Ignore empty extra lines in train.txt

AlexeyAB

2018-05-29 ec766fc3f114dd4faaa3add500776a7bae957ba5

 src/yolo_console_dll.cpp

@@ -1,33 +1,216 @@
#include <iostream>
#include <iomanip> 
#include <string>
#include <vector>
#include <queue>
#include <fstream>
#include <thread>
#include <atomic>
#include <mutex>              // std::mutex, std::unique_lock
#include <condition_variable> // std::condition_variable

//#define OPENCV
#ifdef _WIN32
#define OPENCV
#define GPU
#endif

// To use tracking - uncomment the following line. Tracking is supported only by OpenCV 3.x
//#define TRACK_OPTFLOW

//#include "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.1\include\cuda_runtime.h"
//#pragma comment(lib, "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v9.1/lib/x64/cudart.lib")
//static std::shared_ptr<image_t> device_ptr(NULL, [](void *img) { cudaDeviceReset(); });

#include "yolo_v2_class.hpp"  // imported functions from DLL


#ifdef OPENCV
#include <opencv2/opencv.hpp>       // C++
#pragma comment(lib, "opencv_core249.lib")
#pragma comment(lib, "opencv_imgproc249.lib")
#pragma comment(lib, "opencv_highgui249.lib")
void draw_boxes(cv::Mat mat_img, std::vector<bbox_t> result_vec) {
   for (auto &i : result_vec) {
      cv::rectangle(mat_img, cv::Rect(i.x, i.y, i.w, i.h), cv::Scalar(50, 200, 50), 3);
#include "opencv2/core/version.hpp"
#ifndef CV_VERSION_EPOCH
#include "opencv2/videoio/videoio.hpp"
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_MAJOR)"" CVAUX_STR(CV_VERSION_MINOR)"" CVAUX_STR(CV_VERSION_REVISION)
#pragma comment(lib, "opencv_world" OPENCV_VERSION ".lib")
#ifdef TRACK_OPTFLOW
#pragma comment(lib, "opencv_cudaoptflow" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_cudaimgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
#endif   // TRACK_OPTFLOW
#else
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_EPOCH)""CVAUX_STR(CV_VERSION_MAJOR)""CVAUX_STR(CV_VERSION_MINOR)
#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
#endif   // CV_VERSION_EPOCH

class track_kalman {
public:
   cv::KalmanFilter kf;
   int state_size, meas_size, contr_size;


   track_kalman(int _state_size = 10, int _meas_size = 10, int _contr_size = 0)
      : state_size(_state_size), meas_size(_meas_size), contr_size(_contr_size)
   {
      kf.init(state_size, meas_size, contr_size, CV_32F);

      cv::setIdentity(kf.measurementMatrix);
      cv::setIdentity(kf.measurementNoiseCov, cv::Scalar::all(1e-1));
      cv::setIdentity(kf.processNoiseCov, cv::Scalar::all(1e-5));
      cv::setIdentity(kf.errorCovPost, cv::Scalar::all(1e-2));
      cv::setIdentity(kf.transitionMatrix);
   }
   cv::imshow("window name", mat_img);
   cv::waitKey(0);

   void set(std::vector<bbox_t> result_vec) {
      for (size_t i = 0; i < result_vec.size() && i < state_size*2; ++i) {
         kf.statePost.at<float>(i * 2 + 0) = result_vec[i].x;
         kf.statePost.at<float>(i * 2 + 1) = result_vec[i].y;
      }
   }

   // Kalman.correct() calculates: statePost = statePre + gain * (z(k)-measurementMatrix*statePre);
   // corrected state (x(k)): x(k)=x'(k)+K(k)*(z(k)-H*x'(k))
   std::vector<bbox_t> correct(std::vector<bbox_t> result_vec) {
      cv::Mat measurement(meas_size, 1, CV_32F);
      for (size_t i = 0; i < result_vec.size() && i < meas_size * 2; ++i) {
         measurement.at<float>(i * 2 + 0) = result_vec[i].x;
         measurement.at<float>(i * 2 + 1) = result_vec[i].y;
      }
      cv::Mat estimated = kf.correct(measurement);
      for (size_t i = 0; i < result_vec.size() && i < meas_size * 2; ++i) {
         result_vec[i].x = estimated.at<float>(i * 2 + 0);
         result_vec[i].y = estimated.at<float>(i * 2 + 1);
      }
      return result_vec;
   }

