ImageToTensorCalculator decides the output tensor size in runtime from the input image size.

PiperOrigin-RevId: 511882195
This commit is contained in:
MediaPipe Team 2023-02-23 14:10:21 -08:00 committed by Copybara-Service
parent 0981367e84
commit 01c64082f1
6 changed files with 88 additions and 42 deletions

View File

@ -1033,6 +1033,7 @@ cc_test(
"@com_google_absl//absl/flags:flag", "@com_google_absl//absl/flags:flag",
"@com_google_absl//absl/memory", "@com_google_absl//absl/memory",
"@com_google_absl//absl/strings", "@com_google_absl//absl/strings",
"@com_google_absl//absl/strings:str_format",
], ],
) )

View File

@ -195,8 +195,9 @@ class ImageToTensorCalculator : public Node {
#endif // MEDIAPIPE_DISABLE_GPU #endif // MEDIAPIPE_DISABLE_GPU
RotatedRect roi = GetRoi(image->width(), image->height(), norm_rect); RotatedRect roi = GetRoi(image->width(), image->height(), norm_rect);
ASSIGN_OR_RETURN(auto padding, PadRoi(options_.output_tensor_width(), const int tensor_width = params_.output_width.value_or(image->width());
options_.output_tensor_height(), const int tensor_height = params_.output_height.value_or(image->height());
ASSIGN_OR_RETURN(auto padding, PadRoi(tensor_width, tensor_height,
options_.keep_aspect_ratio(), &roi)); options_.keep_aspect_ratio(), &roi));
if (kOutLetterboxPadding(cc).IsConnected()) { if (kOutLetterboxPadding(cc).IsConnected()) {
kOutLetterboxPadding(cc).Send(padding); kOutLetterboxPadding(cc).Send(padding);
@ -214,9 +215,8 @@ class ImageToTensorCalculator : public Node {
Tensor::ElementType output_tensor_type = Tensor::ElementType output_tensor_type =
GetOutputTensorType(image->UsesGpu(), params_); GetOutputTensorType(image->UsesGpu(), params_);
Tensor tensor(output_tensor_type, Tensor tensor(output_tensor_type, {1, tensor_height, tensor_width,
{1, params_.output_height, params_.output_width, GetNumOutputChannels(*image)});
GetNumOutputChannels(*image)});
MP_RETURN_IF_ERROR((image->UsesGpu() ? gpu_converter_ : cpu_converter_) MP_RETURN_IF_ERROR((image->UsesGpu() ? gpu_converter_ : cpu_converter_)
->Convert(*image, roi, params_.range_min, ->Convert(*image, roi, params_.range_min,
params_.range_max, params_.range_max,

View File

@ -54,6 +54,8 @@ message ImageToTensorCalculatorOptions {
BORDER_REPLICATE = 2; BORDER_REPLICATE = 2;
} }
// The width and height of output tensor. The output tensor would have the
// input image width/height if not set.
optional int32 output_tensor_width = 1; optional int32 output_tensor_width = 1;
optional int32 output_tensor_height = 2; optional int32 output_tensor_height = 2;

