Added yolov3-tiny.cfg

AlexeyAB

2018-05-06 efcbdb5425f60d80d9633930f007dfe3caae5220

 src/yolo_v2_class.hpp

@@ -1,15 +1,4 @@
#pragma once
#include <memory>
#include <vector>
#include <deque>
#include <algorithm>

#ifdef OPENCV
#include <opencv2/opencv.hpp>       // C++
#include "opencv2/highgui/highgui_c.h" // C
#include "opencv2/imgproc/imgproc_c.h" // C
#endif   // OPENCV

#ifdef YOLODLL_EXPORTS
#if defined(_MSC_VER)
#define YOLODLL_API __declspec(dllexport) 
@@ -29,6 +18,7 @@
   float prob;             // confidence - probability that the object was found correctly
   unsigned int obj_id;    // class of object - from range [0, classes-1]
   unsigned int track_id;     // tracking id for video (0 - untracked, 1 - inf - tracked object)
   unsigned int frames_counter;// counter of frames on which the object was detected
};

struct image_t {
@@ -38,11 +28,25 @@
   float *data;            // pointer to the image data
};

#ifdef __cplusplus
#include <memory>
#include <vector>
#include <deque>
#include <algorithm>

#ifdef OPENCV
#include <opencv2/opencv.hpp>       // C++
#include "opencv2/highgui/highgui_c.h" // C
#include "opencv2/imgproc/imgproc_c.h" // C
#endif   // OPENCV

class Detector {
   std::shared_ptr<void> detector_gpu_ptr;
   std::deque<std::vector<bbox_t>> prev_bbox_vec_deque;
   const int cur_gpu_id;
public:
   float nms = .4;
   bool wait_stream;

   YOLODLL_API Detector(std::string cfg_filename, std::string weight_filename, int gpu_id = 0);
   YOLODLL_API ~Detector();
@@ -51,35 +55,51 @@
   YOLODLL_API std::vector<bbox_t> detect(image_t img, float thresh = 0.2, bool use_mean = false);
   static YOLODLL_API image_t load_image(std::string image_filename);
   static YOLODLL_API void free_image(image_t m);
   YOLODLL_API int get_net_width();
   YOLODLL_API int get_net_height();
   YOLODLL_API int get_net_width() const;
   YOLODLL_API int get_net_height() const;

   YOLODLL_API std::vector<bbox_t> tracking(std::vector<bbox_t> cur_bbox_vec, int const frames_story = 4);
   YOLODLL_API std::vector<bbox_t> tracking_id(std::vector<bbox_t> cur_bbox_vec, bool const change_history = true, 
                                    int const frames_story = 10, int const max_dist = 150);

   std::vector<bbox_t> detect_resized(image_t img, int init_w, int init_h, float thresh = 0.2, bool use_mean = false)
   {
      if (img.data == NULL)
         throw std::runtime_error("Image is empty");
      auto detection_boxes = detect(img, thresh, use_mean);
      float wk = (float)init_w / img.w, hk = (float)init_h / img.h;
      for (auto &i : detection_boxes) i.x *= wk, i.w *= wk, i.y *= hk, i.h *= hk;
      return detection_boxes;
   }

#ifdef OPENCV
   std::vector<bbox_t> detect(cv::Mat mat, float thresh = 0.2, bool use_mean = false)
   {
      if(mat.data == NULL)
         throw std::runtime_error("file not found");
      cv::Mat det_mat;
      cv::resize(mat, det_mat, cv::Size(get_net_width(), get_net_height()));
      auto image_ptr = mat_to_image(det_mat);
      auto detection_boxes = detect(*image_ptr, thresh);
      float wk = (float)mat.cols / det_mat.cols, hk = (float)mat.rows / det_mat.rows;
      for (auto &i : detection_boxes) i.x*=wk, i.w*= wk, i.y*=hk, i.h*=hk;
      return detection_boxes;
         throw std::runtime_error("Image is empty");
      auto image_ptr = mat_to_image_resize(mat);
      return detect_resized(*image_ptr, mat.cols, mat.rows, thresh, use_mean);
   }