   // Kalman.predict() calculates: statePre = TransitionMatrix * statePost;
   // predicted state (x'(k)): x(k)=A*x(k-1)+B*u(k)
   std::vector<bbox_t> predict() {
      std::vector<bbox_t> result_vec;
      cv::Mat control;
      cv::Mat prediction = kf.predict(control);
      for (size_t i = 0; i < prediction.rows && i < state_size * 2; ++i) {
         result_vec[i].x = prediction.at<float>(i * 2 + 0);
         result_vec[i].y = prediction.at<float>(i * 2 + 1);
      }
      return result_vec;
   }

};




class extrapolate_coords_t {
public:
   std::vector<bbox_t> old_result_vec;
   std::vector<float> dx_vec, dy_vec, time_vec;
   std::vector<float> old_dx_vec, old_dy_vec;

   void new_result(std::vector<bbox_t> new_result_vec, float new_time) {
      old_dx_vec = dx_vec;
      old_dy_vec = dy_vec;
      if (old_dx_vec.size() != old_result_vec.size()) std::cout << "old_dx != old_res \n";
      dx_vec = std::vector<float>(new_result_vec.size(), 0);
      dy_vec = std::vector<float>(new_result_vec.size(), 0);
      update_result(new_result_vec, new_time, false);
      old_result_vec = new_result_vec;
      time_vec = std::vector<float>(new_result_vec.size(), new_time);
   }

   void update_result(std::vector<bbox_t> new_result_vec, float new_time, bool update = true) {
      for (size_t i = 0; i < new_result_vec.size(); ++i) {
         for (size_t k = 0; k < old_result_vec.size(); ++k) {
            if (old_result_vec[k].track_id == new_result_vec[i].track_id && old_result_vec[k].obj_id == new_result_vec[i].obj_id) {
               float const delta_time = new_time - time_vec[k];
               if (abs(delta_time) < 1) break;
               size_t index = (update) ? k : i;
               float dx = ((float)new_result_vec[i].x - (float)old_result_vec[k].x) / delta_time;
               float dy = ((float)new_result_vec[i].y - (float)old_result_vec[k].y) / delta_time;
               float old_dx = dx, old_dy = dy;

               // if it's shaking
               if (update) {
                  if (dx * dx_vec[i] < 0) dx = dx / 2;
                  if (dy * dy_vec[i] < 0) dy = dy / 2;
               } else {
                  if (dx * old_dx_vec[k] < 0) dx = dx / 2;
                  if (dy * old_dy_vec[k] < 0) dy = dy / 2;
               }
               dx_vec[index] = dx;
               dy_vec[index] = dy;

               //if (old_dx == dx && old_dy == dy) std::cout << "not shakin \n";
               //else std::cout << "shakin \n";

               if (dx_vec[index] > 1000 || dy_vec[index] > 1000) {
                  //std::cout << "!!! bad dx or dy, dx = " << dx_vec[index] << ", dy = " << dy_vec[index] << 
                  // ", delta_time = " << delta_time << ", update = " << update << std::endl;
                  dx_vec[index] = 0;
                  dy_vec[index] = 0;						
               }
               old_result_vec[k].x = new_result_vec[i].x;
               old_result_vec[k].y = new_result_vec[i].y;
               time_vec[k] = new_time;
               break;
            }
         }
      }
   }

   std::vector<bbox_t> predict(float cur_time) {
      std::vector<bbox_t> result_vec = old_result_vec;
      for (size_t i = 0; i < old_result_vec.size(); ++i) {
         float const delta_time = cur_time - time_vec[i];
         auto &bbox = result_vec[i];
         float new_x = (float) bbox.x + dx_vec[i] * delta_time;
         float new_y = (float) bbox.y + dy_vec[i] * delta_time;
         if (new_x > 0) bbox.x = new_x;
         else bbox.x = 0;
         if (new_y > 0) bbox.y = new_y;
         else bbox.y = 0;
      }
      return result_vec;
   }

