Add a GpuOrigin parameter to TensorConverterCalculator

The parameter superseeds flip_vertically. GpuOrigin works more generally than flip_vertically because CONVENTIONAL works on both iOS (no flip) and Android (yes flip). If not set, the calculator falls back to flip_vertically for backwards compatibility.

Note that web demos actually use TOP_LEFT image orientation, so they shouldn't be flipped, but they still are by CONVENTIONAL. That's being discussed right now.

PiperOrigin-RevId: 553400525

mediapipe/calculators/tensor/BUILD

@@ -620,6 +620,7 @@ mediapipe_proto_library(
     deps = [
         "//mediapipe/framework:calculator_options_proto",
         "//mediapipe/framework:calculator_proto",
+        "//mediapipe/gpu:gpu_origin_proto",
     ],
 )
 
@@ -649,7 +650,11 @@ cc_library(
         "//mediapipe/framework/formats:matrix",
         "//mediapipe/framework/formats:tensor",
         "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/port:statusor",
+        "//mediapipe/gpu:gpu_origin_cc_proto",
         "//mediapipe/util:resource_util",
+        "@com_google_absl//absl/strings:str_format",
     ] + select({
         "//mediapipe/gpu:disable_gpu": [],
         "//conditions:default": ["tensor_converter_calculator_gpu_deps"],

mediapipe/calculators/tensor/tensor_converter_calculator.cc

@@ -15,6 +15,9 @@
 #include <string>
 #include <vector>
 
+#include "absl/status/status.h"
+#include "absl/status/statusor.h"
+#include "absl/strings/str_format.h"
 #include "mediapipe/calculators/tensor/tensor_converter_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame.h"
@@ -22,7 +25,7 @@
 #include "mediapipe/framework/formats/tensor.h"
 #include "mediapipe/framework/port.h"
 #include "mediapipe/framework/port/ret_check.h"
-#include "mediapipe/util/resource_util.h"
+#include "mediapipe/gpu/gpu_origin.pb.h"
 
 #if !MEDIAPIPE_DISABLE_GPU
 #include "mediapipe/gpu/gpu_buffer.h"
@@ -43,12 +46,36 @@
 #endif  // !MEDIAPIPE_DISABLE_GPU
 
 namespace {
+
 constexpr int kWorkgroupSize = 8;  // Block size for GPU shader.
 // Commonly used to compute the number of blocks to launch in a kernel.
 int NumGroups(const int size, const int group_size) {  // NOLINT
   return (size + group_size - 1) / group_size;
 }
 
+absl::StatusOr<bool> ShouldFlipVertically(
+    const mediapipe::TensorConverterCalculatorOptions& options) {
+  if (!options.has_gpu_origin()) {
+    return options.flip_vertically();
+  }
+
+  switch (options.gpu_origin()) {
+    case mediapipe::GpuOrigin::TOP_LEFT:
+      return false;
+    case mediapipe::GpuOrigin::DEFAULT:
+    case mediapipe::GpuOrigin::CONVENTIONAL:
+      // TOP_LEFT on Metal, BOTTOM_LEFT on OpenGL.
+#ifdef __APPLE__
+      return false;
+#else
+      return true;
+#endif
+  }
+
+  return absl::InvalidArgumentError(
+      absl::StrFormat("Unhandled GPU origin %i", options.gpu_origin()));
+}
+
 typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>
     RowMajorMatrixXf;
 typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor>
@@ -58,6 +85,7 @@ constexpr char kImageFrameTag[] = "IMAGE";
 constexpr char kGpuBufferTag[] = "IMAGE_GPU";
 constexpr char kTensorsTag[] = "TENSORS";
 constexpr char kMatrixTag[] = "MATRIX";
+
 }  // namespace
 
 namespace mediapipe {
@@ -593,7 +621,7 @@ absl::Status TensorConverterCalculator::LoadOptions(CalculatorContext* cc) {
   }
 
