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 if (p_windows_size.height / p_cell_size * (p_windows_size.width / p_cell_size / 2 + 1) > 1024) {
145 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
146 "the window dimensions so its size is less or equal to "
147 << 1024 * p_cell_size * p_cell_size * 2 + 1
148 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
149 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
150 std::exit(EXIT_FAILURE);
153 if (m_use_linearkernel) {
154 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
155 std::exit(EXIT_FAILURE);
157 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
158 p_rot_labels_data = DynMem(
159 ((uint(p_windows_size.width) / p_cell_size) * (uint(p_windows_size.height) / p_cell_size)) * sizeof(float));
160 p_rot_labels = cv::Mat(p_windows_size.height / int(p_cell_size), p_windows_size.width / int(p_cell_size), CV_32FC1,
161 p_rot_labels_data.hostMem());
163 p_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.height / p_cell_size)) / 2 + 1,
167 #if defined(CUFFT) || defined(FFTW)
168 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1,
169 uint(p_num_of_feats));
170 p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1, 1);
171 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size) / 2 + 1,
174 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)),
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)), 1);
177 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size), p_num_of_feats);
180 int max = m_use_big_batch ? 2 : p_num_scales;
181 for (int i = 0; i < max; ++i) {
182 if (m_use_big_batch && i == 1) {
183 p_threadctxs.emplace_back(
184 new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats * p_num_scales, p_num_scales));
186 p_threadctxs.emplace_back(new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats, 1));
190 p_current_scale = 1.;
192 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
193 double max_size_ratio =
194 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
195 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
196 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
197 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
199 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
200 std::cout << "init: win size. " << p_windows_size.width << " " << p_windows_size.height << std::endl;
201 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
203 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
205 fft.init(uint(p_windows_size.width / p_cell_size), uint(p_windows_size.height / p_cell_size), uint(p_num_of_feats),
206 uint(p_num_scales), m_use_big_batch);
207 fft.set_window(cosine_window_function(p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size));
209 // window weights, i.e. labels
211 gaussian_shaped_labels(p_output_sigma, p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size), p_yf,
212 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_threadctxs.front()->stream);
215 // obtain a sub-window for training initial model
216 p_threadctxs.front()->patch_feats.clear();
217 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
218 *p_threadctxs.front());
219 fft.forward_window(p_threadctxs.front()->patch_feats, p_model_xf, p_threadctxs.front()->fw_all,
220 m_use_cuda ? p_threadctxs.front()->data_features.deviceMem() : nullptr, p_threadctxs.front()->stream);
221 DEBUG_PRINTM(p_model_xf);
222 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
223 p_threadctxs.front()->model_xf = p_model_xf;
224 p_threadctxs.front()->model_xf.set_stream(p_threadctxs.front()->stream);
225 p_yf.set_stream(p_threadctxs.front()->stream);
226 p_model_xf.set_stream(p_threadctxs.front()->stream);
227 p_xf.set_stream(p_threadctxs.front()->stream);
230 if (m_use_linearkernel) {
231 ComplexMat xfconj = p_model_xf.conj();
232 p_model_alphaf_num = xfconj.mul(p_yf);
233 p_model_alphaf_den = (p_model_xf * xfconj);
235 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
236 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
237 gaussian_correlation(*p_threadctxs.front(), p_threadctxs.front()->model_xf, p_threadctxs.front()->model_xf,
238 p_kernel_sigma, true);
240 gaussian_correlation(*p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
242 DEBUG_PRINTM(p_threadctxs.front()->kf);
243 p_model_alphaf_num = p_yf * p_threadctxs.front()->kf;
244 DEBUG_PRINTM(p_model_alphaf_num);
245 p_model_alphaf_den = p_threadctxs.front()->kf * (p_threadctxs.front()->kf + float(p_lambda));
246 DEBUG_PRINTM(p_model_alphaf_den);
248 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
249 DEBUG_PRINTM(p_model_alphaf);
250 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
252 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
253 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
254 (*it)->model_xf = p_model_xf;
255 (*it)->model_xf.set_stream((*it)->stream);
256 (*it)->model_alphaf = p_model_alphaf;
257 (*it)->model_alphaf.set_stream((*it)->stream);
262 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
264 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