   static std::shared_ptr<image_t> mat_to_image(cv::Mat img)
   std::shared_ptr<image_t> mat_to_image_resize(cv::Mat mat) const
   {
      if (mat.data == NULL) return std::shared_ptr<image_t>(NULL);
      cv::Mat det_mat;
      cv::resize(mat, det_mat, cv::Size(get_net_width(), get_net_height()));
      return mat_to_image(det_mat);
   }

   static std::shared_ptr<image_t> mat_to_image(cv::Mat img_src)
   {
      cv::Mat img;
      cv::cvtColor(img_src, img, cv::COLOR_RGB2BGR);
      std::shared_ptr<image_t> image_ptr(new image_t, [](image_t *img) { free_image(*img); delete img; });
      std::shared_ptr<IplImage> ipl_small = std::make_shared<IplImage>(img);
      *image_ptr = ipl_to_image(ipl_small.get());
      rgbgr_image(*image_ptr);
      return image_ptr;
   }

private:

   static image_t ipl_to_image(IplImage* src)
   {
      unsigned char *data = (unsigned char *)src->imageData;
@@ -88,15 +108,17 @@
      int c = src->nChannels;
      int step = src->widthStep;
      image_t out = make_image_custom(w, h, c);
      int i, j, k, count = 0;;
      int count = 0;

      for (k = 0; k < c; ++k) {
         for (i = 0; i < h; ++i) {
            for (j = 0; j < w; ++j) {
               out.data[count++] = data[i*step + j*c + k] / 255.;
      for (int k = 0; k < c; ++k) {
         for (int i = 0; i < h; ++i) {
            int i_step = i*step;
            for (int j = 0; j < w; ++j) {
               out.data[count++] = data[i_step + j*c + k] / 255.;
            }
         }
      }

      return out;
   }

@@ -117,20 +139,503 @@
      return out;
   }

   static void rgbgr_image(image_t im)
   {
      int i;
      for (i = 0; i < im.w*im.h; ++i) {
         float swap = im.data[i];
         im.data[i] = im.data[i + im.w*im.h * 2];
         im.data[i + im.w*im.h * 2] = swap;
      }
   }

#endif   // OPENCV

   std::deque<std::vector<bbox_t>> prev_bbox_vec_deque;
};



#if defined(TRACK_OPTFLOW) && defined(OPENCV) && defined(GPU)

#include <opencv2/cudaoptflow.hpp>
#include <opencv2/cudaimgproc.hpp>
#include <opencv2/cudaarithm.hpp>
#include <opencv2/core/cuda.hpp>

class Tracker_optflow {
public:
   const int gpu_count;
   const int gpu_id;
   const int flow_error;


   Tracker_optflow(int _gpu_id = 0, int win_size = 9, int max_level = 3, int iterations = 8000, int _flow_error = -1) :
      gpu_count(cv::cuda::getCudaEnabledDeviceCount()), gpu_id(std::min(_gpu_id, gpu_count-1)),
      flow_error((_flow_error > 0)? _flow_error:(win_size*4))
   {
      int const old_gpu_id = cv::cuda::getDevice();
      cv::cuda::setDevice(gpu_id);

      stream = cv::cuda::Stream();

      sync_PyrLKOpticalFlow_gpu = cv::cuda::SparsePyrLKOpticalFlow::create();
      sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(win_size, win_size)); // 9, 15, 21, 31
      sync_PyrLKOpticalFlow_gpu->setMaxLevel(max_level);    // +- 3 pt
      sync_PyrLKOpticalFlow_gpu->setNumIters(iterations);   // 2000, def: 30

      cv::cuda::setDevice(old_gpu_id);
   }

   // just to avoid extra allocations
   cv::cuda::GpuMat src_mat_gpu;
   cv::cuda::GpuMat dst_mat_gpu, dst_grey_gpu;
   cv::cuda::GpuMat prev_pts_flow_gpu, cur_pts_flow_gpu;
   cv::cuda::GpuMat status_gpu, err_gpu;