View File

@ -13,10 +13,13 @@
// limitations under the License. // limitations under the License.
#include <cmath> #include <cmath>
#include <optional>
#include <string>
#include <vector> #include <vector>
#include "absl/flags/flag.h" #include "absl/flags/flag.h"
#include "absl/memory/memory.h" #include "absl/memory/memory.h"
#include "absl/strings/str_format.h"
#include "absl/strings/substitute.h" #include "absl/strings/substitute.h"
#include "mediapipe/calculators/tensor/image_to_tensor_converter.h" #include "mediapipe/calculators/tensor/image_to_tensor_converter.h"
#include "mediapipe/calculators/tensor/image_to_tensor_utils.h" #include "mediapipe/calculators/tensor/image_to_tensor_utils.h"
@ -51,13 +54,12 @@ std::string GetFilePath(absl::string_view filename) {
// Image to tensor test template. // Image to tensor test template.
// No processing/assertions should be done after the function is invoked. // No processing/assertions should be done after the function is invoked.
void RunTestWithInputImagePacket(const Packet& input_image_packet, void RunTestWithInputImagePacket(
cv::Mat expected_result, float range_min, const Packet& input_image_packet, cv::Mat expected_result, float range_min,
float range_max, int tensor_width, float range_max, std::optional<int> tensor_width,
int tensor_height, bool keep_aspect, std::optional<int> tensor_height, bool keep_aspect,
absl::optional<BorderMode> border_mode, absl::optional<BorderMode> border_mode,
const mediapipe::NormalizedRect& roi, const mediapipe::NormalizedRect& roi, bool output_int_tensor) {
bool output_int_tensor) {
std::string border_mode_str; std::string border_mode_str;
if (border_mode) { if (border_mode) {
switch (*border_mode) { switch (*border_mode) {
@ -93,8 +95,9 @@ void RunTestWithInputImagePacket(const Packet& input_image_packet,
})", })",
range_min, range_max); range_min, range_max);
} }
auto graph_config = mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>( auto graph_config =
absl::Substitute(R"( mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(absl::Substitute(
R"(
input_stream: "input_image" input_stream: "input_image"
input_stream: "roi" input_stream: "roi"
node { node {
@ -104,8 +107,8 @@ void RunTestWithInputImagePacket(const Packet& input_image_packet,
output_stream: "TENSORS:tensor" output_stream: "TENSORS:tensor"
options { options {
[mediapipe.ImageToTensorCalculatorOptions.ext] { [mediapipe.ImageToTensorCalculatorOptions.ext] {
output_tensor_width: $0 $0 # output tensor width
output_tensor_height: $1 $1 # output tensor height
keep_aspect_ratio: $2 keep_aspect_ratio: $2
$3 # output range $3 # output range
$4 # border mode $4 # border mode
@ -113,11 +116,16 @@ void RunTestWithInputImagePacket(const Packet& input_image_packet,
} }
} }
)", )",
/*$0=*/tensor_width, /*$0=*/tensor_width.has_value()
/*$1=*/tensor_height, ? absl::StrFormat("output_tensor_width: %d", tensor_width.value())
/*$2=*/keep_aspect ? "true" : "false", : "",
/*$3=*/output_tensor_range, /*$1=*/tensor_height.has_value()
/*$4=*/border_mode_str)); ? absl::StrFormat("output_tensor_height: %d",
tensor_height.value())
: "",
/*$2=*/keep_aspect ? "true" : "false",
/*$3=*/output_tensor_range,
/*$4=*/border_mode_str));
std::vector<Packet> output_packets; std::vector<Packet> output_packets;
tool::AddVectorSink("tensor", &graph_config, &output_packets); tool::AddVectorSink("tensor", &graph_config, &output_packets);
@ -149,18 +157,18 @@ void RunTestWithInputImagePacket(const Packet& input_image_packet,
if (output_int_tensor) { if (output_int_tensor) {
if (range_min < 0) { if (range_min < 0) {
EXPECT_EQ(tensor.element_type(), Tensor::ElementType::kInt8); EXPECT_EQ(tensor.element_type(), Tensor::ElementType::kInt8);
tensor_mat = cv::Mat(tensor_height, tensor_width, tensor_mat = cv::Mat(expected_result.rows, expected_result.cols,
channels == 1 ? CV_8SC1 : CV_8SC3, channels == 1 ? CV_8SC1 : CV_8SC3,
const_cast<int8*>(view.buffer<int8>())); const_cast<int8*>(view.buffer<int8>()));
} else { } else {
EXPECT_EQ(tensor.element_type(), Tensor::ElementType::kUInt8); EXPECT_EQ(tensor.element_type(), Tensor::ElementType::kUInt8);
tensor_mat = cv::Mat(tensor_height, tensor_width, tensor_mat = cv::Mat(expected_result.rows, expected_result.cols,
channels == 1 ? CV_8UC1 : CV_8UC3, channels == 1 ? CV_8UC1 : CV_8UC3,
const_cast<uint8*>(view.buffer<uint8>())); const_cast<uint8*>(view.buffer<uint8>()));
} }
} else { } else {
EXPECT_EQ(tensor.element_type(), Tensor::ElementType::kFloat32); EXPECT_EQ(tensor.element_type(), Tensor::ElementType::kFloat32);
tensor_mat = cv::Mat(tensor_height, tensor_width, tensor_mat = cv::Mat(expected_result.rows, expected_result.cols,
channels == 1 ? CV_32FC1 : CV_32FC3, channels == 1 ? CV_32FC1 : CV_32FC3,
const_cast<float*>(view.buffer<float>())); const_cast<float*>(view.buffer<float>()));
} }
@ -216,9 +224,9 @@ const std::vector<InputType> kInputTypesToTest = {InputType::kImageFrame,
void RunTest(cv::Mat input, cv::Mat expected_result, void RunTest(cv::Mat input, cv::Mat expected_result,
std::vector<std::pair<float, float>> float_ranges, std::vector<std::pair<float, float>> float_ranges,
std::vector<std::pair<int, int>> int_ranges, int tensor_width, std::vector<std::pair<int, int>> int_ranges,
int tensor_height, bool keep_aspect, std::optional<int> tensor_width, std::optional<int> tensor_height,
absl::optional<BorderMode> border_mode, bool keep_aspect, absl::optional<BorderMode> border_mode,
const mediapipe::NormalizedRect& roi) { const mediapipe::NormalizedRect& roi) {
for (auto input_type : kInputTypesToTest) { for (auto input_type : kInputTypesToTest) {
for (auto float_range : float_ranges) { for (auto float_range : float_ranges) {
@ -486,5 +494,18 @@ TEST(ImageToTensorCalculatorTest, NoOpExceptRangeBorderZero) {
BorderMode::kZero, roi); BorderMode::kZero, roi);
} }
TEST(ImageToTensorCalculatorTest, NoOpExceptRangeAndUseInputImageDims) {
mediapipe::NormalizedRect roi;
roi.set_x_center(0.5f);
roi.set_y_center(0.5f);
roi.set_width(1.0f);
roi.set_height(1.0f);
RunTest(GetRgb(GetFilePath("input.jpg")),
GetRgb(GetFilePath("noop_except_range.png")),
/*float_ranges=*/{{-1.0f, 1.0f}},
/*int_ranges=*/{{0, 255}, {-128, 127}},
/*tensor_width=*/std::nullopt, /*tensor_height=*/std::nullopt,
/*keep_aspect=*/false, BorderMode::kZero, roi);
}
} // namespace } // namespace
} // namespace mediapipe } // namespace mediapipe

