Make object detection work

Signed-off-by: Filip Kujawa <filip.j.kujawa@gmail.com>
Change-Id: Iff6978c55b5f2566c2891c979fed7a20a83bf690
diff --git a/y2023/vision/yolov5.cc b/y2023/vision/yolov5.cc
index 17d4ad5..473437c 100644
--- a/y2023/vision/yolov5.cc
+++ b/y2023/vision/yolov5.cc
@@ -7,8 +7,11 @@
 #include <tflite/public/edgetpu.h>
 #include <tflite/public/edgetpu_c.h>
 
+#include <chrono>
 #include <opencv2/dnn.hpp>
+#include <string>
 
+#include "absl/types/span.h"
 #include "gflags/gflags.h"
 #include "glog/logging.h"
 
@@ -23,6 +26,8 @@
 
 DEFINE_int32(nthreads, 6, "Number of threads to use during inference.");
 
+DEFINE_bool(visualize_detections, false, "Display inference output");
+
 namespace y2023 {
 namespace vision {
 
@@ -38,7 +43,7 @@
  private:
   // Convert an OpenCV Mat object to a tensor input
   // that can be fed to the TensorFlow Lite model.
-  void ConvertCVMatToTensor(cv::Mat src, uint8_t *in);
+  void ConvertCVMatToTensor(cv::Mat src, absl::Span<uint8_t> tensor);
 
   // Resizes, converts color space, and converts
   // image data type before inference.
@@ -50,10 +55,10 @@
                                                    const int columns);
 
   // Performs non-maximum suppression to remove overlapping bounding boxes.
-  void NonMaximumSupression(const std::vector<std::vector<float>> &orig_preds,
-                            const int rows, const int columns,
-                            std::vector<Detection> *detections,
-                            std::vector<int> *indices);
+  std::vector<Detection> NonMaximumSupression(
+      const std::vector<std::vector<float>> &orig_preds, const int rows,
+      const int columns, std::vector<Detection> *detections,
+      std::vector<int> *indices);
   // Models
   std::unique_ptr<tflite::FlatBufferModel> model_;
   std::unique_ptr<tflite::Interpreter> interpreter_;
@@ -71,7 +76,7 @@
   int img_width_;
 
   // Input of the interpreter
-  uint8_t *input_8_;
+  absl::Span<uint8_t> input_8_;
 
   // Subtract this offset from class labels to get the actual label.
   static constexpr int kClassIdOffset = 5;
@@ -80,94 +85,69 @@
 std::unique_ptr<YOLOV5> MakeYOLOV5() { return std::make_unique<YOLOV5Impl>(); }
 
 void YOLOV5Impl::LoadModel(const std::string path) {
-  LOG(INFO) << "Load model: start";
-
+  VLOG(1) << "Load model: Start";
 
   tflite::ops::builtin::BuiltinOpResolver resolver;
 
   model_ = tflite::FlatBufferModel::VerifyAndBuildFromFile(path.c_str());
-
-  /*
-  auto model_impl = model_->GetModel();
-  model_impl->subgraphs();
-  LOG(INFO) << model_impl;
-  LOG(INFO) << model_impl->subgraphs();
-  auto subgraphs = model_impl->subgraphs();
-  LOG(INFO) << subgraphs->size();
-  LOG(INFO) << subgraphs->Get(0)->inputs()->size();
-  LOG(INFO) << subgraphs->Get(0)->inputs()->Get(0);
-  (void)subgraphs;
-  */
-
-  LOG(INFO) << "Load model: Build Model from file";
-
   CHECK(model_);
   CHECK(model_->initialized());
+  VLOG(1) << "Load model: Build model from file success";
+
   CHECK_EQ(tflite::InterpreterBuilder(*model_, resolver)(&interpreter_),
            kTfLiteOk);
-  LOG(INFO) << "Load model: Interpreter builder done";
-  /*
-  LOG(INFO) << &interpreter_->primary_subgraph();
-  LOG(INFO) << interpreter_->subgraph(0);
-  LOG(INFO) << interpreter_->subgraphs_size();
-  LOG(INFO) << interpreter_->subgraph(0)->inputs().size();
-  LOG(INFO) << interpreter_->inputs().size();
-  */
-
-  //interpreter_->SetExternalContext(kTfLiteEdgeTpuContext, edgetpu_context.get());
-  // LOG(INFO) << "After set external context";
+  VLOG(1) << "Load model: Interpreter builder success";
 
   size_t num_devices;
   std::unique_ptr<edgetpu_device, decltype(&edgetpu_free_devices)> devices(
       edgetpu_list_devices(&num_devices), &edgetpu_free_devices);
 
