11 #include "fft_cufft.h"
14 #include "fft_opencv.h"
22 #define DEBUG_PRINT(obj) \
24 std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl; \
26 #define DEBUG_PRINTM(obj) \
28 std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl \
29 << (obj) << std::endl; \
32 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
33 double output_sigma_factor, int cell_size)
34 : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
35 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size)
39 KCF_Tracker::KCF_Tracker() : fft(*new FFT()) {}
41 KCF_Tracker::~KCF_Tracker()
46 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
48 // check boundary, enforce min size
49 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
51 if (x2 > img.cols - 1) x2 = img.cols - 1;
53 if (y2 > img.rows - 1) y2 = img.rows - 1;
55 if (x2 - x1 < 2 * p_cell_size) {
56 double diff = (2 * p_cell_size - x2 + x1) / 2.;
57 if (x1 - diff >= 0 && x2 + diff < img.cols) {
60 } else if (x1 - 2 * diff >= 0) {
66 if (y2 - y1 < 2 * p_cell_size) {
67 double diff = (2 * p_cell_size - y2 + y1) / 2.;
68 if (y1 - diff >= 0 && y2 + diff < img.rows) {
71 } else if (y1 - 2 * diff >= 0) {
80 p_pose.cx = x1 + p_pose.w / 2.;
81 p_pose.cy = y1 + p_pose.h / 2.;
83 cv::Mat input_gray, input_rgb = img.clone();
84 if (img.channels() == 3) {
85 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
86 input_gray.convertTo(input_gray, CV_32FC1);
88 img.convertTo(input_gray, CV_32FC1);
90 // don't need too large image
91 if (p_pose.w * p_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
92 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
93 p_resize_image = true;
94 p_pose.scale(p_downscale_factor);
95 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
96 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
97 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
98 if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
99 std::cerr << "Fit size does not fit to hog cell size. The dimensions have to be divisible by HOG cell "
101 << p_cell_size << std::endl;
103 std::exit(EXIT_FAILURE);
105 double tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
106 if (fabs(tmp - fit_size_x) > p_floating_error) p_scale_factor_x = fit_size_x / tmp;
107 tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
108 if (fabs(tmp - fit_size_y) > p_floating_error) p_scale_factor_y = fit_size_y / tmp;
109 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
110 << p_scale_factor_y << std::endl;
112 p_pose.scale_x(p_scale_factor_x);
113 p_pose.scale_y(p_scale_factor_y);
114 if (fabs(p_scale_factor_x - 1) > p_floating_error && fabs(p_scale_factor_y - 1) > p_floating_error) {
115 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
116 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
117 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
119 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y,
121 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
126 // compute win size + fit to fhog cell size
127 p_windows_size.width = int(round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size);
128 p_windows_size.height = int(round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size);
131 if (m_use_color) p_num_of_feats += 3;
132 if (m_use_cnfeat) p_num_of_feats += 10;
133 p_roi_width = p_windows_size.width / p_cell_size;
134 p_roi_height = p_windows_size.height / p_cell_size;
138 for (int i = -p_num_scales / 2; i <= p_num_scales / 2; ++i)
139 p_scales.push_back(std::pow(p_scale_step, i));
141 p_scales.push_back(1.);
144 for (int i = -2*p_angle_step; i <=2*p_angle_step ; i += p_angle_step)
145 p_angles.push_back(i);
147 p_angles.push_back(0);
150 if (p_windows_size.height / p_cell_size * (p_windows_size.width / p_cell_size / 2 + 1) > 1024) {
151 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
152 "the window dimensions so its size is less or equal to "
153 << 1024 * p_cell_size * p_cell_size * 2 + 1
154 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
155 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
156 std::exit(EXIT_FAILURE);
159 if (m_use_linearkernel) {
160 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
161 std::exit(EXIT_FAILURE);
163 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
164 p_rot_labels_data = DynMem(
165 ((uint(p_windows_size.width) / p_cell_size) * (uint(p_windows_size.height) / p_cell_size)) * sizeof(float));