267 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
269 if (p_resize_image) {
271 tmp.scale(p_downscale_factor);
274 } else if (p_fit_to_pw2) {
276 tmp.scale_x(p_scale_factor_x);
277 tmp.scale_y(p_scale_factor_y);
286 BBox_c KCF_Tracker::getBBox()
289 tmp.w *= p_current_scale;
290 tmp.h *= p_current_scale;
292 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
294 tmp.scale_x(1 / p_scale_factor_x);
295 tmp.scale_y(1 / p_scale_factor_y);
301 void KCF_Tracker::track(cv::Mat &img)
303 if (m_debug) std::cout << "NEW FRAME" << '\n';
304 cv::Mat input_gray, input_rgb = img.clone();
305 if (img.channels() == 3) {
306 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
307 input_gray.convertTo(input_gray, CV_32FC1);
309 img.convertTo(input_gray, CV_32FC1);
311 // don't need too large image
312 if (p_resize_image) {
313 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
314 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
315 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
316 fabs(p_scale_factor_y - 1) > p_floating_error) {
317 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
318 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
319 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
321 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
322 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
326 double max_response = -1.;
328 cv::Point2i *max_response_pt = nullptr;
329 cv::Mat *max_response_map = nullptr;
331 if (m_use_multithreading) {
332 std::vector<std::future<void>> async_res(p_scales.size());
333 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
334 uint index = uint(std::distance(p_threadctxs.begin(), it));
335 async_res[index] = std::async(std::launch::async, [this, &input_gray, &input_rgb, index, it]() -> void {
336 return scale_track(*(*it), input_rgb, input_gray, this->p_scales[index]);
339 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
340 uint index = uint(std::distance(p_threadctxs.begin(), it));
341 async_res[index].wait();
342 if ((*it)->max_response > max_response) {
343 max_response = (*it)->max_response;
344 max_response_pt = &(*it)->max_loc;
345 max_response_map = &(*it)->response;
346 scale_index = int(index);
350 uint start = m_use_big_batch ? 1 : 0;
351 uint end = m_use_big_batch ? 2 : uint(p_num_scales);
352 NORMAL_OMP_PARALLEL_FOR
353 for (uint i = start; i < end; ++i) {
354 auto it = p_threadctxs.begin();
356 scale_track(*(*it), input_rgb, input_gray, this->p_scales[i]);
358 if (m_use_big_batch) {
359 for (size_t j = 0; j < p_scales.size(); ++j) {
360 if ((*it)->max_responses[j] > max_response) {
361 max_response = (*it)->max_responses[j];
362 max_response_pt = &(*it)->max_locs[j];
363 max_response_map = &(*it)->response_maps[j];
364 scale_index = int(j);
370 if ((*it)->max_response > max_response) {
371 max_response = (*it)->max_response;
372 max_response_pt = &(*it)->max_loc;
373 max_response_map = &(*it)->response;
374 scale_index = int(i);
381 DEBUG_PRINTM(*max_response_map);
382 DEBUG_PRINT(*max_response_pt);
384 // sub pixel quadratic interpolation from neighbours
385 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
386 max_response_pt->y = max_response_pt->y - max_response_map->rows;
387 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
388 max_response_pt->x = max_response_pt->x - max_response_map->cols;
390 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
391 DEBUG_PRINT(new_location);
393 if (m_use_subpixel_localization) new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
394 DEBUG_PRINT(new_location);
396 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
397 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
399 if (p_pose.cx < 0) p_pose.cx = 0;
400 if (p_pose.cx > (img.cols * p_scale_factor_x) - 1) p_pose.cx = (img.cols * p_scale_factor_x) - 1;
401 if (p_pose.cy < 0) p_pose.cy = 0;
402 if (p_pose.cy > (img.rows * p_scale_factor_y) - 1) p_pose.cy = (img.rows * p_scale_factor_y) - 1;
404 if (p_pose.cx < 0) p_pose.cx = 0;
405 if (p_pose.cx > img.cols - 1) p_pose.cx = img.cols - 1;
406 if (p_pose.cy < 0) p_pose.cy = 0;
407 if (p_pose.cy > img.rows - 1) p_pose.cy = img.rows - 1;
410 // sub grid scale interpolation
411 double new_scale = p_scales[uint(scale_index)];
412 if (m_use_subgrid_scale) new_scale = sub_grid_scale(scale_index);
414 p_current_scale *= new_scale;
416 if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
417 if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
419 // obtain a subwindow for training at newly estimated target position
420 p_threadctxs.front()->patch_feats.clear();
421 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
422 *p_threadctxs.front(), p_current_scale);
423 fft.forward_window(p_threadctxs.front()->patch_feats, p_xf, p_threadctxs.front()->fw_all,
424 m_use_cuda ? p_threadctxs.front()->data_features.deviceMem() : nullptr, p_threadctxs.front()->stream);
426 // subsequent frames, interpolate model
427 p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
429 ComplexMat alphaf_num, alphaf_den;
431 if (m_use_linearkernel) {
