11 #include "fft_cufft.h"
14 #include "fft_opencv.h"
22 #define DEBUG_PRINT(obj) if (m_debug) {std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl;}
23 #define DEBUG_PRINTM(obj) if (m_debug) {std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl << (obj) << std::endl;}
24 #define DEBUG_PRINTD(obj) {std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl << (obj) << std::endl;}
27 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor, double output_sigma_factor, int cell_size) :
29 p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
30 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size) {}
32 KCF_Tracker::KCF_Tracker()
35 KCF_Tracker::~KCF_Tracker()
40 void KCF_Tracker::init(cv::Mat &img, const cv::Rect & bbox, int fit_size_x, int fit_size_y)
42 //check boundary, enforce min size
43 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
45 if (x2 > img.cols-1) x2 = img.cols - 1;
47 if (y2 > img.rows-1) y2 = img.rows - 1;
49 if (x2-x1 < 2*p_cell_size) {
50 double diff = (2*p_cell_size -x2+x1)/2.;
51 if (x1 - diff >= 0 && x2 + diff < img.cols){
54 } else if (x1 - 2*diff >= 0) {
60 if (y2-y1 < 2*p_cell_size) {
61 double diff = (2*p_cell_size -y2+y1)/2.;
62 if (y1 - diff >= 0 && y2 + diff < img.rows){
65 } else if (y1 - 2*diff >= 0) {
74 p_pose.cx = x1 + p_pose.w/2.;
75 p_pose.cy = y1 + p_pose.h /2.;
78 cv::Mat input_gray, input_rgb = img.clone();
79 if (img.channels() == 3){
80 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
81 input_gray.convertTo(input_gray, CV_32FC1);
83 img.convertTo(input_gray, CV_32FC1);
85 // don't need too large image
86 if (p_pose.w * p_pose.h > 100.*100. && (fit_size_x == -1 || fit_size_y == -1)) {
87 std::cout << "resizing image by factor of " << 1/p_downscale_factor << std::endl;
88 p_resize_image = true;
89 p_pose.scale(p_downscale_factor);
90 cv::resize(input_gray, input_gray, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
91 cv::resize(input_rgb, input_rgb, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
92 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
93 if (fit_size_x%p_cell_size != 0 || fit_size_y%p_cell_size != 0) {
94 std::cerr << "Fit size does not fit to hog cell size. The dimensions have to be divisible by HOG cell size, which is: " << p_cell_size << std::endl;;
95 std::exit(EXIT_FAILURE);
97 double tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size ;
98 if (fabs(tmp-fit_size_x) > p_floating_error)
99 p_scale_factor_x = fit_size_x/tmp;
100 tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
101 if (fabs(tmp-fit_size_y) > p_floating_error)
102 p_scale_factor_y = fit_size_y/tmp;
103 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x
104 << " and verticaly by factor of " << p_scale_factor_y << std::endl;
106 p_pose.scale_x(p_scale_factor_x);
107 p_pose.scale_y(p_scale_factor_y);
108 if (fabs(p_scale_factor_x-1) > p_floating_error && fabs(p_scale_factor_y-1) > p_floating_error) {
109 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
110 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
111 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
113 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
114 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
119 //compute win size + fit to fhog cell size
120 p_windows_size[0] = int(round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size);
121 p_windows_size[1] = int(round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size);
125 for (int i = -p_num_scales/2; i <= p_num_scales/2; ++i)
126 p_scales.push_back(std::pow(p_scale_step, i));
128 p_scales.push_back(1.);
131 if (p_windows_size[1]/p_cell_size*(p_windows_size[0]/p_cell_size/2+1) > 1024) {
132 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
133 "the window dimensions so its size is less or equal to " << 1024*p_cell_size*p_cell_size*2+1 <<
134 " pixels . Currently the size of the window is: " << p_windows_size[0] << "x" << p_windows_size[1] <<
135 " which is " << p_windows_size[0]*p_windows_size[1] << " pixels. " << std::endl;
136 std::exit(EXIT_FAILURE);
139 if (m_use_linearkernel){
140 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
141 std::exit(EXIT_FAILURE);
146 if(m_use_color) p_num_of_feats += 3;
147 if(m_use_cnfeat) p_num_of_feats += 10;
148 p_roi_width = p_windows_size[0]/p_cell_size;
149 p_roi_height = p_windows_size[1]/p_cell_size;
151 int max =m_use_big_batch ? 2: p_num_scales;
152 for (int i = 0;i<max;++i) {