166 p_rot_labels = cv::Mat(p_windows_size.height / int(p_cell_size), p_windows_size.width / int(p_cell_size), CV_32FC1,
167 p_rot_labels_data.hostMem());
169 p_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.height / p_cell_size)) / 2 + 1,
173 #if defined(CUFFT) || defined(FFTW)
174 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1,
175 uint(p_num_of_feats));
176 p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1, 1);
177 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size) / 2 + 1,
180 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)),
181 uint(p_num_of_feats));
182 p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)), 1);
183 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size), p_num_of_feats);
186 int max = m_use_big_batch ? 2 : p_num_scales;
187 for (int i = 0; i < max; ++i) {
188 if (m_use_big_batch && i == 1) {
189 p_threadctxs.emplace_back(
190 new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats * p_num_scales, p_num_scales));
192 p_threadctxs.emplace_back(new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats, 1));
196 p_current_scale = 1.;
198 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
199 double max_size_ratio =
200 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
201 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
202 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
203 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
205 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
206 std::cout << "init: win size. " << p_windows_size.width << " " << p_windows_size.height << std::endl;
207 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
209 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
211 fft.init(uint(p_windows_size.width / p_cell_size), uint(p_windows_size.height / p_cell_size), uint(p_num_of_feats),
212 uint(p_num_scales), m_use_big_batch);
213 fft.set_window(cosine_window_function(p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size));
215 // window weights, i.e. labels
217 gaussian_shaped_labels(p_output_sigma, p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size), p_yf,
218 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_threadctxs.front()->stream);
221 // obtain a sub-window for training initial model
222 p_threadctxs.front()->patch_feats.clear();
223 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
224 *p_threadctxs.front());
225 fft.forward_window(p_threadctxs.front()->patch_feats, p_model_xf, p_threadctxs.front()->fw_all,
226 m_use_cuda ? p_threadctxs.front()->data_features.deviceMem() : nullptr, p_threadctxs.front()->stream);
227 DEBUG_PRINTM(p_model_xf);
228 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
229 p_threadctxs.front()->model_xf = p_model_xf;
230 p_threadctxs.front()->model_xf.set_stream(p_threadctxs.front()->stream);
231 p_yf.set_stream(p_threadctxs.front()->stream);
232 p_model_xf.set_stream(p_threadctxs.front()->stream);
233 p_xf.set_stream(p_threadctxs.front()->stream);
236 if (m_use_linearkernel) {
237 ComplexMat xfconj = p_model_xf.conj();
238 p_model_alphaf_num = xfconj.mul(p_yf);
239 p_model_alphaf_den = (p_model_xf * xfconj);
241 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
242 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
243 gaussian_correlation(*p_threadctxs.front(), p_threadctxs.front()->model_xf, p_threadctxs.front()->model_xf,
244 p_kernel_sigma, true);
246 gaussian_correlation(*p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
248 DEBUG_PRINTM(p_threadctxs.front()->kf);
249 p_model_alphaf_num = p_yf * p_threadctxs.front()->kf;
250 DEBUG_PRINTM(p_model_alphaf_num);
251 p_model_alphaf_den = p_threadctxs.front()->kf * (p_threadctxs.front()->kf + float(p_lambda));
252 DEBUG_PRINTM(p_model_alphaf_den);
254 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
255 DEBUG_PRINTM(p_model_alphaf);
256 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
258 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
259 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
260 (*it)->model_xf = p_model_xf;
261 (*it)->model_xf.set_stream((*it)->stream);
262 (*it)->model_alphaf = p_model_alphaf;
263 (*it)->model_alphaf.set_stream((*it)->stream);
268 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
270 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
273 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
275 if (p_resize_image) {
277 tmp.scale(p_downscale_factor);