View File

@ -16,6 +16,7 @@
#define MEDIAPIPE_CALCULATORS_TENSOR_IMAGE_TO_TENSOR_UTILS_H_ #define MEDIAPIPE_CALCULATORS_TENSOR_IMAGE_TO_TENSOR_UTILS_H_
#include <array> #include <array>
#include <optional>
#include "absl/types/optional.h" #include "absl/types/optional.h"
#include "mediapipe/calculators/tensor/image_to_tensor_calculator.pb.h" #include "mediapipe/calculators/tensor/image_to_tensor_calculator.pb.h"
@ -51,8 +52,8 @@ enum class BorderMode { kZero, kReplicate };
// Struct that host commonly accessed parameters used in the // Struct that host commonly accessed parameters used in the
// ImageTo[Batch]TensorCalculator. // ImageTo[Batch]TensorCalculator.
struct OutputTensorParams { struct OutputTensorParams {
int output_height; std::optional<int> output_height;
int output_width; std::optional<int> output_width;
int output_batch; int output_batch;
bool is_float_output; bool is_float_output;
float range_min; float range_min;
@ -161,10 +162,14 @@ absl::Status ValidateOptionOutputDims(const T& options) {
<< "The maximum of the output int tensor range must be less than or " << "The maximum of the output int tensor range must be less than or "
"equal to 127."; "equal to 127.";
} }
RET_CHECK_GT(options.output_tensor_width(), 0) if (options.has_output_tensor_width()) {
<< "Valid output tensor width is required."; RET_CHECK_GT(options.output_tensor_width(), 0)
RET_CHECK_GT(options.output_tensor_height(), 0) << "Valid output tensor width is required.";
<< "Valid output tensor height is required."; }
if (options.has_output_tensor_height()) {
RET_CHECK_GT(options.output_tensor_height(), 0)
<< "Valid output tensor height is required.";
}
return absl::OkStatus(); return absl::OkStatus();
} }
@ -185,8 +190,12 @@ OutputTensorParams GetOutputTensorParams(const T& options) {
params.range_min = options.output_tensor_float_range().min(); params.range_min = options.output_tensor_float_range().min();
params.range_max = options.output_tensor_float_range().max(); params.range_max = options.output_tensor_float_range().max();
} }
params.output_width = options.output_tensor_width(); if (options.has_output_tensor_width()) {
params.output_height = options.output_tensor_height(); params.output_width = options.output_tensor_width();
}
if (options.has_output_tensor_height()) {
params.output_height = options.output_tensor_height();
}
params.is_float_output = options.has_output_tensor_float_range(); params.is_float_output = options.has_output_tensor_float_range();
params.output_batch = 1; params.output_batch = 1;
return params; return params;