   cv::cuda::GpuMat src_grey_gpu;   // used in both functions
   cv::Ptr<cv::cuda::SparsePyrLKOpticalFlow> sync_PyrLKOpticalFlow_gpu;
   cv::cuda::Stream stream;

   std::vector<bbox_t> cur_bbox_vec;
   std::vector<bool> good_bbox_vec_flags;
   cv::Mat prev_pts_flow_cpu;

   void update_cur_bbox_vec(std::vector<bbox_t> _cur_bbox_vec)
   {
      cur_bbox_vec = _cur_bbox_vec;
      good_bbox_vec_flags = std::vector<bool>(cur_bbox_vec.size(), true);
      cv::Mat prev_pts, cur_pts_flow_cpu;

      for (auto &i : cur_bbox_vec) {
         float x_center = (i.x + i.w / 2.0F);
         float y_center = (i.y + i.h / 2.0F);
         prev_pts.push_back(cv::Point2f(x_center, y_center));
      }

      if (prev_pts.rows == 0)
         prev_pts_flow_cpu = cv::Mat();
      else
         cv::transpose(prev_pts, prev_pts_flow_cpu);

      if (prev_pts_flow_gpu.cols < prev_pts_flow_cpu.cols) {
         prev_pts_flow_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), prev_pts_flow_cpu.type());
         cur_pts_flow_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), prev_pts_flow_cpu.type());

         status_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), CV_8UC1);
         err_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), CV_32FC1);
      }

      prev_pts_flow_gpu.upload(cv::Mat(prev_pts_flow_cpu), stream);
   }


   void update_tracking_flow(cv::Mat src_mat, std::vector<bbox_t> _cur_bbox_vec)
   {
      int const old_gpu_id = cv::cuda::getDevice();
      if (old_gpu_id != gpu_id)
         cv::cuda::setDevice(gpu_id);

      if (src_mat.channels() == 3) {
         if (src_mat_gpu.cols == 0) {
            src_mat_gpu = cv::cuda::GpuMat(src_mat.size(), src_mat.type());
            src_grey_gpu = cv::cuda::GpuMat(src_mat.size(), CV_8UC1);
         }

         update_cur_bbox_vec(_cur_bbox_vec);

         //src_grey_gpu.upload(src_mat, stream);   // use BGR
         src_mat_gpu.upload(src_mat, stream);
         cv::cuda::cvtColor(src_mat_gpu, src_grey_gpu, CV_BGR2GRAY, 1, stream);
      }
      if (old_gpu_id != gpu_id)
         cv::cuda::setDevice(old_gpu_id);
   }


   std::vector<bbox_t> tracking_flow(cv::Mat dst_mat, bool check_error = true)
   {
      if (sync_PyrLKOpticalFlow_gpu.empty()) {
         std::cout << "sync_PyrLKOpticalFlow_gpu isn't initialized \n";
         return cur_bbox_vec;
      }

      int const old_gpu_id = cv::cuda::getDevice();
      if(old_gpu_id != gpu_id)
         cv::cuda::setDevice(gpu_id);

      if (dst_mat_gpu.cols == 0) {
         dst_mat_gpu = cv::cuda::GpuMat(dst_mat.size(), dst_mat.type());
         dst_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1);
      }

      //dst_grey_gpu.upload(dst_mat, stream);   // use BGR
      dst_mat_gpu.upload(dst_mat, stream);
      cv::cuda::cvtColor(dst_mat_gpu, dst_grey_gpu, CV_BGR2GRAY, 1, stream);

      if (src_grey_gpu.rows != dst_grey_gpu.rows || src_grey_gpu.cols != dst_grey_gpu.cols) {
         stream.waitForCompletion();
         src_grey_gpu = dst_grey_gpu.clone();
         cv::cuda::setDevice(old_gpu_id);
         return cur_bbox_vec;
      }