-  //const auto &available_tpus =
-    // edgetpu::EdgeTpuManager::GetSingleton()->EnumerateEdgeTpu();
-  //LOG(INFO) << "Available tpus: " << available_tpus.size();
-
-  LOG(INFO) << "Load model: Getting devices";
   CHECK_EQ(num_devices, 1ul);
   const auto &device = devices.get()[0];
-  (void )device;
-  LOG(INFO) << "Load model: Got Device";
+  VLOG(1) << "Load model: Got Devices";
 
-  auto *delegate = edgetpu_create_delegate(device.type, device.path, nullptr, 0);
+  auto *delegate =
+      edgetpu_create_delegate(device.type, device.path, nullptr, 0);
 
   interpreter_->ModifyGraphWithDelegate(delegate);
 
+  VLOG(1) << "Load model: Modify graph with delegate complete";
 
   TfLiteStatus status = interpreter_->AllocateTensors();
   CHECK_EQ(status, kTfLiteOk);
   CHECK(interpreter_);
 
-  LOG(INFO) << "Load model: Allocate tensors success";
+  VLOG(1) << "Load model: Allocate tensors success";
+
   input_ = interpreter_->inputs()[0];
-  LOG(INFO) << "After set inputs";
-  LOG(INFO) << input_;
   TfLiteIntArray *dims = interpreter_->tensor(input_)->dims;
   in_height_ = dims->data[1];
   in_width_ = dims->data[2];
   in_channels_ = dims->data[3];
   in_type_ = interpreter_->tensor(input_)->type;
-  input_8_ = interpreter_->typed_tensor<uint8_t>(input_);
-  
+
+  int tensor_size = 1;
+  for (int i = 0; i < dims->size; i++) {
+    tensor_size *= dims->data[i];
+  }
+  input_8_ =
+      absl::Span(interpreter_->typed_tensor<uint8_t>(input_), tensor_size);
 
   interpreter_->SetNumThreads(FLAGS_nthreads);
 
-  LOG(INFO) << "End of load";
+  VLOG(1) << "Load model: Done";
 }
 
-void YOLOV5Impl::ConvertCVMatToTensor(cv::Mat src, uint8_t *in) {
+void YOLOV5Impl::ConvertCVMatToTensor(cv::Mat src, absl::Span<uint8_t> tensor) {
   CHECK(src.type() == CV_8UC3);
   int n = 0, nc = src.channels(), ne = src.elemSize();
-  LOG(INFO) << "ConvertCVMatToTensor - Rows " << src.rows;
-  LOG(INFO) << "ConvertCVMatToTensor - Cols " << src.cols;
+  VLOG(2) << "ConvertCVMatToTensor: Rows " << src.rows;
+  VLOG(2) << "ConvertCVMatToTensor: Cols " << src.cols;
   for (int y = 0; y < src.rows; ++y) {
+    auto *row_ptr = src.ptr<uint8_t>(y);
     for (int x = 0; x < src.cols; ++x) {
       for (int c = 0; c < nc; ++c) {
-	      (void)ne;
-	      (void)n;
-        in[n++] = src.data[y * src.step + x * ne + c];
+        tensor[n++] = *(row_ptr + x * ne + c);
       }
     }
   }
@@ -192,7 +172,7 @@
   return result_vec;
 }
 
-void YOLOV5Impl::NonMaximumSupression(
+std::vector<Detection> YOLOV5Impl::NonMaximumSupression(
     const std::vector<std::vector<float>> &orig_preds, const int rows,
     const int columns, std::vector<Detection> *detections,
     std::vector<int> *indices)
@@ -204,19 +184,24 @@
 
   for (int i = 0; i < rows; i++) {
     if (orig_preds[i][4] > FLAGS_conf_threshold) {
-      int left = (orig_preds[i][0] - orig_preds[i][2] / 2) * img_width_;
-      int top = (orig_preds[i][1] - orig_preds[i][3] / 2) * img_height_;
-      int w = orig_preds[i][2] * img_width_;
-      int h = orig_preds[i][3] * img_height_;
+      float x = orig_preds[i][0];
+      float y = orig_preds[i][1];
+      float w = orig_preds[i][2];
+      float h = orig_preds[i][3];
+      int left = static_cast<int>((x - 0.5 * w) * img_width_);
+      int top = static_cast<int>((y - 0.5 * h) * img_height_);
+      int width = static_cast<int>(w * img_width_);
+      int height = static_cast<int>(h * img_height_);
 
       for (int j = 5; j < columns; j++) {
         scores.push_back(orig_preds[i][j] * orig_preds[i][4]);
       }
 
       cv::minMaxLoc(scores, nullptr, &confidence, nullptr, &class_id);
+      scores.clear();
       if (confidence > FLAGS_conf_threshold) {
-        Detection detection{cv::Rect(left, top, w, h), confidence,
-                            class_id.x - kClassIdOffset};
+        Detection detection{cv::Rect(left, top, width, height), confidence,
+                            class_id.x};
         detections->push_back(detection);
       }
     }
@@ -230,31 +215,36 @@
     confidences.push_back(d.confidence);
   }
 
