main_vot.cpp

   1 #include <stdlib.h>
   2 #include <getopt.h>
   3 #include <libgen.h>
   4 #include <unistd.h>
   5 #include <iomanip>
   6
   7 #include "kcf.h"
   8 #include "vot.hpp"
   9
  10 double calcAccuracy(std::string line, cv::Rect bb_rect, cv::Rect &groundtruth_rect)
  11 {
  12     std::vector<float> numbers;
  13     std::istringstream s(line);
  14     float x;
  15     char ch;
  16
  17     while (s >> x) {
  18         numbers.push_back(x);
  19         s >> ch;
  20     }
  21     double x1 = std::min(numbers[0], std::min(numbers[2], std::min(numbers[4], numbers[6])));
  22     double x2 = std::max(numbers[0], std::max(numbers[2], std::max(numbers[4], numbers[6])));
  23     double y1 = std::min(numbers[1], std::min(numbers[3], std::min(numbers[5], numbers[7])));
  24     double y2 = std::max(numbers[1], std::max(numbers[3], std::max(numbers[5], numbers[7])));
  25
  26     groundtruth_rect = cv::Rect(x1, y1, x2 - x1, y2 - y1);
  27
  28     double rects_intersection = (groundtruth_rect & bb_rect).area();
  29     double rects_union = (groundtruth_rect | bb_rect).area();
  30     double accuracy = rects_intersection / rects_union;
  31
  32     return accuracy;
  33 }
  34
  35 int main(int argc, char *argv[])
  36 {
  37     //load region, images and prepare for output
  38     std::string region, images, output, video_out;
  39     int visualize_delay = -1, fit_size_x = -1, fit_size_y = -1;
  40     KCF_Tracker tracker;
  41     cv::VideoWriter videoWriter;
  42
  43     while (1) {
  44         int option_index = 0;
  45         static struct option long_options[] = {
  46             {"debug",     no_argument,       0,  'd' },
  47             {"visual_debug", optional_argument,    0, 'p'},
  48             {"help",      no_argument,       0,  'h' },
  49             {"output",    required_argument, 0,  'o' },
  50             {"video_out", optional_argument, 0,  'O' },
  51             {"visualize", optional_argument, 0,  'v' },
  52             {"fit",       optional_argument, 0,  'f' },
  53             {0,           0,                 0,  0 }
  54         };
  55
  56         int c = getopt_long(argc, argv, "dp::hv::f::o:O::", long_options, &option_index);
  57         if (c == -1)
  58             break;
  59
  60         switch (c) {
  61         case 'd':
  62             tracker.m_debug = true;
  63             break;
  64         case 'p':
  65             if (!optarg || *optarg == 'p')
  66                 tracker.m_visual_debug = KCF_Tracker::vd::PATCH;
  67             else if (optarg && *optarg == 'r')
  68                 tracker.m_visual_debug = KCF_Tracker::vd::RESPONSE;
  69             else {
  70                 fprintf(stderr, "Unknown visual debug mode: %c", *optarg);
  71                 return 1;
  72             }
  73             break;
  74         case 'h':
  75             std::cerr << "Usage: \n"
  76                       << argv[0] << " [options]\n"
  77                       << argv[0] << " [options] <directory>\n"
  78                       << argv[0] << " [options] <path/to/region.txt or groundtruth.txt> <path/to/images.txt> [path/to/output.txt]\n"
  79                       << "Options:\n"
  80                       << " --visualize    | -v[delay_ms]\n"
  81                       << " --output       | -o <output.txt>\n"
  82                       << " --fit          | -f[W[xH]]\n"
  83                       << " --debug        | -d\n"
  84                       << " --visual_debug | -p [p|r]\n";
  85             exit(0);
  86             break;
  87         case 'o':
  88             output = optarg;
  89             break;
  90         case 'O':
  91             video_out = optarg ? optarg : "./output.avi";
  92             break;
  93         case 'v':
  94             visualize_delay = optarg ? atol(optarg) : 1;
  95             break;
  96         case 'f':
  97             if (!optarg) {
  98                 fit_size_x = fit_size_y = 0;
  99             } else {
 100                 char tail;
 101                 if (sscanf(optarg, "%d%c", &fit_size_x, &tail) == 1) {
 102                     fit_size_y = fit_size_x;
 103                 } else if (sscanf(optarg, "%dx%d%c", &fit_size_x, &fit_size_y, &tail) != 2) {
 104                     fprintf(stderr, "Cannot parse -f argument: %s\n", optarg);
 105                     return 1;
 106                 }
 107             }
 108             break;
 109         }
 110     }
 111
 112     switch (argc - optind) {
 113     case 1:
 114         if (chdir(argv[optind]) == -1) {
 115             perror(argv[optind]);
 116             exit(1);
 117         }
 118         // Fall through
 119     case 0:
 120         region = access("groundtruth.txt", F_OK) == 0 ? "groundtruth.txt" : "region.txt";
 121         images = "images.txt";
 122         if (output.empty())
 123             output = "output.txt";
 124         break;
 125     case 2:
 126         // Fall through
 127     case 3:
 128         region = std::string(argv[optind + 0]);
 129         images = std::string(argv[optind + 1]);
