Improved speed of yolo_console_dll.cpp - 40 FPS on 4K using GeForce GTX 960

AlexeyAB

2017-09-12 8fa9f44211e5b06a5e4c0859eb31af6a45492222

 src/yolo_console_dll.cpp

@@ -101,6 +101,7 @@
            protocol == "rtsp://" || protocol == "http://" || protocol == "https:/")   // video network stream
         {
            cv::Mat cap_frame, cur_frame, det_frame, write_frame;
            std::shared_ptr<image_t> det_image;
            std::vector<bbox_t> result_vec, thread_result_vec;
            detector.nms = 0.02; // comment it - if track_id is not required
            std::atomic<bool> consumed, videowrite_ready;
@@ -116,9 +117,10 @@
            std::chrono::steady_clock::time_point steady_start, steady_end;
            cv::VideoCapture cap(filename); cap >> cur_frame;
            int const video_fps = cap.get(CV_CAP_PROP_FPS);
            cv::Size const frame_size = cur_frame.size();
            cv::VideoWriter output_video;
            if (save_output_videofile)
               output_video.open(out_videofile, CV_FOURCC('D', 'I', 'V', 'X'), std::max(35, video_fps), cur_frame.size(), true);
               output_video.open(out_videofile, CV_FOURCC('D', 'I', 'V', 'X'), std::max(35, video_fps), frame_size, true);

            while (!cur_frame.empty()) {
               if (t_cap.joinable()) {
@@ -132,7 +134,7 @@
               if(consumed)
               {
                  std::unique_lock<std::mutex> lock(mtx);
                  cur_frame.copyTo(det_frame);
                  det_image = detector.mat_to_image_resize(cur_frame);
                  result_vec = thread_result_vec;
                  result_vec = detector.tracking(result_vec);  // comment it - if track_id is not required
                  consumed = false;
@@ -140,14 +142,14 @@
               // launch thread once
               if (!t_detect.joinable()) {
                  t_detect = std::thread([&]() {
                     cv::Mat current_mat = det_frame.clone();
                     auto current_image = det_image;
                     consumed = true;
                     while (!current_mat.empty()) {
                        auto result = detector.detect(current_mat, 0.24, true);
                     while (current_image.use_count() > 0) {
                        auto result = detector.detect_resized(*current_image, frame_size, 0.24, true);
                        ++fps_det_counter;
                        std::unique_lock<std::mutex> lock(mtx);
                        thread_result_vec = result;
                        current_mat = det_frame.clone();
                        current_image = det_image;
                        consumed = true;
                        cv.notify_all();
                     }