280 } else if (p_fit_to_pw2) {
282 tmp.scale_x(p_scale_factor_x);
283 tmp.scale_y(p_scale_factor_y);
292 BBox_c KCF_Tracker::getBBox()
295 tmp.w *= p_current_scale;
296 tmp.h *= p_current_scale;
297 tmp.a = p_current_angle;
299 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
301 tmp.scale_x(1 / p_scale_factor_x);
302 tmp.scale_y(1 / p_scale_factor_y);
308 void KCF_Tracker::track(cv::Mat &img)
310 if (m_debug) std::cout << "NEW FRAME" << '\n';
311 cv::Mat input_gray, input_rgb = img.clone();
312 if (img.channels() == 3) {
313 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
314 input_gray.convertTo(input_gray, CV_32FC1);
316 img.convertTo(input_gray, CV_32FC1);
318 // don't need too large image
319 if (p_resize_image) {
320 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
321 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
322 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
323 fabs(p_scale_factor_y - 1) > p_floating_error) {
324 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
325 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
326 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
328 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
329 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
333 double max_response = -1.;
335 cv::Point2i *max_response_pt = nullptr;
336 cv::Mat *max_response_map = nullptr;
338 if (m_use_multithreading) {
339 std::vector<std::future<void>> async_res(p_scales.size());
340 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
341 uint index = uint(std::distance(p_threadctxs.begin(), it));
342 async_res[index] = std::async(std::launch::async, [this, &input_gray, &input_rgb, index, it]() -> void {
343 return scale_track(*(*it), input_rgb, input_gray, this->p_scales[index]);
346 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
347 uint index = uint(std::distance(p_threadctxs.begin(), it));
348 async_res[index].wait();
349 if ((*it)->max_response > max_response) {
350 max_response = (*it)->max_response;
351 max_response_pt = &(*it)->max_loc;
352 max_response_map = &(*it)->response;
353 scale_index = int(index);
357 uint start = m_use_big_batch ? 1 : 0;
358 uint end = m_use_big_batch ? 2 : uint(p_num_scales);
359 NORMAL_OMP_PARALLEL_FOR
360 for (uint i = start; i < end; ++i) {
361 auto it = p_threadctxs.begin();
363 scale_track(*(*it), input_rgb, input_gray, this->p_scales[i]);
365 if (m_use_big_batch) {
366 for (size_t j = 0; j < p_scales.size(); ++j) {
367 if ((*it)->max_responses[j] > max_response) {
368 max_response = (*it)->max_responses[j];
369 max_response_pt = &(*it)->max_locs[j];
370 max_response_map = &(*it)->response_maps[j];
371 scale_index = int(j);
372 //#pragma omp parallel for ordered private(patch_feat) schedule(dynamic)
373 // for (size_t i = 0; i < p_scales.size(); ++i) {
374 // std::cout << "CURRENT SCALE: " << p_current_scale * p_scales[i] << std::endl;
375 // for (size_t j = 0; j < p_angles.size(); ++j) {
376 // patch_feat = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0],
377 // p_windows_size[1], p_current_scale * p_scales[i], p_current_angle + p_angles[j]);
378 // ComplexMat zf = fft.forward_window(patch_feat);
381 // if (m_use_linearkernel)
382 // response = fft.inverse((p_model_alphaf * zf).sum_over_channels());
384 // ComplexMat kzf = gaussian_correlation(zf, p_model_xf, p_kernel_sigma);
385 // DEBUG_PRINTM(p_model_alphaf);
386 // DEBUG_PRINTM(kzf);
387 // DEBUG_PRINTM(p_model_alphaf * kzf);
388 // response = fft.inverse(p_model_alphaf * kzf);
390 // if (m_visual_debug) {
391 // cv::Mat copy_response = response.clone();
393 // // crop the spectrum, if it has an odd number of rows or columns
394 // copy_response = copy_response(cv::Rect(0, 0, copy_response.cols & -2,
395 // copy_response.rows & -2));
397 // // rearrange the quadrants of Fourier image so that the origin is at the image center
398 // int cx = copy_response.cols/2;
399 // int cy = copy_response.rows/2;
401 // cv::Mat q0(copy_response, cv::Rect(0, 0, cx, cy)); // Top-Left - Create a ROI per
402 // quadrant cv::Mat q1(copy_response, cv::Rect(cx, 0, cx, cy)); // Top-Right cv::Mat
403 // q2(copy_response, cv::Rect(0, cy, cx, cy)); // Bottom-Left cv::Mat q3(copy_response,
404 // cv::Rect(cx, cy, cx, cy)); // Bottom-Right
406 // cv::Mat tmp; // swap quadrants (Top-Left with Bottom-Right)
411 // q1.copyTo(tmp); // swap quadrant (Top-Right with Bottom-Left)