154 p_scale_vars.emplace_back(new Scale_vars(p_windows_size, p_cell_size, p_num_of_feats, 1, &p_model_xf, &p_yf, true));
156 else if (m_use_big_batch) {
157 p_scale_vars.emplace_back(new Scale_vars(p_windows_size, p_cell_size, p_num_of_feats*p_num_scales, p_num_scales));
160 p_scale_vars.emplace_back(new Scale_vars(p_windows_size, p_cell_size, p_num_of_feats, 1));
164 p_current_scale = 1.;
166 double min_size_ratio = std::max(5.*p_cell_size/p_windows_size[0], 5.*p_cell_size/p_windows_size[1]);
167 double max_size_ratio = std::min(floor((img.cols + p_windows_size[0]/3)/p_cell_size)*p_cell_size/p_windows_size[0], floor((img.rows + p_windows_size[1]/3)/p_cell_size)*p_cell_size/p_windows_size[1]);
168 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
169 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
171 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
172 std::cout << "init: win size. " << p_windows_size[0] << " " << p_windows_size[1] << std::endl;
173 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
175 p_output_sigma = std::sqrt(p_pose.w*p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
177 fft.init(uint(p_windows_size[0]/p_cell_size), uint(p_windows_size[1]/p_cell_size), uint(p_num_of_feats), uint(p_num_scales), m_use_big_batch);
178 fft.set_window(cosine_window_function(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
180 //window weights, i.e. labels
181 fft.forward(gaussian_shaped_labels(p_output_sigma, p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size), p_yf,
182 m_use_cuda ? p_scale_vars.front()->rot_labels_data_d: nullptr, p_scale_vars.front()->stream);
185 //obtain a sub-window for training initial model
186 p_scale_vars.front()->patch_feats.clear();
187 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size[0], p_windows_size[1], *p_scale_vars.front());
188 fft.forward_window(p_scale_vars.front()->patch_feats, p_model_xf, p_scale_vars.front()->fw_all,
189 m_use_cuda ? p_scale_vars.front()->data_features_d : nullptr, p_scale_vars.front()->stream);
190 DEBUG_PRINTM(p_model_xf);
191 #if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
192 p_scale_vars.front()->model_xf = p_model_xf;
193 p_scale_vars.front()->model_xf.set_stream(p_scale_vars.front()->stream);
194 p_yf.set_stream(p_scale_vars.front()->stream);
195 p_model_xf.set_stream(p_scale_vars.front()->stream);
198 if (m_use_linearkernel) {
199 ComplexMat xfconj = p_model_xf.conj();
200 p_model_alphaf_num = xfconj.mul(p_yf);
201 p_model_alphaf_den = (p_model_xf * xfconj);
203 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
204 #if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
205 gaussian_correlation(*p_scale_vars.front(), p_scale_vars.front()->model_xf, p_scale_vars.front()->model_xf, p_kernel_sigma, true);
207 gaussian_correlation(*p_scale_vars.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
209 DEBUG_PRINTM(p_scale_vars.front()->kf);
210 p_model_alphaf_num = p_yf * p_scale_vars.front()->kf;
211 DEBUG_PRINTM(p_model_alphaf_num);
212 p_model_alphaf_den = p_scale_vars.front()->kf * (p_scale_vars.front()->kf + float(p_lambda));
213 DEBUG_PRINTM(p_model_alphaf_den);
215 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
216 DEBUG_PRINTM(p_model_alphaf);
217 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
219 #if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
220 for (auto it = p_scale_vars.begin();it != p_scale_vars.end();++it) {
221 (*it)->model_xf = p_model_xf;
222 (*it)->model_xf.set_stream((*it)->stream);
223 (*it)->model_alphaf = p_model_alphaf;
224 (*it)->model_alphaf.set_stream((*it)->stream);
229 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat & img, int fit_size_x, int fit_size_y)
231 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
234 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
236 if (p_resize_image) {
238 tmp.scale(p_downscale_factor);
241 } else if (p_fit_to_pw2) {
243 tmp.scale_x(p_scale_factor_x);
244 tmp.scale_y(p_scale_factor_y);
253 BBox_c KCF_Tracker::getBBox()
256 tmp.w *= p_current_scale;
257 tmp.h *= p_current_scale;
260 tmp.scale(1/p_downscale_factor);
262 tmp.scale_x(1/p_scale_factor_x);
263 tmp.scale_y(1/p_scale_factor_y);
269 void KCF_Tracker::track(cv::Mat &img)
272 std::cout << "NEW FRAME" << '\n';
273 cv::Mat input_gray, input_rgb = img.clone();
274 if (img.channels() == 3){
275 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
276 input_gray.convertTo(input_gray, CV_32FC1);
278 img.convertTo(input_gray, CV_32FC1);
280 // don't need too large image