      ////sync_PyrLKOpticalFlow_gpu.sparse(src_grey_gpu, dst_grey_gpu, prev_pts_flow_gpu, cur_pts_flow_gpu, status_gpu, &err_gpu);  // OpenCV 2.4.x
      sync_PyrLKOpticalFlow_gpu->calc(src_grey_gpu, dst_grey_gpu, prev_pts_flow_gpu, cur_pts_flow_gpu, status_gpu, err_gpu, stream);   // OpenCV 3.x

      cv::Mat cur_pts_flow_cpu;
      cur_pts_flow_gpu.download(cur_pts_flow_cpu, stream);

      dst_grey_gpu.copyTo(src_grey_gpu, stream);

      cv::Mat err_cpu, status_cpu;
      err_gpu.download(err_cpu, stream);
      status_gpu.download(status_cpu, stream);

      stream.waitForCompletion();

      std::vector<bbox_t> result_bbox_vec;

      if (err_cpu.cols == cur_bbox_vec.size() && status_cpu.cols == cur_bbox_vec.size()) 
      {
         for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
         {
            cv::Point2f cur_key_pt = cur_pts_flow_cpu.at<cv::Point2f>(0, i);
            cv::Point2f prev_key_pt = prev_pts_flow_cpu.at<cv::Point2f>(0, i);

            float moved_x = cur_key_pt.x - prev_key_pt.x;
            float moved_y = cur_key_pt.y - prev_key_pt.y;

            if (abs(moved_x) < 100 && abs(moved_y) < 100 && good_bbox_vec_flags[i])
               if (err_cpu.at<float>(0, i) < flow_error && status_cpu.at<unsigned char>(0, i) != 0 &&
                  ((float)cur_bbox_vec[i].x + moved_x) > 0 && ((float)cur_bbox_vec[i].y + moved_y) > 0)
               {
                  cur_bbox_vec[i].x += moved_x + 0.5;
                  cur_bbox_vec[i].y += moved_y + 0.5;
                  result_bbox_vec.push_back(cur_bbox_vec[i]);
               }
               else good_bbox_vec_flags[i] = false;
            else good_bbox_vec_flags[i] = false;

            //if(!check_error && !good_bbox_vec_flags[i]) result_bbox_vec.push_back(cur_bbox_vec[i]);
         }
      }

      cur_pts_flow_gpu.swap(prev_pts_flow_gpu);
      cur_pts_flow_cpu.copyTo(prev_pts_flow_cpu);

      if (old_gpu_id != gpu_id)
         cv::cuda::setDevice(old_gpu_id);

      return result_bbox_vec;
   }

};

#elif defined(TRACK_OPTFLOW) && defined(OPENCV)

//#include <opencv2/optflow.hpp>
#include <opencv2/video/tracking.hpp>

class Tracker_optflow {
public:
   const int flow_error;


   Tracker_optflow(int win_size = 9, int max_level = 3, int iterations = 8000, int _flow_error = -1) :
      flow_error((_flow_error > 0)? _flow_error:(win_size*4))
   {
      sync_PyrLKOpticalFlow = cv::SparsePyrLKOpticalFlow::create();
      sync_PyrLKOpticalFlow->setWinSize(cv::Size(win_size, win_size));  // 9, 15, 21, 31
      sync_PyrLKOpticalFlow->setMaxLevel(max_level);     // +- 3 pt

   }

   // just to avoid extra allocations
   cv::Mat dst_grey;
   cv::Mat prev_pts_flow, cur_pts_flow;
   cv::Mat status, err;

   cv::Mat src_grey; // used in both functions
   cv::Ptr<cv::SparsePyrLKOpticalFlow> sync_PyrLKOpticalFlow;

   std::vector<bbox_t> cur_bbox_vec;
   std::vector<bool> good_bbox_vec_flags;

   void update_cur_bbox_vec(std::vector<bbox_t> _cur_bbox_vec)
   {
      cur_bbox_vec = _cur_bbox_vec;
      good_bbox_vec_flags = std::vector<bool>(cur_bbox_vec.size(), true);
      cv::Mat prev_pts, cur_pts_flow;

      for (auto &i : cur_bbox_vec) {
         float x_center = (i.x + i.w / 2.0F);
         float y_center = (i.y + i.h / 2.0F);
         prev_pts.push_back(cv::Point2f(x_center, y_center));
      }