   // Get y-flip mode.
-  flip_vertically_ = options.flip_vertically();
+  ASSIGN_OR_RETURN(flip_vertically_, ShouldFlipVertically(options));
 
   // Get row_major_matrix mode.
   row_major_matrix_ = options.row_major_matrix();

mediapipe/calculators/tensor/tensor_converter_calculator.proto

@@ -3,6 +3,7 @@ syntax = "proto2";
 package mediapipe;
 
 import "mediapipe/framework/calculator.proto";
+import "mediapipe/gpu/gpu_origin.proto";
 
 // Full Example:
 //
@@ -43,8 +44,14 @@ message TensorConverterCalculatorOptions {
   // with a coordinate system where the origin is at the bottom-left corner
   // (e.g., in OpenGL) whereas the ML model expects an image with a top-left
   // origin.
+  // Prefer gpu_origin over this field.
   optional bool flip_vertically = 2 [default = false];
 
+  // Determines when the input image should be flipped vertically.
+  // See GpuOrigin.Mode for more information.
+  // If unset, falls back to flip_vertically for backwards compatibility.
+  optional GpuOrigin.Mode gpu_origin = 10;
+
   // Controls how many channels of the input image get passed through to the
   // tensor. Valid values are 1,3,4 only. Ignored for iOS GPU.
   optional int32 max_num_channels = 3 [default = 3];

mediapipe/calculators/tensor/tensor_converter_calculator_test.cc

@@ -259,25 +259,22 @@ TEST_F(TensorConverterCalculatorTest, SetOutputRange) {
   for (std::pair<float, float> range : range_values) {
     CalculatorGraph graph;
     CalculatorGraphConfig graph_config =
-        mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
-            absl::Substitute(R"(
-        input_stream: "input_image"
-        node {
-          calculator: "TensorConverterCalculator"
-          input_stream: "IMAGE:input_image"
-          output_stream: "TENSORS:tensor"
-          options {
-            [mediapipe.TensorConverterCalculatorOptions.ext] {
-              output_tensor_float_range {
-                min: $0
-                max: $1
+        mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(absl::Substitute(
+            R"pb(
+              input_stream: "input_image"
+              node {
+                calculator: "TensorConverterCalculator"
+                input_stream: "IMAGE:input_image"
+                output_stream: "TENSORS:tensor"
+                options {
+                  [mediapipe.TensorConverterCalculatorOptions.ext] {
+                    output_tensor_float_range { min: $0 max: $1 }
+                  }
+                }
               }
-            }
-          }
-        }
-        )",
-                             /*$0=*/range.first,
-                             /*$1=*/range.second));
+            )pb",
+            /*$0=*/range.first,
+            /*$1=*/range.second));
     std::vector<Packet> output_packets;
     tool::AddVectorSink("tensor", &graph_config, &output_packets);
 
@@ -320,4 +317,113 @@ TEST_F(TensorConverterCalculatorTest, SetOutputRange) {
   }
 }
 