281 if (p_resize_image) {
282 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
283 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
284 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x-1) > p_floating_error && fabs(p_scale_factor_y-1) > p_floating_error) {
285 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
286 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
287 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
289 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
290 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
294 double max_response = -1.;
296 cv::Point2i *max_response_pt = nullptr;
297 cv::Mat *max_response_map = nullptr;
299 if(m_use_multithreading) {
300 std::vector<std::future<void>> async_res(p_scales.size());
301 for (auto it = p_scale_vars.begin();it != p_scale_vars.end();++it) {
302 uint index = uint(std::distance(p_scale_vars.begin(), it));
303 async_res[index] = std::async(std::launch::async,
304 [this, &input_gray, &input_rgb, index, it]() -> void
305 {return scale_track(*(*it), input_rgb, input_gray, this->p_scales[index]);});
307 for (auto it = p_scale_vars.begin();it != p_scale_vars.end();++it) {
308 uint index = uint(std::distance(p_scale_vars.begin(), it));
309 async_res[index].wait();
310 if ((*it)->max_response > max_response) {
311 max_response = (*it)->max_response;
312 max_response_pt = & (*it)->max_loc;
313 max_response_map = & (*it)->response;
314 scale_index = int(index);
318 uint start = m_use_big_batch ? 1 : 0;
319 uint end = m_use_big_batch ? 2 : uint(p_num_scales);
320 #pragma omp parallel for schedule(dynamic)
321 for (uint i = start; i < end; ++i) {
322 auto it = p_scale_vars.begin();
324 scale_track(*(*it), input_rgb, input_gray, this->p_scales[i]);
326 if (m_use_big_batch) {
327 for (size_t j = 0;j<p_scales.size();++j) {
328 if ((*it)->max_responses[j] > max_response) {
329 max_response = (*it)->max_responses[j];
330 max_response_pt = & (*it)->max_locs[j];
331 max_response_map = & (*it)->response_maps[j];
332 scale_index = int(j);
338 if ((*it)->max_response > max_response) {
339 max_response = (*it)->max_response;
340 max_response_pt = & (*it)->max_loc;
341 max_response_map = & (*it)->response;
349 DEBUG_PRINTM(*max_response_map);
350 DEBUG_PRINT(*max_response_pt);
352 //sub pixel quadratic interpolation from neighbours
353 if (max_response_pt->y > max_response_map->rows / 2) //wrap around to negative half-space of vertical axis
354 max_response_pt->y = max_response_pt->y - max_response_map->rows;
355 if (max_response_pt->x > max_response_map->cols / 2) //same for horizontal axis
356 max_response_pt->x = max_response_pt->x - max_response_map->cols;
358 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
359 DEBUG_PRINT(new_location);
361 if (m_use_subpixel_localization)
362 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
363 DEBUG_PRINT(new_location);
365 p_pose.cx += p_current_scale*p_cell_size*double(new_location.x);
366 p_pose.cy += p_current_scale*p_cell_size*double(new_location.y);
368 if (p_pose.cx < 0) p_pose.cx = 0;
369 if (p_pose.cx > (img.cols*p_scale_factor_x)-1) p_pose.cx = (img.cols*p_scale_factor_x)-1;
370 if (p_pose.cy < 0) p_pose.cy = 0;
371 if (p_pose.cy > (img.rows*p_scale_factor_y)-1) p_pose.cy = (img.rows*p_scale_factor_y)-1;
373 if (p_pose.cx < 0) p_pose.cx = 0;
374 if (p_pose.cx > img.cols-1) p_pose.cx = img.cols-1;
375 if (p_pose.cy < 0) p_pose.cy = 0;
376 if (p_pose.cy > img.rows-1) p_pose.cy = img.rows-1;
379 //sub grid scale interpolation
380 double new_scale = p_scales[uint(scale_index)];
381 if (m_use_subgrid_scale)
382 new_scale = sub_grid_scale(scale_index);
384 p_current_scale *= new_scale;
386 if (p_current_scale < p_min_max_scale[0])
387 p_current_scale = p_min_max_scale[0];
388 if (p_current_scale > p_min_max_scale[1])
389 p_current_scale = p_min_max_scale[1];
391 //obtain a subwindow for training at newly estimated target position
392 p_scale_vars.front()->patch_feats.clear();
393 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size[0], p_windows_size[1], *p_scale_vars.front(), p_current_scale);
394 fft.forward_window(p_scale_vars.front()->patch_feats, p_scale_vars.front()->xf, p_scale_vars.front()->fw_all,
395 m_use_cuda ? p_scale_vars.front()->data_features_d : nullptr, p_scale_vars.front()->stream);
397 //subsequent frames, interpolate model
398 p_model_xf = p_model_xf *float((1. - p_interp_factor)) + p_scale_vars.front()->xf * float(p_interp_factor);
400 ComplexMat alphaf_num, alphaf_den;