432 ComplexMat xfconj = p_xf.conj();
433 alphaf_num = xfconj.mul(p_yf);
434 alphaf_den = (p_xf * xfconj);
436 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
437 gaussian_correlation(*p_threadctxs.front(), p_xf, p_xf, p_kernel_sigma,
439 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
440 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
441 alphaf_num = p_yf * p_threadctxs.front()->kf;
442 alphaf_den = p_threadctxs.front()->kf * (p_threadctxs.front()->kf + float(p_lambda));
445 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
446 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
447 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
449 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
450 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
451 (*it)->model_xf = p_model_xf;
452 (*it)->model_xf.set_stream((*it)->stream);
453 (*it)->model_alphaf = p_model_alphaf;
454 (*it)->model_alphaf.set_stream((*it)->stream);
459 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray, double scale)
461 if (m_use_big_batch) {
462 vars.patch_feats.clear();
463 BIG_BATCH_OMP_PARALLEL_FOR
464 for (uint i = 0; i < uint(p_num_scales); ++i) {
465 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
466 this->p_windows_size.height, vars, this->p_current_scale * this->p_scales[i]);
469 vars.patch_feats.clear();
470 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
471 this->p_windows_size.height, vars, this->p_current_scale *scale);
474 fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
476 DEBUG_PRINTM(vars.zf);
478 if (m_use_linearkernel) {
479 vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
480 : (p_model_alphaf * vars.zf).sum_over_channels();
481 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
483 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
484 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
485 vars.kzf = vars.model_alphaf * vars.kzf;
487 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
488 DEBUG_PRINTM(this->p_model_alphaf);
489 DEBUG_PRINTM(vars.kzf);
490 vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
492 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
495 DEBUG_PRINTM(vars.response);
497 /* target location is at the maximum response. we must take into
498 account the fact that, if the target doesn't move, the peak
499 will appear at the top-left corner, not at the center (this is
500 discussed in the paper). the responses wrap around cyclically. */
501 if (m_use_big_batch) {
502 cv::split(vars.response, vars.response_maps);
504 for (size_t i = 0; i < p_scales.size(); ++i) {
505 double min_val, max_val;
506 cv::Point2i min_loc, max_loc;
507 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
508 DEBUG_PRINT(max_loc);
509 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
510 vars.max_responses[i] = max_val * weight;
511 vars.max_locs[i] = max_loc;
516 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
518 DEBUG_PRINT(vars.max_loc);
520 double weight = scale < 1. ? scale : 1. / scale;
521 vars.max_response = vars.max_val * weight;
526 // ****************************************************************************
528 void KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y,
529 ThreadCtx &vars, double scale)
531 int size_x_scaled = int(floor(size_x * scale));
532 int size_y_scaled = int(floor(size_y * scale));
534 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
535 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
537 // resize to default size
539 // if we downsample use INTER_AREA interpolation
540 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
542 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
545 // get hog(Histogram of Oriented Gradients) features
546 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
548 // get color rgb features (simple r,g,b channels)
549 std::vector<cv::Mat> color_feat;
550 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
551 // resize to default size
553 // if we downsample use INTER_AREA interpolation
554 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
557 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
562 if (m_use_color && input_rgb.channels() == 3) {
563 // use rgb color space
564 cv::Mat patch_rgb_norm;
565 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
566 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
567 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
568 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
569 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
570 cv::split(patch_rgb_norm, rgb);
571 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
574 if (m_use_cnfeat && input_rgb.channels() == 3) {
575 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
576 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
578 BIG_BATCH_OMP_ORDERED
579 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
583 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
585 cv::Mat labels(dim2, dim1, CV_32FC1);
586 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
587 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