-  (void)indices;
-  // TODO(FILIP): Fix linker error.
-  // cv::dnn::NMSBoxes(boxes, confidences, FLAGS_conf_threshold,
-                  // FLAGS_nms_threshold, *indices);
+  cv::dnn::NMSBoxes(boxes, confidences, FLAGS_conf_threshold,
+                    FLAGS_nms_threshold, *indices);
+
+  std::vector<Detection> filtered_detections;
+  for (size_t i = 0; i < indices->size(); i++) {
+    filtered_detections.push_back((*detections)[(*indices)[i]]);
+  }
+
+  VLOG(1) << "NonMaximumSupression: " << detections->size() - indices->size()
+          << " detections filtered out";
+
+  return filtered_detections;
 }
 
 std::vector<Detection> YOLOV5Impl::ProcessImage(cv::Mat frame) {
+  VLOG(1) << "\n";
+
+  auto start = std::chrono::high_resolution_clock::now();
   img_height_ = frame.rows;
   img_width_ = frame.cols;
 
-  //Preprocess;
   cv::resize(frame, frame, cv::Size(in_height_, in_width_), cv::INTER_CUBIC);
   cv::cvtColor(frame, frame, cv::COLOR_BGR2RGB);
   frame.convertTo(frame, CV_8U);
 
-  LOG(INFO) << "After preprocess - Before convert to tensor";
   ConvertCVMatToTensor(frame, input_8_);
 
-  // Inference
-  LOG(INFO) << "Before Invoke";
   TfLiteStatus status = interpreter_->Invoke();
   CHECK_EQ(status, kTfLiteOk);
 
-  LOG(INFO) << "After invoke, status checked";
-
   int output_tensor_index = interpreter_->outputs()[0];
   TfLiteIntArray *out_dims = interpreter_->tensor(output_tensor_index)->dims;
   int num_rows = out_dims->data[1];
@@ -264,22 +254,48 @@
 
   std::vector<std::vector<float>> orig_preds =
       TensorToVector2D(src_tensor, num_rows, num_columns);
-  LOG(INFO) << "After tensor to vector 2D";
 
   std::vector<int> indices;
   std::vector<Detection> detections;
 
-  NonMaximumSupression(orig_preds, num_rows, num_columns, &detections,
-                       &indices);
-  LOG(INFO) << "After NMS";
-  for (size_t i = 0; i < interpreter_->outputs().size(); i++) {
-        LOG(INFO) << "Detection #" << i << " | " << interpreter_->outputs()[i];
+  std::vector<Detection> filtered_detections;
+  filtered_detections = NonMaximumSupression(orig_preds, num_rows, num_columns,
+                                             &detections, &indices);
+  VLOG(1) << "---";
+  for (size_t i = 0; i < filtered_detections.size(); i++) {
+    VLOG(1) << "Detection #" << i << " | Class ID #"
+            << filtered_detections[i].class_id << "  @ "
+            << filtered_detections[i].confidence << " confidence";
   }
-  if (detections.size() > 0) {
-    LOG(INFO) << "Detection ID: " <<  detections[0].class_id;
-    LOG(INFO) << "Confidence" << detections[0].confidence;
+
+  VLOG(1) << "---";
+
+  auto stop = std::chrono::high_resolution_clock::now();
+
+  VLOG(1) << "Inference time: "
+          << std::chrono::duration_cast<std::chrono::milliseconds>(stop - start)
+                 .count();
+
+  if (FLAGS_visualize_detections) {
+    cv::resize(frame, frame, cv::Size(img_width_, img_height_), 0, 0, true);
+    for (size_t i = 0; i < filtered_detections.size(); i++) {
+      VLOG(1) << "Bounding Box | X: " << filtered_detections[i].box.x
+              << " Y: " << filtered_detections[i].box.y
+              << " W: " << filtered_detections[i].box.width
+              << " H: " << filtered_detections[i].box.height;
+      cv::rectangle(frame, filtered_detections[i].box, cv::Scalar(255, 0, 0),
+                    2);
+      cv::putText(
+          frame, std::to_string(filtered_detections[i].class_id),
+          cv::Point(filtered_detections[i].box.x, filtered_detections[i].box.y),
+          cv::FONT_HERSHEY_COMPLEX, 1.0, cv::Scalar(0, 0, 255), 1, cv::LINE_AA);
+    }
+    cv::cvtColor(frame, frame, cv::COLOR_BGR2RGB);
+    cv::imshow("yolo", frame);
+    cv::waitKey(10);
   }
-  return detections;
+
+  return filtered_detections;
 };
 
 }  // namespace vision