x objective-c++",
+        "-Wno-shorten-64-to-32",
+    ],
+    sdk_frameworks = [
+        "CoreVideo",
+        "Metal",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/objc:mediapipe_framework_ios",
+        "@google_toolbox_for_mac//:GTM_Defines",
+    ],
+)
+
 proto_library(
     name = "gl_context_options_proto",
     srcs = ["gl_context_options.proto"],
@@ -499,6 +517,7 @@ cc_library(
         ":gl_base",
         ":gl_context",
         ":gpu_buffer",
+        ":gpu_buffer_format",
         ":gpu_buffer_multi_pool",
         ":gpu_shared_data_internal",
         ":gpu_service",

mediapipe/gpu/MPPMetalUtil.h

@@ -0,0 +1,49 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef MEDIAPIPE_GPU_MPP_METAL_UTIL_H_
+#define MEDIAPIPE_GPU_MPP_METAL_UTIL_H_
+
+#import <CoreVideo/CVMetalTextureCache.h>
+#import <CoreVideo/CoreVideo.h>
+#import <Metal/Metal.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+@interface MPPMetalUtil : NSObject {
+}
+
+/// Copies a Metal Buffer from source to destination.
+/// Uses blitCommandEncoder and assumes offset of 0.
++ (void)blitMetalBufferTo:(id<MTLBuffer>)destination
+                     from:(id<MTLBuffer>)source
+                 blocking:(bool)blocking
+            commandBuffer:(id<MTLCommandBuffer>)commandBuffer;
+
+/// Copies a Metal Buffer from source to destination.
+/// Simple wrapper for blitCommandEncoder.
+/// Optionally block until operation is completed.
++ (void)blitMetalBufferTo:(id<MTLBuffer>)destination
+        destinationOffset:(int)destinationOffset
+                     from:(id<MTLBuffer>)source
+             sourceOffset:(int)sourceOffset
+                    bytes:(size_t)bytes
+                 blocking:(bool)blocking
+            commandBuffer:(id<MTLCommandBuffer>)commandBuffer;
+
+@end
+
+NS_ASSUME_NONNULL_END
+
+#endif  // MEDIAPIPE_GPU_MPP_METAL_UTIL_H_

mediapipe/gpu/MPPMetalUtil.mm

@@ -0,0 +1,51 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#import "mediapipe/gpu/MPPMetalUtil.h"
+
+@implementation MPPMetalUtil
+
++ (void)blitMetalBufferTo:(id<MTLBuffer>)destination
+                     from:(id<MTLBuffer>)source
+                 blocking:(bool)blocking
+            commandBuffer:(id<MTLCommandBuffer>)commandBuffer {
+  size_t bytes = MIN(destination.length, source.length);
+  [self blitMetalBufferTo:destination
+        destinationOffset:0
+                     from:source
+             sourceOffset:0
+                    bytes:bytes
+                 blocking:blocking
+            commandBuffer:commandBuffer];
+}
+
++ (void)blitMetalBufferTo:(id<MTLBuffer>)destination
+        destinationOffset:(int)destinationOffset
+                     from:(id<MTLBuffer>)source
+             sourceOffset:(int)sourceOffset
+                    bytes:(size_t)bytes
+                 blocking:(bool)blocking
+            commandBuffer:(id<MTLCommandBuffer>)commandBuffer {
+  id<MTLBlitCommandEncoder> blit_command = [commandBuffer blitCommandEncoder];
+  [blit_command copyFromBuffer:source
+                  sourceOffset:sourceOffset
+                      toBuffer:destination
+             destinationOffset:destinationOffset
+                          size:bytes];
+  [blit_command endEncoding];
+  [commandBuffer commit];
+  if (blocking) [commandBuffer waitUntilCompleted];
+}
+
+@end

mediapipe/gpu/gl_calculator_helper_impl.h

@@ -73,7 +73,7 @@ class GlCalculatorHelperImpl {
 #endif  // !MEDIAPIPE_GPU_BUFFER_USE_CV_PIXEL_BUFFER
 
   // Sets default texture filtering parameters.
-  void SetStandardTextureParams(GLenum target);
+  void SetStandardTextureParams(GLenum target, GLint internal_format);
 
   // Create the framebuffer for rendering.
   void CreateFramebuffer();

mediapipe/gpu/gl_calculator_helper_impl_common.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "mediapipe/gpu/gl_calculator_helper_impl.h"
+#include "mediapipe/gpu/gpu_buffer_format.h"
 #include "mediapipe/gpu/gpu_shared_data_internal.h"
 
 namespace mediapipe {
@@ -86,9 +87,21 @@ void GlCalculatorHelperImpl::BindFramebuffer(const GlTexture& dst) {
 #endif
 }
 
-void GlCalculatorHelperImpl::SetStandardTextureParams(GLenum target) {
+void GlCalculatorHelperImpl::SetStandardTextureParams(GLenum target,
-  glTexParameteri(target, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+                                                      GLint internal_format) {
-  glTexParameteri(target, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+  GLint filter;
+  switch (internal_format) {
+    case GL_R32F:
+    case GL_RGBA32F:
+      // 32F (unlike 16f) textures do not support texture filtering
+      // (According to OpenGL ES specification [TEXTURE IMAGE SPECIFICATION])
+      filter = GL_NEAREST;
+      break;
+    default:
+      filter = GL_LINEAR;
+  }
+  glTexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
+  glTexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
   glTexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
   glTexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 }
@@ -136,7 +149,9 @@ GlTexture GlCalculatorHelperImpl::MapGlTextureBuffer(
 
   // TODO: do the params need to be reset here??
   glBindTexture(texture.target(), texture.name());
-  SetStandardTextureParams(texture.target());
+  GlTextureInfo info =
+      GlTextureInfoForGpuBufferFormat(texture_buffer->format(), texture.plane_);
+  SetStandardTextureParams(texture.target(), info.gl_internal_format);
   glBindTexture(texture.target(), 0);
 
   return texture;
@@ -150,7 +165,9 @@ GlTextureBufferSharedPtr GlCalculatorHelperImpl::MakeGlTextureBuffer(
       GpuBufferFormatForImageFormat(image_frame.Format()),
       image_frame.PixelData());
   glBindTexture(GL_TEXTURE_2D, buffer->name_);
-  SetStandardTextureParams(buffer->target_);
+  GlTextureInfo info =
+      GlTextureInfoForGpuBufferFormat(buffer->format_, /*plane=*/0);
+  SetStandardTextureParams(buffer->target_, info.gl_internal_format);
   glBindTexture(GL_TEXTURE_2D, 0);
 
   return buffer;

mediapipe/gpu/gl_calculator_helper_impl_ios.mm

@@ -54,7 +54,7 @@ GlTexture GlCalculatorHelperImpl::CreateSourceTexture(
   glTexImage2D(GL_TEXTURE_2D, 0, info.gl_internal_format, texture.width_,
                texture.height_, 0, info.gl_format, info.gl_type,
                image_frame.PixelData());
-  SetStandardTextureParams(GL_TEXTURE_2D);
+  SetStandardTextureParams(GL_TEXTURE_2D, info.gl_internal_format);
   return texture;
 }
 