402 if (m_use_linearkernel) {
403 ComplexMat xfconj = p_scale_vars.front()->xf.conj();
404 alphaf_num = xfconj.mul(p_yf);
405 alphaf_den = (p_scale_vars.front()->xf * xfconj);
407 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
408 gaussian_correlation(*p_scale_vars.front(), p_scale_vars.front()->xf, p_scale_vars.front()->xf, p_kernel_sigma, true);
409 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
410 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
411 alphaf_num = p_yf * p_scale_vars.front()->kf;
412 alphaf_den = p_scale_vars.front()->kf * (p_scale_vars.front()->kf + float(p_lambda));
415 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
416 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
417 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
419 #if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
420 for (auto it = p_scale_vars.begin(); it != p_scale_vars.end(); ++it) {
421 (*it)->model_xf = p_model_xf;
422 (*it)->model_xf.set_stream((*it)->stream);
423 (*it)->model_alphaf = p_model_alphaf;
424 (*it)->model_alphaf.set_stream((*it)->stream);
429 void KCF_Tracker::scale_track(Scale_vars & vars, cv::Mat & input_rgb, cv::Mat & input_gray, double scale)
431 if (m_use_big_batch) {
432 vars.patch_feats.clear();
433 for (uint i = 0; i < uint(p_num_scales); ++i) {
434 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size[0], this->p_windows_size[1],
435 vars, this->p_current_scale * this->p_scales[i]);
438 vars.patch_feats.clear();
439 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size[0], this->p_windows_size[1],
440 vars, this->p_current_scale * scale);
443 fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features_d : nullptr, vars.stream);
444 DEBUG_PRINTM(vars.zf);
446 if (m_use_linearkernel) {
447 vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels() : (p_model_alphaf * vars.zf).sum_over_channels();
448 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch_d : nullptr, vars.stream);
451 #if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
452 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
453 vars.kzf = vars.model_alphaf * vars.kzf;
455 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
456 DEBUG_PRINTM(this->p_model_alphaf);
457 DEBUG_PRINTM(vars.kzf);
458 vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
460 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch_d : nullptr, vars.stream);
463 DEBUG_PRINTM(vars.response);
465 /* target location is at the maximum response. we must take into
466 account the fact that, if the target doesn't move, the peak
467 will appear at the top-left corner, not at the center (this is
468 discussed in the paper). the responses wrap around cyclically. */
469 if (m_use_big_batch) {
470 cv::split(vars.response,vars.response_maps);
472 for (size_t i = 0; i < p_scales.size(); ++i) {
473 double min_val, max_val;
474 cv::Point2i min_loc, max_loc;
475 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
476 DEBUG_PRINT(max_loc);
477 double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
478 vars.max_responses[i] = max_val*weight;
479 vars.max_locs[i] = max_loc;
484 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
486 DEBUG_PRINT(vars.max_loc);
488 double weight = scale < 1. ? scale : 1./scale;
489 vars.max_response = vars.max_val*weight;
494 // ****************************************************************************
496 void KCF_Tracker::get_features(cv::Mat & input_rgb, cv::Mat & input_gray, int cx, int cy, int size_x, int size_y, Scale_vars &vars, double scale)
498 int size_x_scaled = int(floor(size_x*scale));
499 int size_y_scaled = int(floor(size_y*scale));
501 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
502 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
504 //resize to default size
506 //if we downsample use INTER_AREA interpolation
507 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
509 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
512 // get hog(Histogram of Oriented Gradients) features
513 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
515 //get color rgb features (simple r,g,b channels)
516 std::vector<cv::Mat> color_feat;
517 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
518 //resize to default size
520 //if we downsample use INTER_AREA interpolation
521 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_AREA);
523 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_LINEAR);
527 if (m_use_color && input_rgb.channels() == 3) {