415 // cv::namedWindow("Response map",cv::WINDOW_NORMAL);
416 // cv::resizeWindow("Response map", 128, 128);
417 // cv::imshow("Response map", copy_response);
424 if ((*it)->max_response > max_response) {
425 max_response = (*it)->max_response;
426 max_response_pt = &(*it)->max_loc;
427 max_response_map = &(*it)->response;
428 scale_index = int(i);
435 DEBUG_PRINTM(*max_response_map);
436 DEBUG_PRINT(*max_response_pt);
438 // sub pixel quadratic interpolation from neighbours
439 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
440 max_response_pt->y = max_response_pt->y - max_response_map->rows;
441 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
442 max_response_pt->x = max_response_pt->x - max_response_map->cols;
444 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
445 DEBUG_PRINT(new_location);
447 if (m_use_subpixel_localization) new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
448 DEBUG_PRINT(new_location);
450 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
451 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
453 if (p_pose.cx < 0) p_pose.cx = 0;
454 if (p_pose.cx > (img.cols * p_scale_factor_x) - 1) p_pose.cx = (img.cols * p_scale_factor_x) - 1;
455 if (p_pose.cy < 0) p_pose.cy = 0;
456 if (p_pose.cy > (img.rows * p_scale_factor_y) - 1) p_pose.cy = (img.rows * p_scale_factor_y) - 1;
458 if (p_pose.cx < 0) p_pose.cx = 0;
459 if (p_pose.cx > img.cols - 1) p_pose.cx = img.cols - 1;
460 if (p_pose.cy < 0) p_pose.cy = 0;
461 if (p_pose.cy > img.rows - 1) p_pose.cy = img.rows - 1;
464 // sub grid scale interpolation
465 double new_scale = p_scales[uint(scale_index)];
466 if (m_use_subgrid_scale) new_scale = sub_grid_scale(scale_index);
468 p_current_scale *= new_scale;
470 if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
471 if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
473 // TODO Missing angle_index
474 // int tmp_angle = p_current_angle + p_angles[angle_index];
475 // p_current_angle = tmp_angle < 0 ? -std::abs(tmp_angle)%360 : tmp_angle%360;
477 // obtain a subwindow for training at newly estimated target position
478 p_threadctxs.front()->patch_feats.clear();
479 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
480 *p_threadctxs.front(), p_current_scale, p_current_angle);
481 fft.forward_window(p_threadctxs.front()->patch_feats, p_xf, p_threadctxs.front()->fw_all,
482 m_use_cuda ? p_threadctxs.front()->data_features.deviceMem() : nullptr,
483 p_threadctxs.front()->stream);
485 // subsequent frames, interpolate model
486 p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
488 ComplexMat alphaf_num, alphaf_den;
490 if (m_use_linearkernel) {
491 ComplexMat xfconj = p_xf.conj();
492 alphaf_num = xfconj.mul(p_yf);
493 alphaf_den = (p_xf * xfconj);
495 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
496 gaussian_correlation(*p_threadctxs.front(), p_xf, p_xf, p_kernel_sigma,
498 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
499 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
500 alphaf_num = p_yf * p_threadctxs.front()->kf;
501 alphaf_den = p_threadctxs.front()->kf * (p_threadctxs.front()->kf + float(p_lambda));
504 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
505 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
506 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
508 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
509 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
510 (*it)->model_xf = p_model_xf;
511 (*it)->model_xf.set_stream((*it)->stream);
512 (*it)->model_alphaf = p_model_alphaf;
513 (*it)->model_alphaf.set_stream((*it)->stream);
518 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray, double scale)
520 if (m_use_big_batch) {
521 vars.patch_feats.clear();
522 BIG_BATCH_OMP_PARALLEL_FOR
523 for (uint i = 0; i < uint(p_num_scales); ++i) {
524 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
525 this->p_windows_size.height, vars, this->p_current_scale * this->p_scales[i]);
528 vars.patch_feats.clear();
529 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
530 this->p_windows_size.height, vars, this->p_current_scale *scale);
533 fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
535 DEBUG_PRINTM(vars.zf);
537 if (m_use_linearkernel) {
538 vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
539 : (p_model_alphaf * vars.zf).sum_over_channels();
540 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