@@ -107,7 +107,7 @@ GlTexture GlCalculatorHelperImpl::MapGpuBuffer(
 #endif  // TARGET_OS_OSX
 
   glBindTexture(texture.target(), texture.name());
-  SetStandardTextureParams(texture.target());
+  SetStandardTextureParams(texture.target(), info.gl_internal_format);
 
   return texture;
 }

mediapipe/gpu/gl_context_egl.cc

@@ -110,7 +110,13 @@ GlContext::StatusOrGlContext GlContext::Create(EGLContext share_context,
       eglChooseConfig(display_, config_attr, &config_, 1, &num_configs);
   if (!success) {
     return ::mediapipe::UnknownErrorBuilder(MEDIAPIPE_LOC)
-           << "eglChooseConfig() returned error " << eglGetError();
+           << "eglChooseConfig() returned error " << std::showbase << std::hex
+           << eglGetError();
+  }
+  if (!num_configs) {
+    return ::mediapipe::UnknownErrorBuilder(MEDIAPIPE_LOC)
+           << "eglChooseConfig() returned no matching EGL configuration for "
+           << "RGBA8888 D16 ES" << gl_version << " request. ";
   }
 
   const EGLint context_attr[] = {
@@ -125,7 +131,8 @@ GlContext::StatusOrGlContext GlContext::Create(EGLContext share_context,
   int error = eglGetError();
   RET_CHECK(context_ != EGL_NO_CONTEXT)
       << "Could not create GLES " << gl_version << " context; "
-      << "eglCreateContext() returned error " << error
+      << "eglCreateContext() returned error " << std::showbase << std::hex
+      << error
       << (error == EGL_BAD_CONTEXT
               ? ": external context uses a different version of OpenGL"
               : "");
@@ -143,7 +150,8 @@ GlContext::StatusOrGlContext GlContext::Create(EGLContext share_context,
 
   display_ = eglGetDisplay(EGL_DEFAULT_DISPLAY);
   RET_CHECK(display_ != EGL_NO_DISPLAY)
-      << "eglGetDisplay() returned error " << eglGetError();
+      << "eglGetDisplay() returned error " << std::showbase << std::hex
+      << eglGetError();
 
   EGLBoolean success = eglInitialize(display_, &major, &minor);
   RET_CHECK(success) << "Unable to initialize EGL";
@@ -162,7 +170,8 @@ GlContext::StatusOrGlContext GlContext::Create(EGLContext share_context,
 
   surface_ = eglCreatePbufferSurface(display_, config_, pbuffer_attr);
   RET_CHECK(surface_ != EGL_NO_SURFACE)
-      << "eglCreatePbufferSurface() returned error " << eglGetError();
+      << "eglCreatePbufferSurface() returned error " << std::showbase
+      << std::hex << eglGetError();
 
   return ::mediapipe::OkStatus();
 }
@@ -186,17 +195,21 @@ void GlContext::DestroyContext() {
   if (IsCurrent()) {
     if (!eglMakeCurrent(display_, EGL_NO_SURFACE, EGL_NO_SURFACE,
                         EGL_NO_CONTEXT)) {
-      LOG(ERROR) << "eglMakeCurrent() returned error " << eglGetError();
+      LOG(ERROR) << "eglMakeCurrent() returned error " << std::showbase
+                 << std::hex << eglGetError();
     }
   }
   if (surface_ != EGL_NO_SURFACE) {
     if (!eglDestroySurface(display_, surface_)) {
-      LOG(ERROR) << "eglDestroySurface() returned error " << eglGetError();
+      LOG(ERROR) << "eglDestroySurface() returned error " << std::showbase
+                 << std::hex << eglGetError();
     }
+    surface_ = EGL_NO_SURFACE;
   }
   if (context_ != EGL_NO_CONTEXT) {
     if (!eglDestroyContext(display_, context_)) {
-      LOG(ERROR) << "eglDestroyContext() returned error " << eglGetError();
+      LOG(ERROR) << "eglDestroyContext() returned error " << std::showbase
+                 << std::hex << eglGetError();
     }
     context_ = EGL_NO_CONTEXT;
   }
@@ -245,7 +258,8 @@ void GlContext::GetCurrentContextBinding(GlContext::ContextBinding* binding) {
   EGLBoolean success =
       eglMakeCurrent(display, new_binding.draw_surface,
                      new_binding.read_surface, new_binding.context);
-  RET_CHECK(success) << "eglMakeCurrent() returned error " << eglGetError();
+  RET_CHECK(success) << "eglMakeCurrent() returned error " << std::showbase
+                     << std::hex << eglGetError();
   return ::mediapipe::OkStatus();
 }
 

mediapipe/gpu/gl_texture_buffer.cc

@@ -77,7 +77,8 @@ bool GlTextureBuffer::CreateInternal(const void* data) {
       // TODO: maybe we do not actually have to wait for the
       // consumer sync here. Check docs.
       sync_token->WaitOnGpu();
-      DCHECK(glIsTexture(name_to_delete));
+      DLOG_IF(ERROR, !glIsTexture(name_to_delete))
+          << "Deleting invalid texture id: " << name_to_delete;
       glDeleteTextures(1, &name_to_delete);
     });
   };

mediapipe/gpu/gpu_buffer_test.cc

@@ -0,0 +1,50 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "mediapipe/gpu/gpu_buffer.h"
+
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/gpu/gpu_test_base.h"
+
+namespace mediapipe {
+namespace {
+
+class GpuBufferTest : public GpuTestBase {};
+
+TEST_F(GpuBufferTest, BasicTest) {
+  RunInGlContext([this] {
+    GpuBuffer buffer = gpu_shared_.gpu_buffer_pool.GetBuffer(300, 200);
+    EXPECT_EQ(buffer.width(), 300);
+    EXPECT_EQ(buffer.height(), 200);
+    EXPECT_TRUE(buffer);
+    EXPECT_FALSE(buffer == nullptr);
+
+    GpuBuffer no_buffer;
+    EXPECT_FALSE(no_buffer);
+    EXPECT_TRUE(no_buffer == nullptr);
+
+    GpuBuffer buffer2 = buffer;
+    EXPECT_EQ(buffer, buffer);
+    EXPECT_EQ(buffer, buffer2);
+    EXPECT_NE(buffer, no_buffer);
+
+    buffer = nullptr;
+    EXPECT_TRUE(buffer == nullptr);
+    EXPECT_TRUE(buffer == no_buffer);
+  });
+}
+
+}  // anonymous namespace
+}  // namespace mediapipe

mediapipe/gpu/gpu_shared_data_internal.h

@@ -123,7 +123,7 @@ struct GpuSharedData {
       PlatformGlContext external_context) {
     auto status_or_resources = GpuResources::Create(external_context);
     MEDIAPIPE_CHECK_OK(status_or_resources.status())
-        << "could not create GpuResources";
+        << ": could not create GpuResources";
     return std::move(status_or_resources).ValueOrDie();
   }
 };

mediapipe/gpu/gpu_test_base.h

@@ -0,0 +1,39 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef MEDIAPIPE_GPU_GPU_TEST_BASE_H_
+#define MEDIAPIPE_GPU_GPU_TEST_BASE_H_
+
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/gpu/gl_calculator_helper.h"
+#include "mediapipe/gpu/gpu_shared_data_internal.h"
+
+namespace mediapipe {
+
+class GpuTestBase : public ::testing::Test {
+ protected:
+  GpuTestBase() { helper_.InitializeForTest(&gpu_shared_); }
+
+  void RunInGlContext(std::function<void(void)> gl_func) {
+    helper_.RunInGlContext(std::move(gl_func));
+  }
+
+  GpuSharedData gpu_shared_;
+  GlCalculatorHelper helper_;
+};
+
+}  // namespace mediapipe
+
+#endif  // MEDIAPIPE_GPU_GPU_TEST_BASE_H_

mediapipe/graphs/face_detection/BUILD

@@ -35,6 +35,23 @@ cc_library(
     ],
 )
 