589 double sigma_s = sigma * sigma;
591 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
592 float *row_ptr = labels.ptr<float>(j);
594 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
595 row_ptr[i] = float(std::exp(-0.5 * (y_s + x * x) / sigma_s)); //-1/2*e^((y^2+x^2)/sigma^2)
599 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
601 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
602 tmp.copyTo(p_rot_labels);
604 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
607 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
608 // sanity check, 1 at top left corner
609 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
615 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
617 cv::Mat rot_patch(patch.size(), CV_32FC1);
618 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
620 // circular rotate x-axis
622 // move part that does not rotate over the edge
623 cv::Range orig_range(-x_rot, patch.cols);
624 cv::Range rot_range(0, patch.cols - (-x_rot));
625 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
628 orig_range = cv::Range(0, -x_rot);
629 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
630 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
631 } else if (x_rot > 0) {
632 // move part that does not rotate over the edge
633 cv::Range orig_range(0, patch.cols - x_rot);
634 cv::Range rot_range(x_rot, patch.cols);
635 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
638 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
639 rot_range = cv::Range(0, x_rot);
640 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
641 } else { // zero rotation
642 // move part that does not rotate over the edge
643 cv::Range orig_range(0, patch.cols);
644 cv::Range rot_range(0, patch.cols);
645 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
648 // circular rotate y-axis
650 // move part that does not rotate over the edge
651 cv::Range orig_range(-y_rot, patch.rows);
652 cv::Range rot_range(0, patch.rows - (-y_rot));
653 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
656 orig_range = cv::Range(0, -y_rot);
657 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
658 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
659 } else if (y_rot > 0) {
660 // move part that does not rotate over the edge
661 cv::Range orig_range(0, patch.rows - y_rot);
662 cv::Range rot_range(y_rot, patch.rows);
663 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
666 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
667 rot_range = cv::Range(0, y_rot);
668 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
669 } else { // zero rotation
670 // move part that does not rotate over the edge
671 cv::Range orig_range(0, patch.rows);
672 cv::Range rot_range(0, patch.rows);
673 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
679 // hann window actually (Power-of-cosine windows)
680 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
682 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
683 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
684 for (int i = 0; i < dim1; ++i)
685 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
686 N_inv = 1. / (static_cast<double>(dim2) - 1.);
687 for (int i = 0; i < dim2; ++i)
688 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
689 cv::Mat ret = m2 * m1;
693 // Returns sub-window of image input centered at [cx, cy] coordinates),
694 // with size [width, height]. If any pixels are outside of the image,
695 // they will replicate the values at the borders.
696 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
700 int x1 = cx - width / 2;
701 int y1 = cy - height / 2;
702 int x2 = cx + width / 2;
703 int y2 = cy + height / 2;
706 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
707 patch.create(height, width, input.type());
708 patch.setTo(double(0.f));
712 int top = 0, bottom = 0, left = 0, right = 0;
714 // fit to image coordinates, set border extensions;
723 if (x2 >= input.cols) {
724 right = x2 - input.cols + width % 2;
729 if (y2 >= input.rows) {
730 bottom = y2 - input.rows + height % 2;
735 if (x2 - x1 == 0 || y2 - y1 == 0)
736 patch = cv::Mat::zeros(height, width, CV_32FC1);
738 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
739 cv::BORDER_REPLICATE);
740 // imshow( "copyMakeBorder", patch);
745 assert(patch.cols == width && patch.rows == height);
750 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
751 double sigma, bool auto_correlation)
754 xf.sqr_norm(vars.xf_sqr_norm.deviceMem());
755 if (!auto_correlation) yf.sqr_norm(vars.yf_sqr_norm.deviceMem());
757 xf.sqr_norm(vars.xf_sqr_norm.hostMem());
758 if (auto_correlation) {
759 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
761 yf.sqr_norm(vars.yf_sqr_norm.hostMem());
764 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
765 DEBUG_PRINTM(vars.xyf);
766 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
768 if (auto_correlation)
769 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(),
770 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
772 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.yf_sqr_norm.deviceMem(),