542 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
543 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
544 vars.kzf = vars.model_alphaf * vars.kzf;
546 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
547 DEBUG_PRINTM(this->p_model_alphaf);
548 DEBUG_PRINTM(vars.kzf);
549 vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
551 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
554 DEBUG_PRINTM(vars.response);
556 /* target location is at the maximum response. we must take into
557 account the fact that, if the target doesn't move, the peak
558 will appear at the top-left corner, not at the center (this is
559 discussed in the paper). the responses wrap around cyclically. */
560 if (m_use_big_batch) {
561 cv::split(vars.response, vars.response_maps);
563 for (size_t i = 0; i < p_scales.size(); ++i) {
564 double min_val, max_val;
565 cv::Point2i min_loc, max_loc;
566 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
567 DEBUG_PRINT(max_loc);
568 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
569 vars.max_responses[i] = max_val * weight;
570 vars.max_locs[i] = max_loc;
575 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
577 DEBUG_PRINT(vars.max_loc);
579 double weight = scale < 1. ? scale : 1. / scale;
580 vars.max_response = vars.max_val * weight;
585 // ****************************************************************************
587 void KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y,
588 ThreadCtx &vars, double scale, int angle)
590 int size_x_scaled = int(floor(size_x * scale));
591 int size_y_scaled = int(floor(size_y * scale));
593 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled, angle);
594 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled, angle);
595 if (m_visual_debug) {
596 cv::Mat patch_rgb_copy = patch_rgb.clone();
597 // Check 4 sectors of image if they have same number of black pixels
598 for(int sector = 0; sector < 4; sector++){
600 for (int row = (sector<2?0:1)*patch_rgb_copy.rows/2; row < (patch_rgb_copy.rows-1)/(sector<2?2:1); row++){
601 for (int col = (sector == 0 || sector == 2?0:1)*patch_rgb_copy.cols/2; col < (patch_rgb_copy.cols-1)/((sector == 0 || sector == 2?2:1)); col++){
602 cv::Vec3b pixel = patch_rgb_copy.at<cv::Vec3b>(row,col);
603 if (pixel.val[0] == 0 && pixel.val[1] == 0 && pixel.val[2] == 0)
607 std::cout << blackPixels << std::endl;
609 std::cout << std::endl;
611 cv::line(patch_rgb_copy, cv::Point(0, (patch_rgb_copy.cols-1)/2), cv::Point(patch_rgb_copy.rows-1, (patch_rgb_copy.cols-1)/2),cv::Scalar(0, 255, 0));
612 cv::line(patch_rgb_copy, cv::Point((patch_rgb_copy.rows-1)/2, 0), cv::Point((patch_rgb_copy.rows-1)/2, patch_rgb_copy.cols-1),cv::Scalar(0, 255, 0));
613 cv::imshow("Patch RGB unresized", patch_rgb_copy);
617 // resize to default size
619 // if we downsample use INTER_AREA interpolation
620 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
622 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
624 cv::Point2f center((patch_gray.cols-1)/2., (patch_gray.rows-1)/2.);
625 cv::Mat r = getRotationMatrix2D(center, angle, 1.0);
627 cv::warpAffine(patch_gray, patch_gray, r, cv::Size(patch_gray.cols, patch_gray.rows), cv::BORDER_CONSTANT, 1);
629 // get hog(Histogram of Oriented Gradients) features
630 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
632 // get color rgb features (simple r,g,b channels)
633 std::vector<cv::Mat> color_feat;
634 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
635 // resize to default size
637 // if we downsample use INTER_AREA interpolation
638 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
641 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
644 // cv::imshow("Test", patch_rgb);
646 cv::Point2f center((patch_rgb.cols-1)/2., (patch_rgb.rows-1)/2.);
647 cv::Mat r = getRotationMatrix2D(center, angle, 1.0);
649 cv::warpAffine(patch_rgb, patch_rgb, r, cv::Size(patch_rgb.cols, patch_rgb.rows), cv::BORDER_CONSTANT, 1);
650 cv::Mat patch_rgb_copy = patch_rgb.clone();
652 cv::namedWindow("Patch RGB copy", CV_WINDOW_NORMAL);
653 cv::resizeWindow("Patch RGB copy", 200, 200);
654 cv::putText(patch_rgb_copy, std::to_string(angle), cv::Point(0, patch_rgb_copy.rows-1), cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0,255,0),2,cv::LINE_AA);
655 cv::imshow("Patch RGB copy", patch_rgb_copy);
660 if (m_use_color && input_rgb.channels() == 3) {