};


void draw_boxes(cv::Mat mat_img, std::vector<bbox_t> result_vec, std::vector<std::string> obj_names, 
   int current_det_fps = -1, int current_cap_fps = -1)
{
   int const colors[6][3] = { { 1,0,1 },{ 0,0,1 },{ 0,1,1 },{ 0,1,0 },{ 1,1,0 },{ 1,0,0 } };

   for (auto &i : result_vec) {
      cv::Scalar color = obj_id_to_color(i.obj_id);
      cv::rectangle(mat_img, cv::Rect(i.x, i.y, i.w, i.h), color, 2);
      if (obj_names.size() > i.obj_id) {
         std::string obj_name = obj_names[i.obj_id];
         if (i.track_id > 0) obj_name += " - " + std::to_string(i.track_id);
         cv::Size const text_size = getTextSize(obj_name, cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, 2, 0);
         int const max_width = (text_size.width > i.w + 2) ? text_size.width : (i.w + 2);
         cv::rectangle(mat_img, cv::Point2f(std::max((int)i.x - 1, 0), std::max((int)i.y - 30, 0)), 
            cv::Point2f(std::min((int)i.x + max_width, mat_img.cols-1), std::min((int)i.y, mat_img.rows-1)), 
            color, CV_FILLED, 8, 0);
         putText(mat_img, obj_name, cv::Point2f(i.x, i.y - 10), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(0, 0, 0), 2);
      }
   }
   if (current_det_fps >= 0 && current_cap_fps >= 0) {
      std::string fps_str = "FPS detection: " + std::to_string(current_det_fps) + "   FPS capture: " + std::to_string(current_cap_fps);
      putText(mat_img, fps_str, cv::Point2f(10, 20), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(50, 255, 0), 2);
   }
}
#endif   // OPENCV

void show_result(std::vector<bbox_t> const result_vec, std::vector<std::string> const obj_names) {

void show_console_result(std::vector<bbox_t> const result_vec, std::vector<std::string> const obj_names) {
   for (auto &i : result_vec) {
      if (obj_names.size() > i.obj_id) std::cout << obj_names[i.obj_id] << " - ";
      std::cout << "obj_id = " << i.obj_id << " - x = " << i.x << ", y = " << i.y 
      std::cout << "obj_id = " << i.obj_id << ",  x = " << i.x << ", y = " << i.y 
         << ", w = " << i.w << ", h = " << i.h
         << ", prob = " << i.prob << std::endl;
         << std::setprecision(3) << ", prob = " << i.prob << std::endl;
   }
}

@@ -35,41 +218,270 @@
   std::ifstream file(filename);
   std::vector<std::string> file_lines;
   if (!file.is_open()) return file_lines;
   for(std::string line; file >> line;) file_lines.push_back(line);
   for(std::string line; getline(file, line);) file_lines.push_back(line);
   std::cout << "object names loaded \n";
   return file_lines;
}


int main() 
int main(int argc, char *argv[])
{
   Detector detector("yolo-voc.cfg", "yolo-voc.weights");
   std::string  names_file = "data/coco.names";
   std::string  cfg_file = "cfg/yolov3.cfg";
   std::string  weights_file = "yolov3.weights";
   std::string filename;

   auto obj_names = objects_names_from_file("data/voc.names");
   if (argc > 4) {   //voc.names yolo-voc.cfg yolo-voc.weights test.mp4		
      names_file = argv[1];
      cfg_file = argv[2];
      weights_file = argv[3];
      filename = argv[4];
   }
   else if (argc > 1) filename = argv[1];

   float const thresh = (argc > 5) ? std::stof(argv[5]) : 0.20;

   Detector detector(cfg_file, weights_file);

   auto obj_names = objects_names_from_file(names_file);
   std::string out_videofile = "result.avi";
   bool const save_output_videofile = true;
#ifdef TRACK_OPTFLOW
   Tracker_optflow tracker_flow;
   detector.wait_stream = true;
#endif

   while (true) 
   {
      std::string filename;
      std::cout << "input image filename: ";
      std::cin >> filename;
   {		
      std::cout << "input image or video filename: ";
      if(filename.size() == 0) std::cin >> filename;
      if (filename.size() == 0) break;
      
      try {
#ifdef OPENCV
         cv::Mat mat_img = cv::imread(filename);
         std::vector<bbox_t> result_vec = detector.detect(mat_img);
         draw_boxes(mat_img, result_vec);
         extrapolate_coords_t extrapolate_coords;
         bool extrapolate_flag = false;
         float cur_time_extrapolate = 0, old_time_extrapolate = 0;
         preview_boxes_t large_preview(100, 150, false), small_preview(50, 50, true);
         bool show_small_boxes = false;