+cc_library(
+    name = "desktop_tflite_calculators",
+    deps = [
+        "//mediapipe/calculators/core:flow_limiter_calculator",
+        "//mediapipe/calculators/image:image_transformation_calculator",
+        "//mediapipe/calculators/tflite:ssd_anchors_calculator",
+        "//mediapipe/calculators/tflite:tflite_converter_calculator",
+        "//mediapipe/calculators/tflite:tflite_inference_calculator",
+        "//mediapipe/calculators/tflite:tflite_tensors_to_detections_calculator",
+        "//mediapipe/calculators/util:annotation_overlay_calculator",
+        "//mediapipe/calculators/util:detection_label_id_to_text_calculator",
+        "//mediapipe/calculators/util:detection_letterbox_removal_calculator",
+        "//mediapipe/calculators/util:detections_to_render_data_calculator",
+        "//mediapipe/calculators/util:non_max_suppression_calculator",
+    ],
+)
+
 load(
     "//mediapipe/framework/tool:mediapipe_graph.bzl",
     "mediapipe_binary_graph",

mediapipe/graphs/face_detection/face_detection_desktop_live.pbtxt

@@ -0,0 +1,184 @@
+# MediaPipe graph that performs face detection with TensorFlow Lite on CPU.
+# Used in the examples in
+# mediapipie/examples/desktop/face_detection:face_detection_cpu.
+
+# Images on GPU coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Throttles the images flowing downstream for flow control. It passes through
+# the very first incoming image unaltered, and waits for
+# TfLiteTensorsToDetectionsCalculator downstream in the graph to finish
+# generating the corresponding detections before it passes through another
+# image. All images that come in while waiting are dropped, limiting the number
+# of in-flight images between this calculator and
+# TfLiteTensorsToDetectionsCalculator to 1. This prevents the nodes in between
+# from queuing up incoming images and data excessively, which leads to increased
+# latency and memory usage, unwanted in real-time mobile applications. It also
+# eliminates unnecessarily computation, e.g., a transformed image produced by
+# ImageTransformationCalculator may get dropped downstream if the subsequent
+# TfLiteConverterCalculator or TfLiteInferenceCalculator is still busy
+# processing previous inputs.
+node {
+  calculator: "FlowLimiterCalculator"
+  input_stream: "input_video"
+  input_stream: "FINISHED:detections"
+  input_stream_info: {
+    tag_index: "FINISHED"
+    back_edge: true
+  }
+  output_stream: "throttled_input_video"
+}
+
+# Transforms the input image on CPU to a 128x128 image. To scale the input
+# image, the scale_mode option is set to FIT to preserve the aspect ratio,
+# resulting in potential letterboxing in the transformed image.
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE:throttled_input_video"
+  output_stream: "IMAGE:transformed_input_video_cpu"
+  output_stream: "LETTERBOX_PADDING:letterbox_padding"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 128
+      output_height: 128
+      scale_mode: FIT
+    }
+  }
+}
+
+# Converts the transformed input image on CPU into an image tensor stored as a
+# TfLiteTensor.
+node {
+  calculator: "TfLiteConverterCalculator"
+  input_stream: "IMAGE:transformed_input_video_cpu"
+  output_stream: "TENSORS:image_tensor"
+}
+
+# Runs a TensorFlow Lite model on CPU that takes an image tensor and outputs a
+# vector of tensors representing, for instance, detection boxes/keypoints and
+# scores.
+node {
+  calculator: "TfLiteInferenceCalculator"
+  input_stream: "TENSORS:image_tensor"
+  output_stream: "TENSORS:detection_tensors"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "mediapipe/models/face_detection_front.tflite"
+    }
+  }
+}
+
+# Generates a single side packet containing a vector of SSD anchors based on
+# the specification in the options.
+node {
+  calculator: "SsdAnchorsCalculator"
+  output_side_packet: "anchors"
+  node_options: {
+    [type.googleapis.com/mediapipe.SsdAnchorsCalculatorOptions] {
+      num_layers: 4
+      min_scale: 0.1484375
+      max_scale: 0.75
+      input_size_height: 128
+      input_size_width: 128
+      anchor_offset_x: 0.5
+      anchor_offset_y: 0.5
+      strides: 8
+      strides: 16
+      strides: 16
+      strides: 16
+      aspect_ratios: 1.0
+      fixed_anchor_size: true
+    }
+  }
+}
+
+# Decodes the detection tensors generated by the TensorFlow Lite model, based on
+# the SSD anchors and the specification in the options, into a vector of
+# detections. Each detection describes a detected object.
+node {
+  calculator: "TfLiteTensorsToDetectionsCalculator"
+  input_stream: "TENSORS:detection_tensors"
+  input_side_packet: "ANCHORS:anchors"
+  output_stream: "DETECTIONS:detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToDetectionsCalculatorOptions] {
+      num_classes: 1
+      num_boxes: 896
+      num_coords: 16
+      box_coord_offset: 0
+      keypoint_coord_offset: 4
+      num_keypoints: 6
+      num_values_per_keypoint: 2
+      sigmoid_score: true
+      score_clipping_thresh: 100.0
+      reverse_output_order: true
+      x_scale: 128.0
+      y_scale: 128.0
+      h_scale: 128.0
+      w_scale: 128.0
+      min_score_thresh: 0.75
+    }
+  }
+}
+
+# Performs non-max suppression to remove excessive detections.
+node {
+  calculator: "NonMaxSuppressionCalculator"
+  input_stream: "detections"
+  output_stream: "filtered_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.NonMaxSuppressionCalculatorOptions] {
+      min_suppression_threshold: 0.3
+      overlap_type: INTERSECTION_OVER_UNION
+      algorithm: WEIGHTED
+      return_empty_detections: true
+    }
+  }
+}
+
+# Maps detection label IDs to the corresponding label text ("Face"). The label
+# map is provided in the label_map_path option.
+node {
+  calculator: "DetectionLabelIdToTextCalculator"
+  input_stream: "filtered_detections"
+  output_stream: "labeled_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
+      label_map_path: "mediapipe/models/face_detection_front_labelmap.txt"
+    }
+  }
+}
+
+# Adjusts detection locations (already normalized to [0.f, 1.f]) on the
+# letterboxed image (after image transformation with the FIT scale mode) to the
+# corresponding locations on the same image with the letterbox removed (the
+# input image to the graph before image transformation).
+node {
+  calculator: "DetectionLetterboxRemovalCalculator"
+  input_stream: "DETECTIONS:labeled_detections"
+  input_stream: "LETTERBOX_PADDING:letterbox_padding"
+  output_stream: "DETECTIONS:output_detections"
+}
+
+# Converts the detections to drawing primitives for annotation overlay.
+node {
+  calculator: "DetectionsToRenderDataCalculator"
+  input_stream: "DETECTIONS:output_detections"
+  output_stream: "RENDER_DATA:render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
+      thickness: 4.0
+      color { r: 255 g: 0 b: 0 }
+    }
+  }
+}
+
+# Draws annotations and overlays them on top of the input images.
+node {
+  calculator: "AnnotationOverlayCalculator"
+  input_stream: "INPUT_FRAME:throttled_input_video"
+  input_stream: "render_data"
+  output_stream: "OUTPUT_FRAME:output_video"
+}
+

mediapipe/graphs/face_detection/face_detection_mobile_cpu.pbtxt

@@ -41,7 +41,7 @@ node: {
   output_stream: "input_video_cpu"
 }
 
-# Transforms the input image on GPU to a 128x128 image. To scale the input
+# Transforms the input image on CPU to a 128x128 image. To scale the input
 # image, the scale_mode option is set to FIT to preserve the aspect ratio,
 # resulting in potential letterboxing in the transformed image.
 node: {
@@ -75,7 +75,7 @@ node {
   output_stream: "TENSORS:detection_tensors"
   node_options: {
     [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
-      model_path: "face_detection_front.tflite"
+      model_path: "mediapipe/models/face_detection_front.tflite"
     }
   }
 }
@@ -156,7 +156,7 @@ node {
   output_stream: "labeled_detections"
   node_options: {
     [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
-      label_map_path: "face_detection_front_labelmap.txt"
+      label_map_path: "mediapipe/models/face_detection_front_labelmap.txt"
     }
   }
 }
@@ -179,7 +179,7 @@ node {
   output_stream: "RENDER_DATA:render_data"
   node_options: {
     [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
-      thickness: 10.0
+      thickness: 4.0
       color { r: 255 g: 0 b: 0 }
     }
   }