 130         if (output.empty()) {
 131             if ((argc - optind) == 3)
 132                 output = std::string(argv[optind + 2]);
 133             else
 134                 output = std::string(dirname(argv[optind + 0])) + "/output.txt";
 135         }
 136         break;
 137     default:
 138         std::cerr << "Too many arguments\n";
 139         return 1;
 140     }
 141     VOT vot_io(region, images, output);
 142
 143     // if groundtruth.txt is used use intersection over union (IOU) to calculate tracker accuracy
 144     std::ifstream groundtruth_stream;
 145     if (region.compare("groundtruth.txt") == 0) {
 146         groundtruth_stream.open(region.c_str());
 147         std::string line;
 148         std::getline(groundtruth_stream, line);
 149     }
 150
 151     cv::Mat image;
 152
 153     //img = firts frame, initPos = initial position in the first frame
 154     cv::Rect init_rect = vot_io.getInitRectangle();
 155     vot_io.outputBoundingBox(init_rect);
 156     vot_io.getNextImage(image);
 157
 158     if (!video_out.empty()) {
 159         int codec = CV_FOURCC('M', 'J', 'P', 'G');  // select desired codec (must be available at runtime)
 160         double fps = 25.0;                          // framerate of the created video stream
 161         videoWriter.open(video_out, codec, fps, image.size(), true);
 162     }
 163
 164     tracker.init(image, init_rect, fit_size_x, fit_size_y);
 165
 166
 167     BBox_c bb;
 168     cv::Rect bb_rect;
 169     double avg_time = 0., sum_accuracy = 0.;
 170     int frames = 0;
 171
 172     std::cout << std::fixed << std::setprecision(2);
 173
 174     while (vot_io.getNextImage(image) == 1){
 175         double time_profile_counter = cv::getCPUTickCount();
 176         tracker.track(image);
 177         time_profile_counter = cv::getCPUTickCount() - time_profile_counter;
 178          std::cout << "  -> speed : " <<  time_profile_counter/((double)cvGetTickFrequency()*1000) << "ms per frame, "
 179                       "response : " << tracker.getFilterResponse();
 180         avg_time += time_profile_counter/((double)cvGetTickFrequency()*1000);
 181         frames++;
 182
 183         bb = tracker.getBBox();
 184         bb_rect = cv::Rect(bb.cx - bb.w/2., bb.cy - bb.h/2., bb.w, bb.h);
 185         vot_io.outputBoundingBox(bb_rect);
 186
 187         if (groundtruth_stream.is_open()) {
 188             std::string line;
 189             std::getline(groundtruth_stream, line);
 190
 191             cv::Rect groundtruthRect;
 192             double accuracy = calcAccuracy(line, bb_rect, groundtruthRect);
 193             if (visualize_delay >= 0)
 194                 cv::rectangle(image, groundtruthRect, CV_RGB(255, 0,0), 1);
 195             std::cout << ", accuracy: " << accuracy;
 196             sum_accuracy += accuracy;
 197         }
 198
 199         std::cout << std::endl;
 200
 201         if (visualize_delay >= 0 || !video_out.empty()) {
 202             cv::Point pt(bb.cx, bb.cy);
 203             cv::Size size(bb.w, bb.h);
 204             cv::RotatedRect rotatedRectangle(pt, size, bb.a);
 205
 206             cv::Point2f vertices[4];
 207             rotatedRectangle.points(vertices);
 208
 209             for (int i = 0; i < 4; i++)
 210                 cv::line(image, vertices[i], vertices[(i + 1) % 4], cv::Scalar(0, 255, 0), 2);
 211             if (visualize_delay >= 0) {
 212                 cv::imshow("KCF output", image);
 213                 int ret = cv::waitKey(visualize_delay);
 214                 if ((visualize_delay > 0 && ret != -1 && ret < 128) ||
 215                         (visualize_delay == 0 && (ret == 27 /*esc*/ || ret == 'q')))
 216                     break;
 217             }
 218             if (!video_out.empty())
 219                 videoWriter << image;
 220         }
 221
 222 //        std::stringstream s;
 223 //        std::string ss;
 224 //        int countTmp = frames;
 225 //        s << "imgs" << "/img" << (countTmp/10000);
 226 //        countTmp = countTmp%10000;
 227 //        s << (countTmp/1000);
 228 //        countTmp = countTmp%1000;
 229 //        s << (countTmp/100);
 230 //        countTmp = countTmp%100;
 231 //        s << (countTmp/10);
 232 //        countTmp = countTmp%10;
 233 //        s << (countTmp);
 234 //        s << ".jpg";
 235 //        s >> ss;
 236 //        //set image output parameters
 237 //        std::vector<int> compression_params;
 238 //        compression_params.push_back(CV_IMWRITE_JPEG_QUALITY);
 239 //        compression_params.push_back(90);
 240 //        cv::imwrite(ss.c_str(), image, compression_params);
 241     }
 242
 243     std::cout << "Average processing speed: " << avg_time / frames << "ms (" << 1. / (avg_time / frames) * 1000 << " fps)";
 244     if (groundtruth_stream.is_open()) {
 245         std::cout << "; Average accuracy: " << sum_accuracy/frames << std::endl;
 246         groundtruth_stream.close();
 247     }
 248     if (!video_out.empty())
 249        videoWriter.release();
 250     std::cout << std::endl;
 251
 252     return EXIT_SUCCESS;
 253 }