Skip to content

Commit

Permalink
postprocessing: hailo: Remove hardcoded input tensor sizes
Browse files Browse the repository at this point in the history 
Remove all hard-coded tensor size consts from the Hailo postprocessing
stages. Add a HailoPostProcessingStage::InputTensorSize() helper to
fetch the input tensor size supplied in the JSON file instead.

Signed-off-by: Naushir Patuck <naush@raspberrypi.com>
  • Loading branch information
@naushir
naushir committed Jul 19, 2024
1 parent 7a9c213 commit 18fc2d1
Show file tree
Hide file tree
Showing 6 changed files with 34 additions and 43 deletions.
6 changes: 2 additions & 4 deletions post_processing_stages/hailo/hailo_classifier.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,8 +26,6 @@ using PostProcFuncPtr = void (*)(HailoROIPtr);
#define NAME "hailo_classifier"
#define POSTPROC_LIB "libclassification.so"

static constexpr Size INPUT_TENSOR_SIZE { 224, 224 };

class HailoClassifier : public HailoPostProcessingStage
{
public:
Expand Down Expand Up @@ -90,8 +88,8 @@ bool HailoClassifier::Process(CompletedRequestPtr &completed_request)
if (low_res_info_.pixel_format == libcamera::formats::YUV420)
{
StreamInfo rgb_info;
rgb_info.width = INPUT_TENSOR_SIZE.width;
rgb_info.height = INPUT_TENSOR_SIZE.height;
rgb_info.width = InputTensorSize().width;
rgb_info.height = InputTensorSize().height;
rgb_info.stride = rgb_info.width * 3;

input = allocator_.Allocate(rgb_info.stride * rgb_info.height);
Expand Down
3 changes: 3 additions & 0 deletions post_processing_stages/hailo/hailo_postprocessing_stage.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,9 @@ int HailoPostProcessingStage::configureHailoRT()
}
bindings_ = std::move(bindings_exp.release());

hailo_3d_image_shape_t shape = infer_model_->inputs()[0].shape();
input_tensor_size_ = libcamera::Size(shape.width, shape.height);

return 0;
}

Expand Down
6 changes: 6 additions & 0 deletions post_processing_stages/hailo/hailo_postprocessing_stage.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,11 @@ class HailoPostProcessingStage : public PostProcessingStage
return std::string(HAILO_POSTPROC_LIB_DIR) + "/" + lib;
}

const libcamera::Size &InputTensorSize() const
{
return input_tensor_size_;
}

hailo_status DispatchJob(const uint8_t *input, hailort::AsyncInferJob &job, std::vector<OutTensor> &output_tensors);
HailoROIPtr MakeROI(const std::vector<OutTensor> &output_tensors) const;

Expand All @@ -125,4 +130,5 @@ class HailoPostProcessingStage : public PostProcessingStage
std::string hef_file_;
hailort::ConfiguredInferModel::Bindings bindings_;
std::chrono::time_point<std::chrono::steady_clock> last_frame_;
libcamera::Size input_tensor_size_;
};
15 changes: 4 additions & 11 deletions post_processing_stages/hailo/hailo_yolo_inference.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,13 +34,6 @@ namespace fs = std::filesystem;
#define POSTPROC_LIB "libyolo_post.so"
#define POSTPROC_LIB_NMS "libyolo_hailortpp_post.so"

namespace
{

constexpr Size INPUT_TENSOR_SIZE { 640, 640 };

} // namespace

class YoloInference : public HailoPostProcessingStage
{
public:
Expand Down Expand Up @@ -147,9 +140,9 @@ bool YoloInference::Process(CompletedRequestPtr &completed_request)
return false;
}

if (low_res_info_.width != INPUT_TENSOR_SIZE.width || low_res_info_.height != INPUT_TENSOR_SIZE.height)
if (low_res_info_.width != InputTensorSize().width || low_res_info_.height != InputTensorSize().height)
{
LOG_ERROR("Wrong low res size, expecting " << INPUT_TENSOR_SIZE.toString());
LOG_ERROR("Wrong low res size, expecting " << InputTensorSize().toString());
return false;
}

Expand All @@ -161,8 +154,8 @@ bool YoloInference::Process(CompletedRequestPtr &completed_request)
if (low_res_info_.pixel_format == libcamera::formats::YUV420)
{
StreamInfo rgb_info;
rgb_info.width = INPUT_TENSOR_SIZE.width;
rgb_info.height = INPUT_TENSOR_SIZE.height;
rgb_info.width = InputTensorSize().width;
rgb_info.height = InputTensorSize().height;
rgb_info.stride = rgb_info.width * 3;

input = allocator_.Allocate(rgb_info.stride * rgb_info.height);
Expand Down
32 changes: 15 additions & 17 deletions post_processing_stages/hailo/hailo_yolov5_segmentation.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,8 +38,6 @@ namespace fs = std::filesystem;
namespace
{

constexpr Size INPUT_TENSOR_SIZE { 640, 640 };

template <typename T>
class MessageQueue
{
Expand Down Expand Up @@ -305,9 +303,9 @@ bool YoloSegmentation::Process(CompletedRequestPtr &completed_request)
return false;
}

if (low_res_info_.width != INPUT_TENSOR_SIZE.width || low_res_info_.height != INPUT_TENSOR_SIZE.height)
if (low_res_info_.width != InputTensorSize().width || low_res_info_.height != InputTensorSize().height)
{
LOG_ERROR("Wrong low res size, expecting " << INPUT_TENSOR_SIZE.toString());
LOG_ERROR("Wrong low res size, expecting " << InputTensorSize().toString());
return false;
}