         std::string const file_ext = filename.substr(filename.find_last_of(".") + 1);
         std::string const protocol = filename.substr(0, 7);
         if (file_ext == "avi" || file_ext == "mp4" || file_ext == "mjpg" || file_ext == "mov" ||  // video file
            protocol == "rtmp://" || protocol == "rtsp://" || protocol == "http://" || protocol == "https:/")  // video network stream
         {
            cv::Mat cap_frame, cur_frame, det_frame, write_frame;
            std::queue<cv::Mat> track_optflow_queue;
            int passed_flow_frames = 0;
            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;
            bool exit_flag = false;
            consumed = true;
            videowrite_ready = true;
            std::atomic<int> fps_det_counter, fps_cap_counter;
            fps_det_counter = 0;
            fps_cap_counter = 0;
            int current_det_fps = 0, current_cap_fps = 0;
            std::thread t_detect, t_cap, t_videowrite;
            std::mutex mtx;
            std::condition_variable cv_detected, cv_pre_tracked;
            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), frame_size, true);

            while (!cur_frame.empty()) 
            {
               // always sync
               if (t_cap.joinable()) {
                  t_cap.join();
                  ++fps_cap_counter;
                  cur_frame = cap_frame.clone();
               }
               t_cap = std::thread([&]() { cap >> cap_frame; });
               ++cur_time_extrapolate;

               // swap result bouned-boxes and input-frame
               if(consumed)
               {
                  std::unique_lock<std::mutex> lock(mtx);
                  det_image = detector.mat_to_image_resize(cur_frame);
                  auto old_result_vec = detector.tracking_id(result_vec);
                  auto detected_result_vec = thread_result_vec;
                  result_vec = detected_result_vec;
#ifdef TRACK_OPTFLOW
                  // track optical flow
                  if (track_optflow_queue.size() > 0) {
                     //std::cout << "\n !!!! all = " << track_optflow_queue.size() << ", cur = " << passed_flow_frames << std::endl;
                     cv::Mat first_frame = track_optflow_queue.front();
                     tracker_flow.update_tracking_flow(track_optflow_queue.front(), result_vec);

                     while (track_optflow_queue.size() > 1) {
                        track_optflow_queue.pop();
                        result_vec = tracker_flow.tracking_flow(track_optflow_queue.front(), true);
                     }
                     track_optflow_queue.pop();
                     passed_flow_frames = 0;

                     result_vec = detector.tracking_id(result_vec);
                     auto tmp_result_vec = detector.tracking_id(detected_result_vec, false);
                     small_preview.set(first_frame, tmp_result_vec);

                     extrapolate_coords.new_result(tmp_result_vec, old_time_extrapolate);
                     old_time_extrapolate = cur_time_extrapolate;
                     extrapolate_coords.update_result(result_vec, cur_time_extrapolate - 1);
                  }
#else
                  result_vec = detector.tracking_id(result_vec);  // comment it - if track_id is not required					
                  extrapolate_coords.new_result(result_vec, cur_time_extrapolate - 1);
#endif
                  // add old tracked objects
                  for (auto &i : old_result_vec) {
                     auto it = std::find_if(result_vec.begin(), result_vec.end(),
                        [&i](bbox_t const& b) { return b.track_id == i.track_id && b.obj_id == i.obj_id; });
                     bool track_id_absent = (it == result_vec.end());
                     if (track_id_absent) {
                        if (i.frames_counter-- > 1)
                           result_vec.push_back(i);
                     }
                     else {
                        it->frames_counter = std::min((unsigned)3, i.frames_counter + 1);
                     }
                  }
#ifdef TRACK_OPTFLOW
                  tracker_flow.update_cur_bbox_vec(result_vec);
                  result_vec = tracker_flow.tracking_flow(cur_frame, true);   // track optical flow
#endif
                  consumed = false;
                  cv_pre_tracked.notify_all();
               }
               // launch thread once - Detection
               if (!t_detect.joinable()) {
                  t_detect = std::thread([&]() {
                     auto current_image = det_image;
                     consumed = true;
                     while (current_image.use_count() > 0 && !exit_flag) {
                        auto result = detector.detect_resized(*current_image, frame_size.width, frame_size.height, 
                           thresh, false);   // true
                        ++fps_det_counter;
                        std::unique_lock<std::mutex> lock(mtx);
                        thread_result_vec = result;
                        consumed = true;
                        cv_detected.notify_all();
                        if (detector.wait_stream) {
                           while (consumed && !exit_flag) cv_pre_tracked.wait(lock);
                        }
                        current_image = det_image;
                     }
                  });
               }
               //while (!consumed); // sync detection

