Added CUDA-streams to Darknet-Yolo forward inference

AlexeyAB

2017-12-31 9d23aad8696268e8ce3a94fee9490fd1db000dc8

 src/yolo_v2_class.hpp

@@ -8,6 +8,11 @@
#include <opencv2/opencv.hpp>       // C++
#include "opencv2/highgui/highgui_c.h" // C
#include "opencv2/imgproc/imgproc_c.h" // C

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

#ifdef YOLODLL_EXPORTS
@@ -41,6 +46,7 @@

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

@@ -93,6 +99,7 @@
   }

private:

   static image_t ipl_to_image(IplImage* src)
   {
      unsigned char *data = (unsigned char *)src->imageData;
@@ -142,8 +149,165 @@

#endif   // OPENCV

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


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

class Tracker_optflow {
public:


   // just to avoid extra allocations
   cv::cuda::GpuMat src_mat_gpu;
   cv::cuda::GpuMat dst_mat_gpu, dst_grey_gpu;
   cv::cuda::GpuMat tmp_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;

   void update_tracking_flow(cv::Mat src_mat, int gpu_id = 0)
   {
      int const old_gpu_id = cv::cuda::getDevice();
      static const int gpu_count = cv::cuda::getCudaEnabledDeviceCount();
      if (gpu_count > gpu_id)
         cv::cuda::setDevice(gpu_id);

      cv::cuda::Stream stream;

      if (sync_PyrLKOpticalFlow_gpu.empty()) {
         sync_PyrLKOpticalFlow_gpu = cv::cuda::SparsePyrLKOpticalFlow::create();

         //sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(31, 31)); //sync_PyrLKOpticalFlow_gpu.winSize = cv::Size(31, 31);
         //sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(15, 15)); //sync_PyrLKOpticalFlow_gpu.winSize = cv::Size(15, 15);
         sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(21, 21));
         sync_PyrLKOpticalFlow_gpu->setMaxLevel(3);   //sync_PyrLKOpticalFlow_gpu.maxLevel = 8; // +-32 points // def: 3
         sync_PyrLKOpticalFlow_gpu->setNumIters(6000);   //sync_PyrLKOpticalFlow_gpu.iters = 8000; // def: 30
         //??? //sync_PyrLKOpticalFlow_gpu.getMinEigenVals = true;
         //std::cout << "sync_PyrLKOpticalFlow_gpu.maxLevel: " << sync_PyrLKOpticalFlow_gpu.maxLevel << std::endl;
         //std::cout << "sync_PyrLKOpticalFlow_gpu.iters: " << sync_PyrLKOpticalFlow_gpu.iters << std::endl;
         //std::cout << "sync_PyrLKOpticalFlow_gpu.winSize: " << sync_PyrLKOpticalFlow_gpu.winSize << std::endl;
      }

      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);
         }

         src_mat_gpu.upload(src_mat, stream);
         cv::cuda::cvtColor(src_mat_gpu, src_grey_gpu, CV_BGR2GRAY, 0, stream);
         //std::cout << " \n\n OK !!! \n\n";
      }
      cv::cuda::setDevice(old_gpu_id);
   }


   std::vector<bbox_t> tracking_flow(cv::Mat dst_mat, std::vector<bbox_t> cur_bbox_vec, int gpu_id = 0)
   {
      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();
      static const int gpu_count = cv::cuda::getCudaEnabledDeviceCount();
      if (gpu_count > gpu_id)
         cv::cuda::setDevice(gpu_id);

      cv::cuda::Stream stream;

      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);
         tmp_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1);
      }


      dst_mat_gpu.upload(dst_mat, stream);

      cv::cuda::cvtColor(dst_mat_gpu, dst_grey_gpu, CV_BGR2GRAY, 0, 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;
      }

      cv::Mat prev_pts, prev_pts_flow_cpu, cur_pts_flow_cpu;

      for (auto &i : cur_bbox_vec) {
         float x_center = (i.x + i.w / 2);
         float y_center = (i.y + i.h / 2);
         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);


      dst_grey_gpu.copyTo(tmp_grey_gpu, stream);

      //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
                                                                                                      //std::cout << "\n 1-e \n";

      cur_pts_flow_gpu.download(cur_pts_flow_cpu, stream);

      tmp_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;

      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 (err_cpu.cols > i &&  status_cpu.cols > i)
            if (abs(moved_x) < 100 && abs(moved_y) < 100)
               //if (err_cpu.at<float>(0, i) < 60 && status_cpu.at<unsigned char>(0, i) != 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]);
               }
      }

      cv::cuda::setDevice(old_gpu_id);

      return result_bbox_vec;
   }

};
#else

class Tracker_optflow {};

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