+TEST_F(TensorConverterCalculatorTest, FlipVertically) {
+  CalculatorGraph graph;
+  CalculatorGraphConfig graph_config =
+      mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"pb(
+        input_stream: "input_image"
+        node {
+          calculator: "TensorConverterCalculator"
+          input_stream: "IMAGE:input_image"
+          output_stream: "TENSORS:tensor"
+          options {
+            [mediapipe.TensorConverterCalculatorOptions.ext] {
+              flip_vertically: true
+              output_tensor_float_range { min: 0 max: 255 }
+            }
+          }
+        }
+      )pb");
+  std::vector<Packet> output_packets;
+  tool::AddVectorSink("tensor", &graph_config, &output_packets);
+
+  // Run the graph.
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  MP_ASSERT_OK(graph.StartRun({}));
+  auto input_image = absl::make_unique<ImageFrame>(ImageFormat::GRAY8, 1, 2);
+  cv::Mat mat = mediapipe::formats::MatView(input_image.get());
+  constexpr uint8_t kY0Value = 100;
+  constexpr uint8_t kY1Value = 200;
+  mat.at<uint8_t>(0, 0) = kY0Value;
+  mat.at<uint8_t>(1, 0) = kY1Value;  // Note: y, x!
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "input_image", Adopt(input_image.release()).At(Timestamp(0))));
+
+  // Wait until the calculator finishes processing.
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  ASSERT_THAT(output_packets.size(), Eq(1));
+
+  // Get and process results.
+  const std::vector<Tensor>& tensor_vec =
+      output_packets[0].Get<std::vector<Tensor>>();
+  EXPECT_THAT(tensor_vec.size(), Eq(1));
+
+  const Tensor* tensor = &tensor_vec[0];
+
+  EXPECT_THAT(tensor->element_type(), Eq(Tensor::ElementType::kFloat32));
+  const float* dataf = tensor->GetCpuReadView().buffer<float>();
+  EXPECT_EQ(kY1Value, static_cast<int>(roundf(dataf[0])));  // Y0, Y1 flipped!
+  EXPECT_EQ(kY0Value, static_cast<int>(roundf(dataf[1])));
+
+  // Fully close graph at end, otherwise calculator+tensors are destroyed
+  // after calling WaitUntilDone().
+  MP_ASSERT_OK(graph.CloseInputStream("input_image"));
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
+TEST_F(TensorConverterCalculatorTest, GpuOriginOverridesFlipVertically) {
+  CalculatorGraph graph;
+  CalculatorGraphConfig graph_config =
+      mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(R"pb(
+        input_stream: "input_image"
+        node {
+          calculator: "TensorConverterCalculator"
+          input_stream: "IMAGE:input_image"
+          output_stream: "TENSORS:tensor"
+          options {
+            [mediapipe.TensorConverterCalculatorOptions.ext] {
+              flip_vertically: true
+              gpu_origin: TOP_LEFT
+              output_tensor_float_range { min: 0 max: 255 }
+            }
+          }
+        }
+      )pb");
+  std::vector<Packet> output_packets;
+  tool::AddVectorSink("tensor", &graph_config, &output_packets);
+
+  // Run the graph.
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  MP_ASSERT_OK(graph.StartRun({}));
+  auto input_image = absl::make_unique<ImageFrame>(ImageFormat::GRAY8, 1, 2);
+  cv::Mat mat = mediapipe::formats::MatView(input_image.get());
+  constexpr uint8_t kY0Value = 100;
+  constexpr uint8_t kY1Value = 200;
+  mat.at<uint8_t>(0, 0) = kY0Value;
+  mat.at<uint8_t>(1, 0) = kY1Value;  // Note: y, x!
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "input_image", Adopt(input_image.release()).At(Timestamp(0))));
+
+  // Wait until the calculator finishes processing.
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  EXPECT_THAT(output_packets.size(), Eq(1));
+
+  // Get and process results.
+  const std::vector<Tensor>& tensor_vec =
+      output_packets[0].Get<std::vector<Tensor>>();
+  EXPECT_THAT(tensor_vec.size(), Eq(1));
+
+  const Tensor* tensor = &tensor_vec[0];
+
+  EXPECT_THAT(tensor->element_type(), Eq(Tensor::ElementType::kFloat32));
+  const float* dataf = tensor->GetCpuReadView().buffer<float>();
+  EXPECT_EQ(kY0Value, static_cast<int>(roundf(dataf[0])));  // Not flipped!
+  EXPECT_EQ(kY1Value, static_cast<int>(roundf(dataf[1])));
+
+  // Fully close graph at end, otherwise calculator+tensors are destroyed
+  // after calling WaitUntilDone().
+  MP_ASSERT_OK(graph.CloseInputStream("input_image"));
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
 }  // namespace mediapipe