               if (!cur_frame.empty()) {
                  steady_end = std::chrono::steady_clock::now();
                  if (std::chrono::duration<double>(steady_end - steady_start).count() >= 1) {
                     current_det_fps = fps_det_counter;
                     current_cap_fps = fps_cap_counter;
                     steady_start = steady_end;
                     fps_det_counter = 0;
                     fps_cap_counter = 0;
                  }

                  large_preview.set(cur_frame, result_vec);
#ifdef TRACK_OPTFLOW
                  ++passed_flow_frames;
                  track_optflow_queue.push(cur_frame.clone());
                  result_vec = tracker_flow.tracking_flow(cur_frame);   // track optical flow
                  extrapolate_coords.update_result(result_vec, cur_time_extrapolate);
                  small_preview.draw(cur_frame, show_small_boxes);
#endif						
                  auto result_vec_draw = result_vec;
                  if (extrapolate_flag) {
                     result_vec_draw = extrapolate_coords.predict(cur_time_extrapolate);
                     cv::putText(cur_frame, "extrapolate", cv::Point2f(10, 40), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.0, cv::Scalar(50, 50, 0), 2);
                  }
                  draw_boxes(cur_frame, result_vec_draw, obj_names, current_det_fps, current_cap_fps);
                  //show_console_result(result_vec, obj_names);
                  large_preview.draw(cur_frame);

                  cv::imshow("window name", cur_frame);
                  int key = cv::waitKey(3);  // 3 or 16ms
                  if (key == 'f') show_small_boxes = !show_small_boxes;
                  if (key == 'p') while (true) if(cv::waitKey(100) == 'p') break;
                  if (key == 'e') extrapolate_flag = !extrapolate_flag;
                  if (key == 27) { exit_flag = true; break; }

                  if (output_video.isOpened() && videowrite_ready) {
                     if (t_videowrite.joinable()) t_videowrite.join();
                     write_frame = cur_frame.clone();
                     videowrite_ready = false;
                     t_videowrite = std::thread([&]() { 
                         output_video << write_frame; videowrite_ready = true;
                     });
                  }
               }

#ifndef TRACK_OPTFLOW
               // wait detection result for video-file only (not for net-cam)
               if (protocol != "rtsp://" && protocol != "http://" && protocol != "https:/") {
                  std::unique_lock<std::mutex> lock(mtx);
                  while (!consumed) cv_detected.wait(lock);
               }
#endif
            }
            exit_flag = true;
            if (t_cap.joinable()) t_cap.join();
            if (t_detect.joinable()) t_detect.join();
            if (t_videowrite.joinable()) t_videowrite.join();
            std::cout << "Video ended \n";
            break;
         }
         else if (file_ext == "txt") { // list of image files
            std::ifstream file(filename);
            if (!file.is_open()) std::cout << "File not found! \n";
            else 
               for (std::string line; file >> line;) {
                  std::cout << line << std::endl;
                  cv::Mat mat_img = cv::imread(line);
                  std::vector<bbox_t> result_vec = detector.detect(mat_img);
                  show_console_result(result_vec, obj_names);
                  //draw_boxes(mat_img, result_vec, obj_names);
                  //cv::imwrite("res_" + line, mat_img);
               }
				
         }
         else {   // image file
            cv::Mat mat_img = cv::imread(filename);
				
            auto start = std::chrono::steady_clock::now();
            std::vector<bbox_t> result_vec = detector.detect(mat_img);
            auto end = std::chrono::steady_clock::now();
            std::chrono::duration<double> spent = end - start;
            std::cout << " Time: " << spent.count() << " sec \n";

            //result_vec = detector.tracking_id(result_vec);   // comment it - if track_id is not required
            draw_boxes(mat_img, result_vec, obj_names);
            cv::imshow("window name", mat_img);
            show_console_result(result_vec, obj_names);
            cv::waitKey(0);
         }
#else
         //std::vector<bbox_t> result_vec = detector.detect(filename);

         auto img = detector.load_image(filename);
         std::vector<bbox_t> result_vec = detector.detect(img);
         detector.free_image(img);
#endif
         show_result(result_vec, obj_names);
         show_console_result(result_vec, obj_names);
#endif			
      }
      catch (std::exception &e) { std::cerr << "exception: " << e.what() << "\n"; getchar(); }
      catch (...) { std::cerr << "unknown exception \n"; getchar(); }
      filename.clear();
   }

   return 0;