mediapipe/graphs/face_detection/face_detection_mobile_gpu.pbtxt

@@ -62,10 +62,10 @@ node {
 node {
   calculator: "TfLiteInferenceCalculator"
   input_stream: "TENSORS_GPU:image_tensor"
-  output_stream: "TENSORS:detection_tensors"
+  output_stream: "TENSORS_GPU:detection_tensors"
   node_options: {
     [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
-      model_path: "face_detection_front.tflite"
+      model_path: "mediapipe/models/face_detection_front.tflite"
     }
   }
 }
@@ -99,7 +99,7 @@ node {
 # detections. Each detection describes a detected object.
 node {
   calculator: "TfLiteTensorsToDetectionsCalculator"
-  input_stream: "TENSORS:detection_tensors"
+  input_stream: "TENSORS_GPU:detection_tensors"
   input_side_packet: "ANCHORS:anchors"
   output_stream: "DETECTIONS:detections"
   node_options: {
@@ -146,7 +146,7 @@ node {
   output_stream: "labeled_detections"
   node_options: {
     [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
-      label_map_path: "face_detection_front_labelmap.txt"
+      label_map_path: "mediapipe/models/face_detection_front_labelmap.txt"
     }
   }
 }
@@ -169,7 +169,7 @@ node {
   output_stream: "RENDER_DATA:render_data"
   node_options: {
     [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
-      thickness: 10.0
+      thickness: 4.0
       color { r: 255 g: 0 b: 0 }
     }
   }

mediapipe/graphs/hair_segmentation/hair_segmentation_mobile_gpu.pbtxt

@@ -111,7 +111,7 @@ node {
   input_side_packet: "CUSTOM_OP_RESOLVER:op_resolver"
   node_options: {
     [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
-      model_path: "hair_segmentation.tflite"
+      model_path: "mediapipe/models/hair_segmentation.tflite"
       use_gpu: true
     }
   }

mediapipe/graphs/hand_tracking/BUILD

@@ -19,73 +19,35 @@ package(default_visibility = ["//visibility:public"])
 load(
     "//mediapipe/framework/tool:mediapipe_graph.bzl",
     "mediapipe_binary_graph",
-    "mediapipe_simple_subgraph",
 )
 
-mediapipe_simple_subgraph(
+cc_library(
-    name = "hand_detection_gpu",
+    name = "desktop_tflite_calculators",
-    graph = "hand_detection_gpu.pbtxt",
-    register_as = "HandDetectionSubgraph",
     deps = [
-        "//mediapipe/calculators/image:image_properties_calculator",
+        "//mediapipe/calculators/core:flow_limiter_calculator",
-        "//mediapipe/calculators/image:image_transformation_calculator",
+        "//mediapipe/calculators/core:gate_calculator",
-        "//mediapipe/calculators/tflite:ssd_anchors_calculator",
+        "//mediapipe/calculators/core:immediate_mux_calculator",
-        "//mediapipe/calculators/tflite:tflite_converter_calculator",
+        "//mediapipe/calculators/core:merge_calculator",
-        "//mediapipe/calculators/tflite:tflite_custom_op_resolver_calculator",
+        "//mediapipe/calculators/core:packet_inner_join_calculator",
-        "//mediapipe/calculators/tflite:tflite_inference_calculator",
+        "//mediapipe/calculators/core:previous_loopback_calculator",
-        "//mediapipe/calculators/tflite:tflite_tensors_to_detections_calculator",
+        "//mediapipe/calculators/video:opencv_video_decoder_calculator",
-        "//mediapipe/calculators/util:detection_label_id_to_text_calculator",
+        "//mediapipe/calculators/video:opencv_video_encoder_calculator",
-        "//mediapipe/calculators/util:detection_letterbox_removal_calculator",
+        "//mediapipe/graphs/hand_tracking/subgraphs:hand_detection_cpu",
-        "//mediapipe/calculators/util:detections_to_rects_calculator",
+        "//mediapipe/graphs/hand_tracking/subgraphs:hand_landmark_cpu",
-        "//mediapipe/calculators/util:non_max_suppression_calculator",
+        "//mediapipe/graphs/hand_tracking/subgraphs:renderer_cpu",
-        "//mediapipe/calculators/util:rect_transformation_calculator",
-    ],
-)
-
-mediapipe_simple_subgraph(
-    name = "hand_landmark_gpu",
-    graph = "hand_landmark_gpu.pbtxt",
-    register_as = "HandLandmarkSubgraph",
-    deps = [
-        "//mediapipe/calculators/core:split_vector_calculator",
-        "//mediapipe/calculators/image:image_cropping_calculator",
-        "//mediapipe/calculators/image:image_properties_calculator",
-        "//mediapipe/calculators/image:image_transformation_calculator",
-        "//mediapipe/calculators/tflite:tflite_converter_calculator",
-        "//mediapipe/calculators/tflite:tflite_inference_calculator",
-        "//mediapipe/calculators/tflite:tflite_tensors_to_floats_calculator",
-        "//mediapipe/calculators/tflite:tflite_tensors_to_landmarks_calculator",
-        "//mediapipe/calculators/util:detections_to_rects_calculator",
-        "//mediapipe/calculators/util:landmark_letterbox_removal_calculator",
-        "//mediapipe/calculators/util:landmark_projection_calculator",
-        "//mediapipe/calculators/util:landmarks_to_detection_calculator",
-        "//mediapipe/calculators/util:rect_transformation_calculator",
-        "//mediapipe/calculators/util:thresholding_calculator",
-    ],
-)
-
-mediapipe_simple_subgraph(
-    name = "renderer_gpu",
-    graph = "renderer_gpu.pbtxt",
-    register_as = "RendererSubgraph",
-    deps = [
-        "//mediapipe/calculators/util:annotation_overlay_calculator",
-        "//mediapipe/calculators/util:detections_to_render_data_calculator",
-        "//mediapipe/calculators/util:landmarks_to_render_data_calculator",
-        "//mediapipe/calculators/util:rect_to_render_data_calculator",
     ],
 )
 
 cc_library(
     name = "mobile_calculators",
     deps = [
-        ":hand_detection_gpu",
-        ":hand_landmark_gpu",
-        ":renderer_gpu",
         "//mediapipe/calculators/core:flow_limiter_calculator",
         "//mediapipe/calculators/core:gate_calculator",
         "//mediapipe/calculators/core:merge_calculator",
         "//mediapipe/calculators/core:previous_loopback_calculator",
+        "//mediapipe/graphs/hand_tracking/subgraphs:hand_detection_gpu",
+        "//mediapipe/graphs/hand_tracking/subgraphs:hand_landmark_gpu",
+        "//mediapipe/graphs/hand_tracking/subgraphs:renderer_gpu",
     ],
 )
 
@@ -99,9 +61,9 @@ mediapipe_binary_graph(
 cc_library(
     name = "detection_mobile_calculators",
     deps = [
-        ":hand_detection_gpu",
-        ":renderer_gpu",
         "//mediapipe/calculators/core:flow_limiter_calculator",
+        "//mediapipe/graphs/hand_tracking/subgraphs:hand_detection_gpu",
+        "//mediapipe/graphs/hand_tracking/subgraphs:renderer_gpu",
     ],
 )
 