528 //use rgb color space
529 cv::Mat patch_rgb_norm;
530 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
531 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
532 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
533 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
534 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
535 cv::split(patch_rgb_norm, rgb);
536 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
539 if (m_use_cnfeat && input_rgb.channels() == 3) {
540 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
541 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
543 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
547 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
549 cv::Mat labels(dim2, dim1, CV_32FC1);
550 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
551 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
553 double sigma_s = sigma*sigma;
555 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j){
556 float * row_ptr = labels.ptr<float>(j);
558 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i){
559 row_ptr[i] = float(std::exp(-0.5 * (y_s + x*x) / sigma_s));//-1/2*e^((y^2+x^2)/sigma^2)
563 //rotate so that 1 is at top-left corner (see KCF paper for explanation)
565 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
566 tmp.copyTo(p_scale_vars.front()->rot_labels);
568 assert(p_scale_vars[0].rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
571 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
572 //sanity check, 1 at top left corner
573 assert(rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
579 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
581 cv::Mat rot_patch(patch.size(), CV_32FC1);
582 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
584 //circular rotate x-axis
586 //move part that does not rotate over the edge
587 cv::Range orig_range(-x_rot, patch.cols);
588 cv::Range rot_range(0, patch.cols - (-x_rot));
589 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
592 orig_range = cv::Range(0, -x_rot);
593 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
594 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
595 }else if (x_rot > 0){
596 //move part that does not rotate over the edge
597 cv::Range orig_range(0, patch.cols - x_rot);
598 cv::Range rot_range(x_rot, patch.cols);
599 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
602 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
603 rot_range = cv::Range(0, x_rot);
604 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
605 }else { //zero rotation
606 //move part that does not rotate over the edge
607 cv::Range orig_range(0, patch.cols);
608 cv::Range rot_range(0, patch.cols);
609 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
612 //circular rotate y-axis
614 //move part that does not rotate over the edge
615 cv::Range orig_range(-y_rot, patch.rows);
616 cv::Range rot_range(0, patch.rows - (-y_rot));
617 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
620 orig_range = cv::Range(0, -y_rot);
621 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
622 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
623 }else if (y_rot > 0){
624 //move part that does not rotate over the edge
625 cv::Range orig_range(0, patch.rows - y_rot);
626 cv::Range rot_range(y_rot, patch.rows);
627 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
630 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
631 rot_range = cv::Range(0, y_rot);
632 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
633 }else { //zero rotation
634 //move part that does not rotate over the edge
635 cv::Range orig_range(0, patch.rows);
636 cv::Range rot_range(0, patch.rows);
637 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
643 //hann window actually (Power-of-cosine windows)
644 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
646 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
647 double N_inv = 1./(static_cast<double>(dim1)-1.);
648 for (int i = 0; i < dim1; ++i)
649 m1.at<float>(i) = float(0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
650 N_inv = 1./(static_cast<double>(dim2)-1.);
651 for (int i = 0; i < dim2; ++i)
652 m2.at<float>(i) = float(0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
657 // Returns sub-window of image input centered at [cx, cy] coordinates),
658 // with size [width, height]. If any pixels are outside of the image,
659 // they will replicate the values at the borders.