      if (prev_pts.rows == 0)
         prev_pts_flow = cv::Mat();
      else
         cv::transpose(prev_pts, prev_pts_flow);
   }


   void update_tracking_flow(cv::Mat new_src_mat, std::vector<bbox_t> _cur_bbox_vec)
   {
      if (new_src_mat.channels() == 3) {

         update_cur_bbox_vec(_cur_bbox_vec);

         cv::cvtColor(new_src_mat, src_grey, CV_BGR2GRAY, 1);
      }
   }


   std::vector<bbox_t> tracking_flow(cv::Mat new_dst_mat, bool check_error = true)
   {
      if (sync_PyrLKOpticalFlow.empty()) {
         std::cout << "sync_PyrLKOpticalFlow isn't initialized \n";
         return cur_bbox_vec;
      }

      cv::cvtColor(new_dst_mat, dst_grey, CV_BGR2GRAY, 1);

      if (src_grey.rows != dst_grey.rows || src_grey.cols != dst_grey.cols) {
         src_grey = dst_grey.clone();
         return cur_bbox_vec;
      }

      if (prev_pts_flow.cols < 1) {
         return cur_bbox_vec;
      }

      ////sync_PyrLKOpticalFlow_gpu.sparse(src_grey_gpu, dst_grey_gpu, prev_pts_flow_gpu, cur_pts_flow_gpu, status_gpu, &err_gpu);  // OpenCV 2.4.x
      sync_PyrLKOpticalFlow->calc(src_grey, dst_grey, prev_pts_flow, cur_pts_flow, status, err);   // OpenCV 3.x

      dst_grey.copyTo(src_grey);

      std::vector<bbox_t> result_bbox_vec;

      if (err.rows == cur_bbox_vec.size() && status.rows == cur_bbox_vec.size())
      {
         for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
         {
            cv::Point2f cur_key_pt = cur_pts_flow.at<cv::Point2f>(0, i);
            cv::Point2f prev_key_pt = prev_pts_flow.at<cv::Point2f>(0, i);

            float moved_x = cur_key_pt.x - prev_key_pt.x;
            float moved_y = cur_key_pt.y - prev_key_pt.y;

            if (abs(moved_x) < 100 && abs(moved_y) < 100 && good_bbox_vec_flags[i])
               if (err.at<float>(0, i) < flow_error && status.at<unsigned char>(0, i) != 0 &&
                  ((float)cur_bbox_vec[i].x + moved_x) > 0 && ((float)cur_bbox_vec[i].y + moved_y) > 0)
               {
                  cur_bbox_vec[i].x += moved_x + 0.5;
                  cur_bbox_vec[i].y += moved_y + 0.5;
                  result_bbox_vec.push_back(cur_bbox_vec[i]);
               }
               else good_bbox_vec_flags[i] = false;
            else good_bbox_vec_flags[i] = false;

            //if(!check_error && !good_bbox_vec_flags[i]) result_bbox_vec.push_back(cur_bbox_vec[i]);
         }
      }

      prev_pts_flow = cur_pts_flow.clone();

      return result_bbox_vec;
   }

};
#else

class Tracker_optflow {};

#endif   // defined(TRACK_OPTFLOW) && defined(OPENCV)


#ifdef OPENCV

static cv::Scalar obj_id_to_color(int obj_id) {
   int const colors[6][3] = { { 1,0,1 },{ 0,0,1 },{ 0,1,1 },{ 0,1,0 },{ 1,1,0 },{ 1,0,0 } };
   int const offset = obj_id * 123457 % 6;
   int const color_scale = 150 + (obj_id * 123457) % 100;
   cv::Scalar color(colors[offset][0], colors[offset][1], colors[offset][2]);
   color *= color_scale;
   return color;
}

class preview_boxes_t {
   enum { frames_history = 30 }; // how long to keep the history saved