mediapipe/graphs/hand_tracking/hand_detection_desktop.pbtxt

@@ -0,0 +1,62 @@
+# MediaPipe graph that performs hand detection on desktop with TensorFlow Lite
+# on CPU.
+# Used in the example in
+# mediapipie/examples/desktop/hand_tracking:hand_detection_tflite.
+
+# max_queue_size limits the number of packets enqueued on any input stream
+# by throttling inputs to the graph. This makes the graph only process one
+# frame per time.
+max_queue_size: 1
+
+# Decodes an input video file into images and a video header.
+node {
+  calculator: "OpenCvVideoDecoderCalculator"
+  input_side_packet: "INPUT_FILE_PATH:input_video_path"
+  output_stream: "VIDEO:input_video"
+  output_stream: "VIDEO_PRESTREAM:input_video_header"
+}
+
+# Performs hand detection model on the input frames. See
+# hand_detection_cpu.pbtxt for the detail of the sub-graph.
+node {
+  calculator: "HandDetectionSubgraph"
+  input_stream: "input_video"
+  output_stream: "DETECTIONS:output_detections"
+}
+
+# Converts the detections to drawing primitives for annotation overlay.
+node {
+  calculator: "DetectionsToRenderDataCalculator"
+  input_stream: "DETECTIONS:output_detections"
+  output_stream: "RENDER_DATA:render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
+      thickness: 4.0
+      color { r: 0 g: 255 b: 0 }
+    }
+  }
+}
+
+# Draws annotations and overlays them on top of the original image coming into
+# the graph.
+node {
+  calculator: "AnnotationOverlayCalculator"
+  input_stream: "INPUT_FRAME:input_video"
+  input_stream: "render_data"
+  output_stream: "OUTPUT_FRAME:output_video"
+}
+
+# Encodes the annotated images into a video file, adopting properties specified
+# in the input video header, e.g., video framerate.
+node {
+  calculator: "OpenCvVideoEncoderCalculator"
+  input_stream: "VIDEO:output_video"
+  input_stream: "VIDEO_PRESTREAM:input_video_header"
+  input_side_packet: "OUTPUT_FILE_PATH:output_video_path"
+  node_options: {
+    [type.googleapis.com/mediapipe.OpenCvVideoEncoderCalculatorOptions]: {
+      codec: "avc1"
+      video_format: "mp4"
+    }
+  }
+}

mediapipe/graphs/hand_tracking/hand_detection_desktop_live.pbtxt

@@ -0,0 +1,38 @@
+# MediaPipe graph that performs hand detection on desktop with TensorFlow Lite
+# on CPU.
+# Used in the example in
+# mediapipie/examples/desktop/hand_tracking:hand_detection_cpu.
+
+# Images coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Performs hand detection model on the input frames. See
+# hand_detection_cpu.pbtxt for the detail of the sub-graph.
+node {
+  calculator: "HandDetectionSubgraph"
+  input_stream: "input_video"
+  output_stream: "DETECTIONS:output_detections"
+}
+
+# Converts the detections to drawing primitives for annotation overlay.
+node {
+  calculator: "DetectionsToRenderDataCalculator"
+  input_stream: "DETECTIONS:output_detections"
+  output_stream: "RENDER_DATA:render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
+      thickness: 4.0
+      color { r: 0 g: 255 b: 0 }
+    }
+  }
+}
+
+# Draws annotations and overlays them on top of the original image coming into
+# the graph.
+node {
+  calculator: "AnnotationOverlayCalculator"
+  input_stream: "INPUT_FRAME:input_video"
+  input_stream: "render_data"
+  output_stream: "OUTPUT_FRAME:output_video"
+}

mediapipe/graphs/hand_tracking/hand_tracking_desktop.pbtxt

@@ -0,0 +1,126 @@
+# MediaPipe graph that performs hand tracking on desktop with TensorFlow Lite
+# on CPU.
+# Used in the example in
+# mediapipie/examples/desktop/hand_tracking:hand_tracking_tflite.
+
+# max_queue_size limits the number of packets enqueued on any input stream
+# by throttling inputs to the graph. This makes the graph only process one
+# frame per time.
+max_queue_size: 1
+
+# Decodes an input video file into images and a video header.
+node {
+  calculator: "OpenCvVideoDecoderCalculator"
+  input_side_packet: "INPUT_FILE_PATH:input_video_path"
+  output_stream: "VIDEO:input_video"
+  output_stream: "VIDEO_PRESTREAM:input_video_header"
+}
+
+# Caches a hand-presence decision fed back from HandLandmarkSubgraph, and upon
+# the arrival of the next input image sends out the cached decision with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand-presence decision. Note that upon the arrival
+# of the very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:input_video"
+  input_stream: "LOOP:hand_presence"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_hand_presence"
+}
+
+# Drops the incoming image if HandLandmarkSubgraph was able to identify hand
+# presence in the previous image. Otherwise, passes the incoming image through
+# to trigger a new round of hand detection in HandDetectionSubgraph.
+node {
+  calculator: "GateCalculator"
+  input_stream: "input_video"
+  input_stream: "DISALLOW:prev_hand_presence"
+  output_stream: "hand_detection_input_video"
+
+  node_options: {
+    [type.googleapis.com/mediapipe.GateCalculatorOptions] {
+      empty_packets_as_allow: true
+    }
+  }
+}
+
+# Subgraph that detections hands (see hand_detection_cpu.pbtxt).
+node {
+  calculator: "HandDetectionSubgraph"
+  input_stream: "hand_detection_input_video"
+  output_stream: "DETECTIONS:palm_detections"
+  output_stream: "NORM_RECT:hand_rect_from_palm_detections"
+}
+
+# Subgraph that localizes hand landmarks (see hand_landmark_cpu.pbtxt).
+node {
+  calculator: "HandLandmarkSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "NORM_RECT:hand_rect"
+  output_stream: "LANDMARKS:hand_landmarks"
+  output_stream: "NORM_RECT:hand_rect_from_landmarks"
+  output_stream: "PRESENCE:hand_presence"
+}
+
+# Caches a hand rectangle fed back from HandLandmarkSubgraph, and upon the
+# arrival of the next input image sends out the cached rectangle with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand rectangle. Note that upon the arrival of the
+# very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:input_video"
+  input_stream: "LOOP:hand_rect_from_landmarks"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_hand_rect_from_landmarks"
+}
+
+# Merges a stream of hand rectangles generated by HandDetectionSubgraph and that
+# generated by HandLandmarkSubgraph into a single output stream by selecting
+# between one of the two streams. The former is selected if the incoming packet
+# is not empty, i.e., hand detection is performed on the current image by
+# HandDetectionSubgraph (because HandLandmarkSubgraph could not identify hand
+# presence in the previous image). Otherwise, the latter is selected, which is
+# never empty because HandLandmarkSubgraphs processes all images (that went
+# through FlowLimiterCaculator).
+node {
+  calculator: "MergeCalculator"
+  input_stream: "hand_rect_from_palm_detections"
+  input_stream: "prev_hand_rect_from_landmarks"
+  output_stream: "hand_rect"
+}
+
+# Subgraph that renders annotations and overlays them on top of the input
+# images (see renderer_cpu.pbtxt).
+node {
+  calculator: "RendererSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "LANDMARKS:hand_landmarks"
+  input_stream: "NORM_RECT:hand_rect"
+  input_stream: "DETECTIONS:palm_detections"
+  output_stream: "IMAGE:output_video"
+}
+
+# Encodes the annotated images into a video file, adopting properties specified
+# in the input video header, e.g., video framerate.
+node {
+  calculator: "OpenCvVideoEncoderCalculator"
+  input_stream: "VIDEO:output_video"
+  input_stream: "VIDEO_PRESTREAM:input_video_header"
+  input_side_packet: "OUTPUT_FILE_PATH:output_video_path"
+  node_options: {
+    [type.googleapis.com/mediapipe.OpenCvVideoEncoderCalculatorOptions]: {
+      codec: "avc1"
+      video_format: "mp4"
+    }
+  }
+}