View File

@ -14,6 +14,8 @@
#include "mediapipe/calculators/tensor/image_to_tensor_utils.h" #include "mediapipe/calculators/tensor/image_to_tensor_utils.h"
#include <optional>
#include "mediapipe/framework/formats/rect.pb.h" #include "mediapipe/framework/formats/rect.pb.h"
#include "mediapipe/framework/port/gtest.h" #include "mediapipe/framework/port/gtest.h"
#include "mediapipe/framework/port/parse_text_proto.h" #include "mediapipe/framework/port/parse_text_proto.h"
@ -172,6 +174,10 @@ constexpr char kValidIntProto[] = R"(
output_tensor_height: 200 output_tensor_height: 200
)"; )";
constexpr char kValidNoTensorDimsProto[] = R"(
output_tensor_float_range { min: 0 max: 255 }
)";
TEST(ValidateOptionOutputDims, ImageToTensorCalcOptions) { TEST(ValidateOptionOutputDims, ImageToTensorCalcOptions) {
const auto float_options = const auto float_options =
mediapipe::ParseTextProtoOrDie<mediapipe::ImageToTensorCalculatorOptions>( mediapipe::ParseTextProtoOrDie<mediapipe::ImageToTensorCalculatorOptions>(
@ -193,13 +199,6 @@ TEST(ValidateOptionOutputDims, EmptyProto) {
ValidateOptionOutputDims(options), ValidateOptionOutputDims(options),
StatusIs(absl::StatusCode::kInternal, StatusIs(absl::StatusCode::kInternal,
HasSubstr("Valid output float tensor range is required"))); HasSubstr("Valid output float tensor range is required")));
// Output width/height is not set.
options.mutable_output_tensor_float_range()->set_min(0.0);
options.mutable_output_tensor_float_range()->set_max(1.0);
EXPECT_THAT(ValidateOptionOutputDims(options),
StatusIs(absl::StatusCode::kInternal,
HasSubstr("Valid output tensor width is required")));
} }
TEST(GetOutputTensorParams, ImageToTensorCalcOptionsSetValues) { TEST(GetOutputTensorParams, ImageToTensorCalcOptionsSetValues) {
@ -215,6 +214,20 @@ TEST(GetOutputTensorParams, ImageToTensorCalcOptionsSetValues) {
EXPECT_EQ(params2.output_height, 200); EXPECT_EQ(params2.output_height, 200);
} }
TEST(GetOutputTensorParams, ImageToTensorCalcOptionsNoTensorDims) {
// Test valid option for ImageToTensorCalculatorOptions without output
// width/height.
const auto options =
mediapipe::ParseTextProtoOrDie<mediapipe::ImageToTensorCalculatorOptions>(
kValidNoTensorDimsProto);
const auto params3 = GetOutputTensorParams(options);
EXPECT_EQ(params3.range_min, 0.0f);
EXPECT_EQ(params3.range_max, 255.0f);
EXPECT_EQ(params3.output_batch, 1);
EXPECT_EQ(params3.output_width, std::nullopt);
EXPECT_EQ(params3.output_height, std::nullopt);
}
TEST(GetBorderMode, GetBorderMode) { TEST(GetBorderMode, GetBorderMode) {
// Default to REPLICATE. // Default to REPLICATE.
auto border_mode = auto border_mode =