661 // use rgb color space
662 cv::Mat patch_rgb_norm;
663 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
664 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
665 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
666 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
667 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
668 cv::split(patch_rgb_norm, rgb);
669 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
672 if (m_use_cnfeat && input_rgb.channels() == 3) {
673 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
674 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
676 BIG_BATCH_OMP_ORDERED
677 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
681 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
683 cv::Mat labels(dim2, dim1, CV_32FC1);
684 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
685 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
687 double sigma_s = sigma * sigma;
689 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
690 float *row_ptr = labels.ptr<float>(j);
692 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
693 row_ptr[i] = float(std::exp(-0.5 * (y_s + x * x) / sigma_s)); //-1/2*e^((y^2+x^2)/sigma^2)
697 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
699 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
700 tmp.copyTo(p_rot_labels);
702 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
705 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
706 // sanity check, 1 at top left corner
707 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
713 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
715 cv::Mat rot_patch(patch.size(), CV_32FC1);
716 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
718 // circular rotate x-axis
720 // move part that does not rotate over the edge
721 cv::Range orig_range(-x_rot, patch.cols);
722 cv::Range rot_range(0, patch.cols - (-x_rot));
723 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
726 orig_range = cv::Range(0, -x_rot);
727 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
728 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
729 } else if (x_rot > 0) {
730 // move part that does not rotate over the edge
731 cv::Range orig_range(0, patch.cols - x_rot);
732 cv::Range rot_range(x_rot, patch.cols);
733 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
736 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
737 rot_range = cv::Range(0, x_rot);
738 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
739 } else { // zero rotation
740 // move part that does not rotate over the edge
741 cv::Range orig_range(0, patch.cols);
742 cv::Range rot_range(0, patch.cols);
743 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
746 // circular rotate y-axis
748 // move part that does not rotate over the edge
749 cv::Range orig_range(-y_rot, patch.rows);
750 cv::Range rot_range(0, patch.rows - (-y_rot));
751 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
754 orig_range = cv::Range(0, -y_rot);
755 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
756 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
757 } else if (y_rot > 0) {
758 // move part that does not rotate over the edge
759 cv::Range orig_range(0, patch.rows - y_rot);
760 cv::Range rot_range(y_rot, patch.rows);
761 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
764 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
765 rot_range = cv::Range(0, y_rot);
766 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
767 } else { // zero rotation
768 // move part that does not rotate over the edge
769 cv::Range orig_range(0, patch.rows);
770 cv::Range rot_range(0, patch.rows);
771 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
777 // hann window actually (Power-of-cosine windows)
778 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
780 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
781 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
782 for (int i = 0; i < dim1; ++i)
783 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
784 N_inv = 1. / (static_cast<double>(dim2) - 1.);
785 for (int i = 0; i < dim2; ++i)
786 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
787 cv::Mat ret = m2 * m1;
791 // Returns sub-window of image input centered at [cx, cy] coordinates),
792 // with size [width, height]. If any pixels are outside of the image,
793 // they will replicate the values at the borders.