mediapipe/graphs/hand_tracking/hand_tracking_desktop_live.pbtxt

@@ -0,0 +1,103 @@
+# MediaPipe graph that performs hand tracking on desktop with TensorFlow Lite
+# on CPU.
+# Used in the example in
+# mediapipie/examples/desktop/hand_tracking:hand_tracking_cpu.
+
+# Images coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Caches a hand-presence decision fed back from HandLandmarkSubgraph, and upon
+# the arrival of the next input image sends out the cached decision with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand-presence decision. Note that upon the arrival
+# of the very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:input_video"
+  input_stream: "LOOP:hand_presence"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_hand_presence"
+}
+
+# Drops the incoming image if HandLandmarkSubgraph was able to identify hand
+# presence in the previous image. Otherwise, passes the incoming image through
+# to trigger a new round of hand detection in HandDetectionSubgraph.
+node {
+  calculator: "GateCalculator"
+  input_stream: "input_video"
+  input_stream: "DISALLOW:prev_hand_presence"
+  output_stream: "hand_detection_input_video"
+
+  node_options: {
+    [type.googleapis.com/mediapipe.GateCalculatorOptions] {
+      empty_packets_as_allow: true
+    }
+  }
+}
+
+# Subgraph that detections hands (see hand_detection_cpu.pbtxt).
+node {
+  calculator: "HandDetectionSubgraph"
+  input_stream: "hand_detection_input_video"
+  output_stream: "DETECTIONS:palm_detections"
+  output_stream: "NORM_RECT:hand_rect_from_palm_detections"
+}
+
+# Subgraph that localizes hand landmarks (see hand_landmark_cpu.pbtxt).
+node {
+  calculator: "HandLandmarkSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "NORM_RECT:hand_rect"
+  output_stream: "LANDMARKS:hand_landmarks"
+  output_stream: "NORM_RECT:hand_rect_from_landmarks"
+  output_stream: "PRESENCE:hand_presence"
+}
+
+# Caches a hand rectangle fed back from HandLandmarkSubgraph, and upon the
+# arrival of the next input image sends out the cached rectangle with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand rectangle. Note that upon the arrival of the
+# very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:input_video"
+  input_stream: "LOOP:hand_rect_from_landmarks"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_hand_rect_from_landmarks"
+}
+
+# Merges a stream of hand rectangles generated by HandDetectionSubgraph and that
+# generated by HandLandmarkSubgraph into a single output stream by selecting
+# between one of the two streams. The former is selected if the incoming packet
+# is not empty, i.e., hand detection is performed on the current image by
+# HandDetectionSubgraph (because HandLandmarkSubgraph could not identify hand
+# presence in the previous image). Otherwise, the latter is selected, which is
+# never empty because HandLandmarkSubgraphs processes all images (that went
+# through FlowLimiterCaculator).
+node {
+  calculator: "MergeCalculator"
+  input_stream: "hand_rect_from_palm_detections"
+  input_stream: "prev_hand_rect_from_landmarks"
+  output_stream: "hand_rect"
+}
+
+# Subgraph that renders annotations and overlays them on top of the input
+# images (see renderer_cpu.pbtxt).
+node {
+  calculator: "RendererSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "LANDMARKS:hand_landmarks"
+  input_stream: "NORM_RECT:hand_rect"
+  input_stream: "DETECTIONS:palm_detections"
+  output_stream: "IMAGE:output_video"
+}
+

mediapipe/graphs/hand_tracking/subgraphs/BUILD

@@ -0,0 +1,132 @@
+# Copyright 2019 The MediaPipe Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+licenses(["notice"])  # Apache 2.0
+
+package(default_visibility = ["//visibility:public"])
+
+load(
+    "//mediapipe/framework/tool:mediapipe_graph.bzl",
+    "mediapipe_simple_subgraph",
+)
+
+mediapipe_simple_subgraph(
+    name = "hand_detection_cpu",
+    graph = "hand_detection_cpu.pbtxt",
+    register_as = "HandDetectionSubgraph",
+    deps = [
+        "//mediapipe/calculators/image:image_properties_calculator",
+        "//mediapipe/calculators/image:image_transformation_calculator",
+        "//mediapipe/calculators/tflite:ssd_anchors_calculator",
+        "//mediapipe/calculators/tflite:tflite_converter_calculator",
+        "//mediapipe/calculators/tflite:tflite_custom_op_resolver_calculator",
+        "//mediapipe/calculators/tflite:tflite_inference_calculator",
+        "//mediapipe/calculators/tflite:tflite_tensors_to_detections_calculator",
+        "//mediapipe/calculators/util:detection_label_id_to_text_calculator",
+        "//mediapipe/calculators/util:detection_letterbox_removal_calculator",
+        "//mediapipe/calculators/util:detections_to_rects_calculator",
+        "//mediapipe/calculators/util:detections_to_render_data_calculator",
+        "//mediapipe/calculators/util:non_max_suppression_calculator",
+        "//mediapipe/calculators/util:rect_transformation_calculator",
+    ],
+)
+
+mediapipe_simple_subgraph(
+    name = "hand_landmark_cpu",
+    graph = "hand_landmark_cpu.pbtxt",
+    register_as = "HandLandmarkSubgraph",
+    deps = [
+        "//mediapipe/calculators/core:split_vector_calculator",
+        "//mediapipe/calculators/image:image_cropping_calculator",
+        "//mediapipe/calculators/image:image_properties_calculator",
+        "//mediapipe/calculators/image:image_transformation_calculator",
+        "//mediapipe/calculators/tflite:tflite_converter_calculator",
+        "//mediapipe/calculators/tflite:tflite_inference_calculator",
+        "//mediapipe/calculators/tflite:tflite_tensors_to_floats_calculator",
+        "//mediapipe/calculators/tflite:tflite_tensors_to_landmarks_calculator",
+        "//mediapipe/calculators/util:detections_to_rects_calculator",
+        "//mediapipe/calculators/util:landmark_letterbox_removal_calculator",
+        "//mediapipe/calculators/util:landmark_projection_calculator",
+        "//mediapipe/calculators/util:landmarks_to_detection_calculator",
+        "//mediapipe/calculators/util:landmarks_to_render_data_calculator",
+        "//mediapipe/calculators/util:rect_transformation_calculator",
+        "//mediapipe/calculators/util:thresholding_calculator",
+    ],
+)
+
+mediapipe_simple_subgraph(
+    name = "renderer_cpu",
+    graph = "renderer_cpu.pbtxt",
+    register_as = "RendererSubgraph",
+    deps = [
+        "//mediapipe/calculators/util:annotation_overlay_calculator",
+        "//mediapipe/calculators/util:detections_to_render_data_calculator",
+        "//mediapipe/calculators/util:landmarks_to_render_data_calculator",
+        "//mediapipe/calculators/util:rect_to_render_data_calculator",
+    ],
+)
+
+mediapipe_simple_subgraph(
+    name = "hand_detection_gpu",
+    graph = "hand_detection_gpu.pbtxt",
+    register_as = "HandDetectionSubgraph",
+    deps = [
+        "//mediapipe/calculators/image:image_properties_calculator",
+        "//mediapipe/calculators/image:image_transformation_calculator",
+        "//mediapipe/calculators/tflite:ssd_anchors_calculator",
+        "//mediapipe/calculators/tflite:tflite_converter_calculator",
+        "//mediapipe/calculators/tflite:tflite_custom_op_resolver_calculator",
+        "//mediapipe/calculators/tflite:tflite_inference_calculator",
+        "//mediapipe/calculators/tflite:tflite_tensors_to_detections_calculator",
+        "//mediapipe/calculators/util:detection_label_id_to_text_calculator",
+        "//mediapipe/calculators/util:detection_letterbox_removal_calculator",
+        "//mediapipe/calculators/util:detections_to_rects_calculator",
+        "//mediapipe/calculators/util:non_max_suppression_calculator",
+        "//mediapipe/calculators/util:rect_transformation_calculator",
+    ],
+)
+
+mediapipe_simple_subgraph(
+    name = "hand_landmark_gpu",
+    graph = "hand_landmark_gpu.pbtxt",
+    register_as = "HandLandmarkSubgraph",
+    deps = [
+        "//mediapipe/calculators/core:split_vector_calculator",
+        "//mediapipe/calculators/image:image_cropping_calculator",
+        "//mediapipe/calculators/image:image_properties_calculator",
+        "//mediapipe/calculators/image:image_transformation_calculator",
+        "//mediapipe/calculators/tflite:tflite_converter_calculator",
+        "//mediapipe/calculators/tflite:tflite_inference_calculator",
+        "//mediapipe/calculators/tflite:tflite_tensors_to_floats_calculator",
+        "//mediapipe/calculators/tflite:tflite_tensors_to_landmarks_calculator",
+        "//mediapipe/calculators/util:detections_to_rects_calculator",
+        "//mediapipe/calculators/util:landmark_letterbox_removal_calculator",
+        "//mediapipe/calculators/util:landmark_projection_calculator",
+        "//mediapipe/calculators/util:landmarks_to_detection_calculator",
+        "//mediapipe/calculators/util:rect_transformation_calculator",
+        "//mediapipe/calculators/util:thresholding_calculator",
+    ],
+)
+
+mediapipe_simple_subgraph(
+    name = "renderer_gpu",
+    graph = "renderer_gpu.pbtxt",
+    register_as = "RendererSubgraph",
+    deps = [
+        "//mediapipe/calculators/util:annotation_overlay_calculator",
+        "//mediapipe/calculators/util:detections_to_render_data_calculator",
+        "//mediapipe/calculators/util:landmarks_to_render_data_calculator",
+        "//mediapipe/calculators/util:rect_to_render_data_calculator",
+    ],
+)