660 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat & input, int cx, int cy, int width, int height)
664 int x1 = cx - width/2;
665 int y1 = cy - height/2;
666 int x2 = cx + width/2;
667 int y2 = cy + height/2;
670 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
671 patch.create(height, width, input.type());
672 patch.setTo(double(0.f));
676 int top = 0, bottom = 0, left = 0, right = 0;
678 //fit to image coordinates, set border extensions;
687 if (x2 >= input.cols) {
688 right = x2 - input.cols + width % 2;
693 if (y2 >= input.rows) {
694 bottom = y2 - input.rows + height % 2;
699 if (x2 - x1 == 0 || y2 - y1 == 0)
700 patch = cv::Mat::zeros(height, width, CV_32FC1);
703 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right, cv::BORDER_REPLICATE);
704 // imshow( "copyMakeBorder", patch);
709 assert(patch.cols == width && patch.rows == height);
714 void KCF_Tracker::gaussian_correlation(struct Scale_vars & vars, const ComplexMat & xf, const ComplexMat & yf, double sigma, bool auto_correlation)
717 xf.sqr_norm(vars.xf_sqr_norm_d);
718 if (!auto_correlation)
719 yf.sqr_norm(vars.yf_sqr_norm_d);
721 xf.sqr_norm(vars.xf_sqr_norm);
722 if (auto_correlation){
723 vars.yf_sqr_norm[0] = vars.xf_sqr_norm[0];
725 yf.sqr_norm(vars.yf_sqr_norm);
728 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
729 DEBUG_PRINTM(vars.xyf);
730 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features_d : nullptr, vars.stream);
733 cuda_gaussian_correlation(vars.data_i_features, vars.gauss_corr_res_d, vars.xf_sqr_norm_d, vars.xf_sqr_norm_d,
734 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
736 cuda_gaussian_correlation(vars.data_i_features, vars.gauss_corr_res_d, vars.xf_sqr_norm_d, vars.yf_sqr_norm_d,
737 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
739 //ifft2 and sum over 3rd dimension, we dont care about individual channels
740 DEBUG_PRINTM(vars.ifft2_res);
742 if (xf.channels() != p_num_scales*p_num_of_feats)
743 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
745 xy_sum.create(vars.ifft2_res.size(), CV_32FC(int(p_scales.size())));
747 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
748 float * row_ptr = vars.ifft2_res.ptr<float>(y);
749 float * row_ptr_sum = xy_sum.ptr<float>(y);
750 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
751 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
752 row_ptr_sum[(x*xy_sum.channels())+sum_ch] += std::accumulate(row_ptr + x*vars.ifft2_res.channels() + sum_ch*(vars.ifft2_res.channels()/xy_sum.channels()),
753 (row_ptr + x*vars.ifft2_res.channels() + (sum_ch+1)*(vars.ifft2_res.channels()/xy_sum.channels())), 0.f);
757 DEBUG_PRINTM(xy_sum);
759 std::vector<cv::Mat> scales;
760 cv::split(xy_sum,scales);
762 float numel_xf_inv = 1.f/(xf.cols * xf.rows * (xf.channels()/xf.n_scales));
763 for (uint i = 0; i < uint(xf.n_scales); ++i){
764 cv::Mat in_roi(vars.in_all, cv::Rect(0, int(i)*scales[0].rows, scales[0].cols, scales[0].rows));
765 cv::exp(- 1. / (sigma * sigma) * cv::max((double(vars.xf_sqr_norm[i] + vars.yf_sqr_norm[0]) - 2 * scales[i]) * double(numel_xf_inv), 0), in_roi);
766 DEBUG_PRINTM(in_roi);
769 DEBUG_PRINTM(vars.in_all);
770 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res_d : nullptr, vars.stream);
774 float get_response_circular(cv::Point2i & pt, cv::Mat & response)
779 x = response.cols + x;
781 y = response.rows + y;
782 if (x >= response.cols)
783 x = x - response.cols;
784 if (y >= response.rows)
785 y = y - response.rows;
787 return response.at<float>(y,x);