794 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height, int angle)
798 int x1 = cx - width/2;
799 int y1 = cy - height/2;
800 int x2 = cx + width/2;
801 int y2 = cy + height/2;
803 // std::cout << "Original coordinates x1: " << x1 << " y1: " << y1 << " x2: " << x2 << " y2: " << y2 << std::endl;
805 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
806 patch.create(height, width, input.type());
807 patch.setTo(double(0.f));
811 int top = 0, bottom = 0, left = 0, right = 0;
813 // fit to image coordinates, set border extensions;
822 if (x2 >= input.cols) {
823 right = x2 - input.cols + width % 2;
828 if (y2 >= input.rows) {
829 bottom = y2 - input.rows + height % 2;
834 // cv::Mat input_copy;
835 // cv::Point2f center(x2-x1, y2-y1);
836 // cv::Mat r = getRotationMatrix2D(center, angle, 1.0);
838 // cv::warpAffine(input, input_copy, r, cv::Size(input.cols, input.rows), cv::BORDER_CONSTANT, 1);
840 if (x2 - x1 == 0 || y2 - y1 == 0)
841 patch = cv::Mat::zeros(height, width, CV_32FC1);
843 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
844 cv::BORDER_REPLICATE);
845 // imshow( "copyMakeBorder", patch);
850 assert(patch.cols == width && patch.rows == height);
855 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
856 double sigma, bool auto_correlation)
859 xf.sqr_norm(vars.xf_sqr_norm.deviceMem());
860 if (!auto_correlation) yf.sqr_norm(vars.yf_sqr_norm.deviceMem());
862 xf.sqr_norm(vars.xf_sqr_norm.hostMem());
863 if (auto_correlation) {
864 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
866 yf.sqr_norm(vars.yf_sqr_norm.hostMem());
869 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
870 DEBUG_PRINTM(vars.xyf);
871 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
873 if (auto_correlation)
874 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(),
875 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
877 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.yf_sqr_norm.deviceMem(),
878 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
880 // ifft2 and sum over 3rd dimension, we dont care about individual channels
881 DEBUG_PRINTM(vars.ifft2_res);
883 if (xf.channels() != p_num_scales * p_num_of_feats)
884 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
886 xy_sum.create(vars.ifft2_res.size(), CV_32FC(int(p_scales.size())));
888 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
889 float *row_ptr = vars.ifft2_res.ptr<float>(y);
890 float *row_ptr_sum = xy_sum.ptr<float>(y);
891 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
892 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
893 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
894 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
895 (row_ptr + x * vars.ifft2_res.channels() +
896 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
901 DEBUG_PRINTM(xy_sum);
903 std::vector<cv::Mat> scales;
904 cv::split(xy_sum, scales);
906 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
907 for (uint i = 0; i < uint(xf.n_scales); ++i) {
908 cv::Mat in_roi(vars.in_all, cv::Rect(0, int(i) * scales[0].rows, scales[0].cols, scales[0].rows));
910 -1. / (sigma * sigma) *
911 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
913 DEBUG_PRINTM(in_roi);
916 DEBUG_PRINTM(vars.in_all);
917 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
922 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
926 if (x < 0) x = response.cols + x;
927 if (y < 0) y = response.rows + y;
928 if (x >= response.cols) x = x - response.cols;
929 if (y >= response.rows) y = y - response.rows;
931 return response.at<float>(y, x);