mediapipe/graphs/hand_tracking/subgraphs/hand_detection_cpu.pbtxt

@@ -0,0 +1,193 @@
+# MediaPipe hand detection subgraph.
+
+type: "HandDetectionSubgraph"
+
+input_stream: "input_video"
+output_stream: "DETECTIONS:palm_detections"
+output_stream: "NORM_RECT:hand_rect_from_palm_detections"
+
+# Transforms the input image on CPU to a 256x256 image. To scale the input
+# image, the scale_mode option is set to FIT to preserve the aspect ratio,
+# resulting in potential letterboxing in the transformed image.
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE:input_video"
+  output_stream: "IMAGE:transformed_input_video"
+  output_stream: "LETTERBOX_PADDING:letterbox_padding"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 256
+      output_height: 256
+      scale_mode: FIT
+    }
+  }
+}
+
+# Generates a single side packet containing a TensorFlow Lite op resolver that
+# supports custom ops needed by the model used in this graph.
+node {
+  calculator: "TfLiteCustomOpResolverCalculator"
+  output_side_packet: "op_resolver"
+}
+
+# Converts the transformed input image on CPU into an image tensor as a
+# TfLiteTensor. The zero_center option is set to true to normalize the
+# pixel values to [-1.f, 1.f] as opposed to [0.f, 1.f].
+node {
+  calculator: "TfLiteConverterCalculator"
+  input_stream: "IMAGE:transformed_input_video"
+  output_stream: "TENSORS:image_tensor"
+}
+
+# Runs a TensorFlow Lite model on CPU that takes an image tensor and outputs a
+# vector of tensors representing, for instance, detection boxes/keypoints and
+# scores.
+node {
+  calculator: "TfLiteInferenceCalculator"
+  input_stream: "TENSORS:image_tensor"
+  output_stream: "TENSORS:detection_tensors"
+  input_side_packet: "CUSTOM_OP_RESOLVER:op_resolver"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "mediapipe/models/palm_detection.tflite"
+    }
+  }
+}
+
+# Generates a single side packet containing a vector of SSD anchors based on
+# the specification in the options.
+node {
+  calculator: "SsdAnchorsCalculator"
+  output_side_packet: "anchors"
+  node_options: {
+    [type.googleapis.com/mediapipe.SsdAnchorsCalculatorOptions] {
+      num_layers: 5
+      min_scale: 0.1171875
+      max_scale: 0.75
+      input_size_height: 256
+      input_size_width: 256
+      anchor_offset_x: 0.5
+      anchor_offset_y: 0.5
+      strides: 8
+      strides: 16
+      strides: 32
+      strides: 32
+      strides: 32
+      aspect_ratios: 1.0
+      fixed_anchor_size: true
+    }
+  }
+}
+
+# Decodes the detection tensors generated by the TensorFlow Lite model, based on
+# the SSD anchors and the specification in the options, into a vector of
+# detections. Each detection describes a detected object.
+node {
+  calculator: "TfLiteTensorsToDetectionsCalculator"
+  input_stream: "TENSORS:detection_tensors"
+  input_side_packet: "ANCHORS:anchors"
+  output_stream: "DETECTIONS:detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToDetectionsCalculatorOptions] {
+      num_classes: 1
+      num_boxes: 2944
+      num_coords: 18
+      box_coord_offset: 0
+      keypoint_coord_offset: 4
+      num_keypoints: 7
+      num_values_per_keypoint: 2
+      sigmoid_score: true
+      score_clipping_thresh: 100.0
+      reverse_output_order: true
+
+      x_scale: 256.0
+      y_scale: 256.0
+      h_scale: 256.0
+      w_scale: 256.0
+      min_score_thresh: 0.5
+    }
+  }
+}
+
+# Performs non-max suppression to remove excessive detections.
+node {
+  calculator: "NonMaxSuppressionCalculator"
+  input_stream: "detections"
+  output_stream: "filtered_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.NonMaxSuppressionCalculatorOptions] {
+      min_suppression_threshold: 0.3
+      min_score_threshold: 0.5
+      overlap_type: INTERSECTION_OVER_UNION
+      algorithm: WEIGHTED
+      return_empty_detections: true
+    }
+  }
+}
+
+# Maps detection label IDs to the corresponding label text. The label map is
+# provided in the label_map_path option.
+node {
+  calculator: "DetectionLabelIdToTextCalculator"
+  input_stream: "filtered_detections"
+  output_stream: "labeled_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
+      label_map_path: "mediapipe/models/palm_detection_labelmap.txt"
+    }
+  }
+}
+
+# Adjusts detection locations (already normalized to [0.f, 1.f]) on the
+# letterboxed image (after image transformation with the FIT scale mode) to the
+# corresponding locations on the same image with the letterbox removed (the
+# input image to the graph before image transformation).
+node {
+  calculator: "DetectionLetterboxRemovalCalculator"
+  input_stream: "DETECTIONS:labeled_detections"
+  input_stream: "LETTERBOX_PADDING:letterbox_padding"
+  output_stream: "DETECTIONS:palm_detections"
+}
+
+# Extracts image size from the input images.
+node {
+  calculator: "ImagePropertiesCalculator"
+  input_stream: "IMAGE:input_video"
+  output_stream: "SIZE:image_size"
+}
+
+# Converts results of palm detection into a rectangle (normalized by image size)
+# that encloses the palm and is rotated such that the line connecting center of
+# the wrist and MCP of the middle finger is aligned with the Y-axis of the
+# rectangle.
+node {
+  calculator: "DetectionsToRectsCalculator"
+  input_stream: "DETECTIONS:palm_detections"
+  input_stream: "IMAGE_SIZE:image_size"
+  output_stream: "NORM_RECT:palm_rect"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRectsCalculatorOptions] {
+      rotation_vector_start_keypoint_index: 0  # Center of wrist.
+      rotation_vector_end_keypoint_index: 2  # MCP of middle finger.
+      rotation_vector_target_angle_degrees: 90
+      output_zero_rect_for_empty_detections: true
+    }
+  }
+}
+
+# Expands and shifts the rectangle that contains the palm so that it's likely
+# to cover the entire hand.
+node {
+  calculator: "RectTransformationCalculator"
+  input_stream: "NORM_RECT:palm_rect"
+  input_stream: "IMAGE_SIZE:image_size"
+  output_stream: "hand_rect_from_palm_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.RectTransformationCalculatorOptions] {
+      scale_x: 2.6
+      scale_y: 2.6
+      shift_y: -0.5
+      square_long: true
+    }
+  }
+}