   struct preview_box_track_t {
      unsigned int track_id, obj_id, last_showed_frames_ago;
      bool current_detection;
      bbox_t bbox;
      cv::Mat mat_obj, mat_resized_obj;
      preview_box_track_t() : track_id(0), obj_id(0), last_showed_frames_ago(frames_history), current_detection(false) {}
   };
   std::vector<preview_box_track_t> preview_box_track_id;
   size_t const preview_box_size, bottom_offset;
   bool const one_off_detections;
public:
   preview_boxes_t(size_t _preview_box_size = 100, size_t _bottom_offset = 100, bool _one_off_detections = false) :
      preview_box_size(_preview_box_size), bottom_offset(_bottom_offset), one_off_detections(_one_off_detections)
   {}

   void set(cv::Mat src_mat, std::vector<bbox_t> result_vec)
   {
      size_t const count_preview_boxes = src_mat.cols / preview_box_size;
      if (preview_box_track_id.size() != count_preview_boxes) preview_box_track_id.resize(count_preview_boxes);

      // increment frames history
      for (auto &i : preview_box_track_id)
         i.last_showed_frames_ago = std::min((unsigned)frames_history, i.last_showed_frames_ago + 1);

      // occupy empty boxes
      for (auto &k : result_vec) {
         bool found = false;
         // find the same (track_id)
         for (auto &i : preview_box_track_id) {
            if (i.track_id == k.track_id) {
               if (!one_off_detections) i.last_showed_frames_ago = 0; // for tracked objects
               found = true;
               break;
            }
         }
         if (!found) {
            // find empty box
            for (auto &i : preview_box_track_id) {
               if (i.last_showed_frames_ago == frames_history) {
                  if (!one_off_detections && k.frames_counter == 0) break; // don't show if obj isn't tracked yet
                  i.track_id = k.track_id;
                  i.obj_id = k.obj_id;
                  i.bbox = k;
                  i.last_showed_frames_ago = 0;
                  break;
               }
            }
         }
      }

      // draw preview box (from old or current frame)
      for (size_t i = 0; i < preview_box_track_id.size(); ++i)
      {
         // get object image
         cv::Mat dst = preview_box_track_id[i].mat_resized_obj;
         preview_box_track_id[i].current_detection = false;

         for (auto &k : result_vec) {
            if (preview_box_track_id[i].track_id == k.track_id) {
               if (one_off_detections && preview_box_track_id[i].last_showed_frames_ago > 0) {
                  preview_box_track_id[i].last_showed_frames_ago = frames_history; break;
               }
               bbox_t b = k;
               cv::Rect r(b.x, b.y, b.w, b.h);
               cv::Rect img_rect(cv::Point2i(0, 0), src_mat.size());
               cv::Rect rect_roi = r & img_rect;
               if (rect_roi.width > 1 || rect_roi.height > 1) {
                  cv::Mat roi = src_mat(rect_roi);
                  cv::resize(roi, dst, cv::Size(preview_box_size, preview_box_size), cv::INTER_NEAREST);
                  preview_box_track_id[i].mat_obj = roi.clone();
                  preview_box_track_id[i].mat_resized_obj = dst.clone();
                  preview_box_track_id[i].current_detection = true;
                  preview_box_track_id[i].bbox = k;
               }
               break;
            }
         }
      }
   }


   void draw(cv::Mat draw_mat, bool show_small_boxes = false)
   {
      // draw preview box (from old or current frame)
      for (size_t i = 0; i < preview_box_track_id.size(); ++i)
      {
         auto &prev_box = preview_box_track_id[i];

         // draw object image
         cv::Mat dst = prev_box.mat_resized_obj;
         if (prev_box.last_showed_frames_ago < frames_history &&
            dst.size() == cv::Size(preview_box_size, preview_box_size))
         {
            cv::Rect dst_rect_roi(cv::Point2i(i * preview_box_size, draw_mat.rows - bottom_offset), dst.size());
            cv::Mat dst_roi = draw_mat(dst_rect_roi);
            dst.copyTo(dst_roi);

            cv::Scalar color = obj_id_to_color(prev_box.obj_id);
            int thickness = (prev_box.current_detection) ? 5 : 1;
            cv::rectangle(draw_mat, dst_rect_roi, color, thickness);