773 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
775 // ifft2 and sum over 3rd dimension, we dont care about individual channels
776 DEBUG_PRINTM(vars.ifft2_res);
778 if (xf.channels() != p_num_scales * p_num_of_feats)
779 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
781 xy_sum.create(vars.ifft2_res.size(), CV_32FC(int(p_scales.size())));
783 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
784 float *row_ptr = vars.ifft2_res.ptr<float>(y);
785 float *row_ptr_sum = xy_sum.ptr<float>(y);
786 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
787 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
788 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
789 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
790 (row_ptr + x * vars.ifft2_res.channels() +
791 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
796 DEBUG_PRINTM(xy_sum);
798 std::vector<cv::Mat> scales;
799 cv::split(xy_sum, scales);
801 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
802 for (uint i = 0; i < uint(xf.n_scales); ++i) {
803 cv::Mat in_roi(vars.in_all, cv::Rect(0, int(i) * scales[0].rows, scales[0].cols, scales[0].rows));
805 -1. / (sigma * sigma) *
806 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
808 DEBUG_PRINTM(in_roi);
811 DEBUG_PRINTM(vars.in_all);
812 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
817 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
821 if (x < 0) x = response.cols + x;
822 if (y < 0) y = response.rows + y;
823 if (x >= response.cols) x = x - response.cols;
824 if (y >= response.rows) y = y - response.rows;
826 return response.at<float>(y, x);
829 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
831 // find neighbourhood of max_loc (response is circular)
835 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);
836 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
837 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);
840 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
841 cv::Mat A = (cv::Mat_<float>(9, 6) <<
842 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
843 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
844 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
845 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
846 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
847 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
848 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
849 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
850 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);
851 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
852 get_response_circular(p1, response),
853 get_response_circular(p2, response),
854 get_response_circular(p3, response),
855 get_response_circular(p4, response),
856 get_response_circular(p5, response),
857 get_response_circular(p6, response),
858 get_response_circular(p7, response),
859 get_response_circular(p8, response),
860 get_response_circular(max_loc, response));
863 cv::solve(A, fval, x, cv::DECOMP_SVD);
865 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);
867 cv::Point2f sub_peak(max_loc.x, max_loc.y);
868 if (b > 0 || b < 0) {
869 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
870 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
876 double KCF_Tracker::sub_grid_scale(int index)
879 if (index < 0 || index > int(p_scales.size()) - 1) {
880 // interpolate from all values
881 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
882 A.create(int(p_scales.size()), 3, CV_32FC1);
883 fval.create(int(p_scales.size()), 1, CV_32FC1);
884 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
885 uint i = uint(std::distance(p_threadctxs.begin(), it));
887 A.at<float>(j, 0) = float(p_scales[i] * p_scales[i]);
888 A.at<float>(j, 1) = float(p_scales[i]);
889 A.at<float>(j, 2) = 1;
891 m_use_big_batch ? float(p_threadctxs.back()->max_responses[i]) : float((*it)->max_response);
894 // only from neighbours
895 if (index == 0 || index == int(p_scales.size()) - 1) return p_scales[uint(index)];
897 A = (cv::Mat_<float>(3, 3) << p_scales[uint(index) - 1] * p_scales[uint(index) - 1], p_scales[uint(index) - 1],
898 1, p_scales[uint(index)] * p_scales[uint(index)], p_scales[uint(index)], 1,
899 p_scales[uint(index) + 1] * p_scales[uint(index) + 1], p_scales[uint(index) + 1], 1);
900 auto it1 = p_threadctxs.begin();
901 std::advance(it1, index - 1);
902 auto it2 = p_threadctxs.begin();
903 std::advance(it2, index);
904 auto it3 = p_threadctxs.begin();
905 std::advance(it3, index + 1);
906 fval = (cv::Mat_<float>(3, 1) << (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index) - 1]
907 : (*it1)->max_response),
908 (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index)] : (*it2)->max_response),
909 (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index) + 1] : (*it3)->max_response));
913 cv::solve(A, fval, x, cv::DECOMP_SVD);
914 float a = x.at<float>(0), b = x.at<float>(1);
915 double scale = p_scales[uint(index)];
916 if (a > 0 || a < 0) scale = double(-b / (2 * a));