mediapipe/graphs/hand_tracking/subgraphs/hand_detection_gpu.pbtxt

@@ -49,11 +49,11 @@ node {
 node {
   calculator: "TfLiteInferenceCalculator"
   input_stream: "TENSORS_GPU:image_tensor"
-  output_stream: "TENSORS:detection_tensors"
+  output_stream: "TENSORS_GPU:detection_tensors"
   input_side_packet: "CUSTOM_OP_RESOLVER:opresolver"
   node_options: {
     [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
-      model_path: "palm_detection.tflite"
+      model_path: "mediapipe/models/palm_detection.tflite"
       use_gpu: true
     }
   }
@@ -89,7 +89,7 @@ node {
 # detections. Each detection describes a detected object.
 node {
   calculator: "TfLiteTensorsToDetectionsCalculator"
-  input_stream: "TENSORS:detection_tensors"
+  input_stream: "TENSORS_GPU:detection_tensors"
   input_side_packet: "ANCHORS:anchors"
   output_stream: "DETECTIONS:detections"
   node_options: {
@@ -137,7 +137,7 @@ node {
   output_stream: "labeled_detections"
   node_options: {
     [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
-      label_map_path: "palm_detection_labelmap.txt"
+      label_map_path: "mediapipe/models/palm_detection_labelmap.txt"
     }
   }
 }

mediapipe/graphs/hand_tracking/subgraphs/hand_landmark_cpu.pbtxt

@@ -0,0 +1,185 @@
+# MediaPipe hand landmark localization subgraph.
+
+type: "HandLandmarkSubgraph"
+
+input_stream: "IMAGE:input_video"
+input_stream: "NORM_RECT:hand_rect"
+output_stream: "LANDMARKS:hand_landmarks"
+output_stream: "NORM_RECT:hand_rect_for_next_frame"
+output_stream: "PRESENCE:hand_presence"
+
+# Crops the rectangle that contains a hand from the input image.
+node {
+  calculator: "ImageCroppingCalculator"
+  input_stream: "IMAGE:input_video"
+  input_stream: "NORM_RECT:hand_rect"
+  output_stream: "IMAGE:hand_image"
+}
+
+# Transforms the input image on CPU to a 256x256 image. To scale the input
+# image, the scale_mode option is set to FIT to preserve the aspect ratio,
+# resulting in potential letterboxing in the transformed image.
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE:hand_image"
+  output_stream: "IMAGE:transformed_input_video"
+  output_stream: "LETTERBOX_PADDING:letterbox_padding"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 256
+      output_height: 256
+      scale_mode: FIT
+    }
+  }
+}
+
+# Converts the transformed input image on GPU into an image tensor stored in
+# tflite::gpu::GlBuffer. The zero_center option is set to true to normalize the
+# pixel values to [-1.f, 1.f] as opposed to [0.f, 1.f]. The flip_vertically
+# option is set to true to account for the descrepancy between the
+# representation of the input image (origin at the bottom-left corner, the
+# OpenGL convention) and what the model used in this graph is expecting (origin
+# at the top-left corner).
+node {
+  calculator: "TfLiteConverterCalculator"
+  input_stream: "IMAGE:transformed_input_video"
+  output_stream: "TENSORS:image_tensor"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteConverterCalculatorOptions] {
+      zero_center: false
+    }
+  }
+}
+
+# Runs a TensorFlow Lite model on GPU that takes an image tensor and outputs a
+# vector of tensors representing, for instance, detection boxes/keypoints and
+# scores.
+node {
+  calculator: "TfLiteInferenceCalculator"
+  input_stream: "TENSORS:image_tensor"
+  output_stream: "TENSORS:output_tensors"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "mediapipe/models/hand_landmark.tflite"
+    }
+  }
+}
+
+# Splits a vector of TFLite tensors to multiple vectors according to the ranges
+# specified in option.
+node {
+  calculator: "SplitTfLiteTensorVectorCalculator"
+  input_stream: "output_tensors"
+  output_stream: "landmark_tensors"
+  output_stream: "hand_flag_tensor"
+  node_options: {
+    [type.googleapis.com/mediapipe.SplitVectorCalculatorOptions] {
+      ranges: { begin: 0 end: 1 }
+      ranges: { begin: 1 end: 2 }
+    }
+  }
+}
+
+# Converts the hand-flag tensor into a float that represents the confidence
+# score of hand presence.
+node {
+  calculator: "TfLiteTensorsToFloatsCalculator"
+  input_stream: "TENSORS:hand_flag_tensor"
+  output_stream: "FLOAT:hand_presence_score"
+}
+
+# Applies a threshold to the confidence score to determine whether a hand is
+# present.
+node {
+  calculator: "ThresholdingCalculator"
+  input_stream: "FLOAT:hand_presence_score"
+  output_stream: "FLAG:hand_presence"
+  node_options: {
+    [type.googleapis.com/mediapipe.ThresholdingCalculatorOptions] {
+      threshold: 0.1
+    }
+  }
+}
+
+# Decodes the landmark tensors into a vector of lanmarks, where the landmark
+# coordinates are normalized by the size of the input image to the model.
+node {
+  calculator: "TfLiteTensorsToLandmarksCalculator"
+  input_stream: "TENSORS:landmark_tensors"
+  output_stream: "NORM_LANDMARKS:landmarks"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToLandmarksCalculatorOptions] {
+      num_landmarks: 21
+      input_image_width: 256
+      input_image_height: 256
+    }
+  }
+}
+
+# Adjusts landmarks (already normalized to [0.f, 1.f]) on the letterboxed hand
+# image (after image transformation with the FIT scale mode) to the
+# corresponding locations on the same image with the letterbox removed (hand
+# image before image transformation).
+node {
+  calculator: "LandmarkLetterboxRemovalCalculator"
+  input_stream: "LANDMARKS:landmarks"
+  input_stream: "LETTERBOX_PADDING:letterbox_padding"
+  output_stream: "LANDMARKS:scaled_landmarks"
+}
+
+# Projects the landmarks from the cropped hand image to the corresponding
+# locations on the full image before cropping (input to the graph).
+node {
+  calculator: "LandmarkProjectionCalculator"
+  input_stream: "NORM_LANDMARKS:scaled_landmarks"
+  input_stream: "NORM_RECT:hand_rect"
+  output_stream: "NORM_LANDMARKS:hand_landmarks"
+}
+
+# Extracts image size from the input images.
+node {
+  calculator: "ImagePropertiesCalculator"
+  input_stream: "IMAGE:input_video"
+  output_stream: "SIZE:image_size"
+}
+
+# Converts hand landmarks to a detection that tightly encloses all landmarks.
+node {
+  calculator: "LandmarksToDetectionCalculator"
+  input_stream: "NORM_LANDMARKS:hand_landmarks"
+  output_stream: "DETECTION:hand_detection"
+}
+
+# Converts the hand detection into a rectangle (normalized by image size)
+# that encloses the hand and is rotated such that the line connecting center of
+# the wrist and MCP of the middle finger is aligned with the Y-axis of the
+# rectangle.
+node {
+  calculator: "DetectionsToRectsCalculator"
+  input_stream: "DETECTION:hand_detection"
+  input_stream: "IMAGE_SIZE:image_size"
+  output_stream: "NORM_RECT:hand_rect_from_landmarks"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRectsCalculatorOptions] {
+      rotation_vector_start_keypoint_index: 0  # Center of wrist.
+      rotation_vector_end_keypoint_index: 9  # MCP of middle finger.
+      rotation_vector_target_angle_degrees: 90
+    }
+  }
+}
+
+# Expands the hand rectangle so that in the next video frame it's likely to
+# still contain the hand even with some motion.
+node {
+  calculator: "RectTransformationCalculator"
+  input_stream: "NORM_RECT:hand_rect_from_landmarks"
+  input_stream: "IMAGE_SIZE:image_size"
+  output_stream: "hand_rect_for_next_frame"
+  node_options: {
+    [type.googleapis.com/mediapipe.RectTransformationCalculatorOptions] {
+      scale_x: 1.6
+      scale_y: 1.6
+      square_long: true
+    }
+  }
+}