            unsigned int const track_id = prev_box.track_id;
            std::string track_id_str = (track_id > 0) ? std::to_string(track_id) : "";
            putText(draw_mat, track_id_str, dst_rect_roi.tl() - cv::Point2i(-4, 5), cv::FONT_HERSHEY_COMPLEX_SMALL, 0.9, cv::Scalar(0, 0, 0), 2);

            std::string size_str = std::to_string(prev_box.bbox.w) + "x" + std::to_string(prev_box.bbox.h);
            putText(draw_mat, size_str, dst_rect_roi.tl() + cv::Point2i(0, 12), cv::FONT_HERSHEY_COMPLEX_SMALL, 0.8, cv::Scalar(0, 0, 0), 1);

            if (!one_off_detections && prev_box.current_detection) {
               cv::line(draw_mat, dst_rect_roi.tl() + cv::Point2i(preview_box_size, 0),
                  cv::Point2i(prev_box.bbox.x, prev_box.bbox.y + prev_box.bbox.h),
                  color);
            }

            if (one_off_detections && show_small_boxes) {
               cv::Rect src_rect_roi(cv::Point2i(prev_box.bbox.x, prev_box.bbox.y),
                  cv::Size(prev_box.bbox.w, prev_box.bbox.h));
               unsigned int const color_history = (255 * prev_box.last_showed_frames_ago) / frames_history;
               color = cv::Scalar(255 - 3 * color_history, 255 - 2 * color_history, 255 - 1 * color_history);
               if (prev_box.mat_obj.size() == src_rect_roi.size()) {
                  prev_box.mat_obj.copyTo(draw_mat(src_rect_roi));
               }
               cv::rectangle(draw_mat, src_rect_roi, color, thickness);
               putText(draw_mat, track_id_str, src_rect_roi.tl() - cv::Point2i(0, 10), cv::FONT_HERSHEY_COMPLEX_SMALL, 0.8, cv::Scalar(0, 0, 0), 1);
            }
         }
      }
   }
};
#endif   // OPENCV

//extern "C" {
#endif   // __cplusplus

/*
   // C - wrappers
   YOLODLL_API void create_detector(char const* cfg_filename, char const* weight_filename, int gpu_id);
   YOLODLL_API void delete_detector();
   YOLODLL_API bbox_t* detect_custom(image_t img, float thresh, bool use_mean, int *result_size);
   YOLODLL_API bbox_t* detect_resized(image_t img, int init_w, int init_h, float thresh, bool use_mean, int *result_size);
   YOLODLL_API bbox_t* detect(image_t img, int *result_size);
   YOLODLL_API image_t load_img(char *image_filename);
   YOLODLL_API void free_img(image_t m);

#ifdef __cplusplus
}  // extern "C"

static std::shared_ptr<void> c_detector_ptr;
static std::vector<bbox_t> c_result_vec;

void create_detector(char const* cfg_filename, char const* weight_filename, int gpu_id) {
   c_detector_ptr = std::make_shared<YOLODLL_API Detector>(cfg_filename, weight_filename, gpu_id);
}

void delete_detector() { c_detector_ptr.reset(); }

bbox_t* detect_custom(image_t img, float thresh, bool use_mean, int *result_size) {
   c_result_vec = static_cast<Detector*>(c_detector_ptr.get())->detect(img, thresh, use_mean);
   *result_size = c_result_vec.size();
   return c_result_vec.data();
}

bbox_t* detect_resized(image_t img, int init_w, int init_h, float thresh, bool use_mean, int *result_size) {
   c_result_vec = static_cast<Detector*>(c_detector_ptr.get())->detect_resized(img, init_w, init_h, thresh, use_mean);
   *result_size = c_result_vec.size();
   return c_result_vec.data();
}

bbox_t* detect(image_t img, int *result_size) {
   return detect_custom(img, 0.24, true, result_size);
}

image_t load_img(char *image_filename) {
   return static_cast<Detector*>(c_detector_ptr.get())->load_image(image_filename);
}
void free_img(image_t m) {
   static_cast<Detector*>(c_detector_ptr.get())->free_image(m);
}

#endif   // __cplusplus
*/