Expand All @@ -318,8 +316,8 @@ bool YoloSegmentation::Process(CompletedRequestPtr &completed_request)
if (low_res_info_.pixel_format == libcamera::formats::YUV420)
{
StreamInfo rgb_info;
rgb_info.width = INPUT_TENSOR_SIZE.width;
rgb_info.height = INPUT_TENSOR_SIZE.height;
rgb_info.width = InputTensorSize().width;
rgb_info.height = InputTensorSize().height;
rgb_info.stride = rgb_info.width * 3;

input = allocator_.Allocate(rgb_info.stride * rgb_info.height);
Expand Down Expand Up @@ -390,19 +388,19 @@ bool YoloSegmentation::runInference(uint8_t *input, uint32_t *output)
filter(roi, yolo_params_);

// RGB -> BGR for cv::imshow
for (unsigned int j = 0; j < INPUT_TENSOR_SIZE.height; j++)
for (unsigned int j = 0; j < InputTensorSize().height; j++)
{
uint8_t *ptr = input + j * INPUT_TENSOR_SIZE.width * 3;
for (unsigned int i = 0; i < INPUT_TENSOR_SIZE.width; i++)
uint8_t *ptr = input + j * InputTensorSize().width * 3;
for (unsigned int i = 0; i < InputTensorSize().width; i++)
{
const uint8_t t = ptr[0];
ptr[0] = ptr[2], ptr[2] = t;
ptr += 3;
}
}

cv::Mat image(INPUT_TENSOR_SIZE.height, INPUT_TENSOR_SIZE.width, CV_8UC3, (void *)input,
INPUT_TENSOR_SIZE.width * 3);
cv::Mat image(InputTensorSize().height, InputTensorSize().width, CV_8UC3, (void *)input,
InputTensorSize().width * 3);

std::vector<HailoDetectionPtr> detections = hailo_common::get_hailo_detections(roi);
for (auto &detection : detections)
Expand All @@ -412,10 +410,10 @@ bool YoloSegmentation::runInference(uint8_t *input, uint32_t *output)

auto bbox = detection->get_bbox();

cv::rectangle(image, cv::Point2f(bbox.xmin() * float(INPUT_TENSOR_SIZE.width),
bbox.ymin() * float(INPUT_TENSOR_SIZE.height)),
cv::Point2f(bbox.xmax() * float(INPUT_TENSOR_SIZE.width),
bbox.ymax() * float(INPUT_TENSOR_SIZE.height)),
cv::rectangle(image, cv::Point2f(bbox.xmin() * float(InputTensorSize().width),
bbox.ymin() * float(InputTensorSize().height)),
cv::Point2f(bbox.xmax() * float(InputTensorSize().width),
bbox.ymax() * float(InputTensorSize().height)),
cv::Scalar(0, 0, 255), 1);

draw_all(image, detection, 0);
Expand All @@ -439,8 +437,8 @@ void YoloSegmentation::displayThread()
{
current_image = std::move(msg.payload);

cv::Mat image(INPUT_TENSOR_SIZE.height, INPUT_TENSOR_SIZE.width, CV_8UC3, (void *)current_image.get(),
INPUT_TENSOR_SIZE.width * 3);
cv::Mat image(InputTensorSize().height, InputTensorSize().width, CV_8UC3, (void *)current_image.get(),
InputTensorSize().width * 3);

cv::imshow("Segmentation Results", image);
cv::waitKey(1);
Expand Down
15 changes: 4 additions & 11 deletions post_processing_stages/hailo/hailo_yolov8_pose.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -24,13 +24,6 @@ using PostProcFuncPtr = std::pair<std::vector<KeyPt>, std::vector<PairPairs>>(*)
#define NAME "hailo_yolo_pose"
#define POSTPROC_LIB "libyolov8pose_post.so"

namespace
{

constexpr Size INPUT_TENSOR_SIZE { 640, 640 };

} // namespace

class YoloPose : public HailoPostProcessingStage
{
public:
Expand Down Expand Up @@ -78,9 +71,9 @@ bool YoloPose::Process(CompletedRequestPtr &completed_request)
return false;
}

if (low_res_info_.width != INPUT_TENSOR_SIZE.width || low_res_info_.height != INPUT_TENSOR_SIZE.height)
if (low_res_info_.width != InputTensorSize().width || low_res_info_.height != InputTensorSize().height)
{
LOG_ERROR("Wrong low res size, expecting " << INPUT_TENSOR_SIZE.toString());
LOG_ERROR("Wrong low res size, expecting " << InputTensorSize().toString());
return false;
}

Expand All @@ -92,8 +85,8 @@ bool YoloPose::Process(CompletedRequestPtr &completed_request)
if (low_res_info_.pixel_format == libcamera::formats::YUV420)
{
StreamInfo rgb_info;
rgb_info.width = INPUT_TENSOR_SIZE.width;
rgb_info.height = INPUT_TENSOR_SIZE.height;
rgb_info.width = InputTensorSize().width;
rgb_info.height = InputTensorSize().height;
rgb_info.stride = rgb_info.width * 3;

input = allocator_.Allocate(rgb_info.stride * rgb_info.height);
Expand Down

0 comments on commit 18fc2d1

Please sign in to comment.