934 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
936 // find neighbourhood of max_loc (response is circular)
940 cv::Point2i p1(max_loc.x - 1, max_loc.y - 1), p2(max_loc.x, max_loc.y - 1), p3(max_loc.x + 1, max_loc.y - 1);
941 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
942 cv::Point2i p6(max_loc.x - 1, max_loc.y + 1), p7(max_loc.x, max_loc.y + 1), p8(max_loc.x + 1, max_loc.y + 1);
945 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
946 cv::Mat A = (cv::Mat_<float>(9, 6) <<
947 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
948 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
949 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
950 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
951 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
952 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
953 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
954 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
955 max_loc.x*max_loc.x, max_loc.x*max_loc.y, max_loc.y*max_loc.y, max_loc.x, max_loc.y, 1.f);
956 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
957 get_response_circular(p1, response),
958 get_response_circular(p2, response),
959 get_response_circular(p3, response),
960 get_response_circular(p4, response),
961 get_response_circular(p5, response),
962 get_response_circular(p6, response),
963 get_response_circular(p7, response),
964 get_response_circular(p8, response),
965 get_response_circular(max_loc, response));
968 cv::solve(A, fval, x, cv::DECOMP_SVD);
970 float a = x.at<float>(0), b = x.at<float>(1), c = x.at<float>(2), d = x.at<float>(3), e = x.at<float>(4);
972 cv::Point2f sub_peak(max_loc.x, max_loc.y);
973 if (b > 0 || b < 0) {
974 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
975 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
981 double KCF_Tracker::sub_grid_scale(int index)
984 if (index < 0 || index > int(p_scales.size()) - 1) {
985 // interpolate from all values
986 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
987 A.create(int(p_scales.size()), 3, CV_32FC1);
988 fval.create(int(p_scales.size()), 1, CV_32FC1);
989 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
990 uint i = uint(std::distance(p_threadctxs.begin(), it));
992 A.at<float>(j, 0) = float(p_scales[i] * p_scales[i]);
993 A.at<float>(j, 1) = float(p_scales[i]);
994 A.at<float>(j, 2) = 1;
996 m_use_big_batch ? float(p_threadctxs.back()->max_responses[i]) : float((*it)->max_response);
999 // only from neighbours
1000 if (index == 0 || index == int(p_scales.size()) - 1) return p_scales[uint(index)];
1002 A = (cv::Mat_<float>(3, 3) << p_scales[uint(index) - 1] * p_scales[uint(index) - 1], p_scales[uint(index) - 1],
1003 1, p_scales[uint(index)] * p_scales[uint(index)], p_scales[uint(index)], 1,
1004 p_scales[uint(index) + 1] * p_scales[uint(index) + 1], p_scales[uint(index) + 1], 1);
1005 auto it1 = p_threadctxs.begin();
1006 std::advance(it1, index - 1);
1007 auto it2 = p_threadctxs.begin();
1008 std::advance(it2, index);
1009 auto it3 = p_threadctxs.begin();
1010 std::advance(it3, index + 1);
1011 fval = (cv::Mat_<float>(3, 1) << (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index) - 1]
1012 : (*it1)->max_response),
1013 (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index)] : (*it2)->max_response),
1014 (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index) + 1] : (*it3)->max_response));
1018 cv::solve(A, fval, x, cv::DECOMP_SVD);
1019 float a = x.at<float>(0), b = x.at<float>(1);
1020 double scale = p_scales[uint(index)];
1021 if (a > 0 || a < 0) scale = double(-b / (2 * a));