790 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point & max_loc, cv::Mat & response)
792 //find neighbourhood of max_loc (response is circular)
796 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);
797 cv::Point2i p4(max_loc.x-1, max_loc.y), p5(max_loc.x+1, max_loc.y);
798 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);
801 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
802 cv::Mat A = (cv::Mat_<float>(9, 6) <<
803 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
804 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
805 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
806 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
807 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
808 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
809 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
810 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
811 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);
812 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
813 get_response_circular(p1, response),
814 get_response_circular(p2, response),
815 get_response_circular(p3, response),
816 get_response_circular(p4, response),
817 get_response_circular(p5, response),
818 get_response_circular(p6, response),
819 get_response_circular(p7, response),
820 get_response_circular(p8, response),
821 get_response_circular(max_loc, response));
824 cv::solve(A, fval, x, cv::DECOMP_SVD);
826 float a = x.at<float>(0), b = x.at<float>(1), c = x.at<float>(2),
827 d = x.at<float>(3), e = x.at<float>(4);
829 cv::Point2f sub_peak(max_loc.x, max_loc.y);
830 if (b > 0 || b < 0) {
831 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
832 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
838 double KCF_Tracker::sub_grid_scale(int index)
841 if (index < 0 || index > int(p_scales.size())-1) {
842 // interpolate from all values
843 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
844 A.create(int(p_scales.size()), 3, CV_32FC1);
845 fval.create(int(p_scales.size()), 1, CV_32FC1);
846 for (auto it = p_scale_vars.begin(); it != p_scale_vars.end(); ++it) {
847 uint i = uint(std::distance(p_scale_vars.begin(), it));
849 A.at<float>(j, 0) = float(p_scales[i] * p_scales[i]);
850 A.at<float>(j, 1) = float(p_scales[i]);
851 A.at<float>(j, 2) = 1;
852 fval.at<float>(j) = m_use_big_batch ? float(p_scale_vars.back()->max_responses[i]) : float((*it)->max_response);
855 //only from neighbours
856 if (index == 0 || index == int(p_scales.size())-1)
857 return p_scales[uint(index)];
859 A = (cv::Mat_<float>(3, 3) <<
860 p_scales[uint(index)-1] * p_scales[uint(index)-1], p_scales[uint(index)-1], 1,
861 p_scales[uint(index)] * p_scales[uint(index)], p_scales[uint(index)], 1,
862 p_scales[uint(index)+1] * p_scales[uint(index)+1], p_scales[uint(index)+1], 1);
863 auto it1 = p_scale_vars.begin();
864 std::advance(it1, index-1);
865 auto it2 = p_scale_vars.begin();
866 std::advance(it2, index);
867 auto it3 = p_scale_vars.begin();
868 std::advance(it3, index+1);
869 fval = (cv::Mat_<float>(3, 1) << (m_use_big_batch ? p_scale_vars.back()->max_responses[uint(index)-1] : (*it1)->max_response),
870 (m_use_big_batch ? p_scale_vars.back()->max_responses[uint(index)] : (*it2)->max_response),
871 (m_use_big_batch ? p_scale_vars.back()->max_responses[uint(index)+1] : (*it3)->max_response));
875 cv::solve(A, fval, x, cv::DECOMP_SVD);
876 float a = x.at<float>(0), b = x.at<float>(1);
877 double scale = p_scales[uint(index)];
879 scale = double(-b / (2 * a));