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;}
25 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor, double output_sigma_factor, int cell_size) :
27 p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
28 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size) {}
30 KCF_Tracker::KCF_Tracker()
33 KCF_Tracker::~KCF_Tracker()
37 for (int i = 0;i < p_num_scales;++i) {
38 CudaSafeCall(cudaFreeHost(scale_vars[i].xf_sqr_norm));
39 CudaSafeCall(cudaFreeHost(scale_vars[i].yf_sqr_norm));
40 CudaSafeCall(cudaFree(scale_vars[i].gauss_corr_res));
43 for (int i = 0;i < p_num_scales;++i) {
44 free(scale_vars[i].xf_sqr_norm);
45 free(scale_vars[i].yf_sqr_norm);
50 void KCF_Tracker::init(cv::Mat &img, const cv::Rect & bbox, int fit_size_x, int fit_size_y)
52 //check boundary, enforce min size
53 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
55 if (x2 > img.cols-1) x2 = img.cols - 1;
57 if (y2 > img.rows-1) y2 = img.rows - 1;
59 if (x2-x1 < 2*p_cell_size) {
60 double diff = (2*p_cell_size -x2+x1)/2.;
61 if (x1 - diff >= 0 && x2 + diff < img.cols){
64 } else if (x1 - 2*diff >= 0) {
70 if (y2-y1 < 2*p_cell_size) {
71 double diff = (2*p_cell_size -y2+y1)/2.;
72 if (y1 - diff >= 0 && y2 + diff < img.rows){
75 } else if (y1 - 2*diff >= 0) {
84 p_pose.cx = x1 + p_pose.w/2.;
85 p_pose.cy = y1 + p_pose.h /2.;
88 cv::Mat input_gray, input_rgb = img.clone();
89 if (img.channels() == 3){
90 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
91 input_gray.convertTo(input_gray, CV_32FC1);
93 img.convertTo(input_gray, CV_32FC1);
95 // don't need too large image
96 if (p_pose.w * p_pose.h > 100.*100. && (fit_size_x == -1 || fit_size_y == -1)) {
97 std::cout << "resizing image by factor of " << 1/p_downscale_factor << std::endl;
98 p_resize_image = true;
99 p_pose.scale(p_downscale_factor);
100 cv::resize(input_gray, input_gray, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
101 cv::resize(input_rgb, input_rgb, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
102 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
103 if (fit_size_x%p_cell_size != 0 || fit_size_y%p_cell_size != 0) {
104 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;;
105 std::exit(EXIT_FAILURE);
108 if (( tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_x)
109 p_scale_factor_x = fit_size_x/tmp;
110 if (( tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_y)
111 p_scale_factor_y = fit_size_y/tmp;
112 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x
113 << " and verticaly by factor of " << p_scale_factor_y << std::endl;
115 p_pose.scale_x(p_scale_factor_x);
116 p_pose.scale_y(p_scale_factor_y);
117 if (p_scale_factor_x != 1 && p_scale_factor_y != 1) {
118 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
119 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
120 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
122 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
123 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
128 //compute win size + fit to fhog cell size
129 p_windows_size[0] = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
130 p_windows_size[1] = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
134 for (int i = -p_num_scales/2; i <= p_num_scales/2; ++i)
135 p_scales.push_back(std::pow(p_scale_step, i));
137 p_scales.push_back(1.);
139 for (int i = 0;i<p_num_scales;++i) {
140 scale_vars.push_back(Scale_vars());
144 if(m_use_color) p_num_of_feats += 3;
145 if(m_use_cnfeat) p_num_of_feats += 10;
146 p_roi_width = p_windows_size[0]/p_cell_size;
147 p_roi_height = p_windows_size[1]/p_cell_size;
151 p_current_scale = 1.;
153 double min_size_ratio = std::max(5.*p_cell_size/p_windows_size[0], 5.*p_cell_size/p_windows_size[1]);
154 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]);
155 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
156 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
158 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
159 std::cout << "init: win size. " << p_windows_size[0] << " " << p_windows_size[1] << std::endl;
160 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
162 p_output_sigma = std::sqrt(p_pose.w*p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
164 fft.init(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size, p_num_of_feats, p_num_scales, m_use_big_batch);
165 fft.set_window(cosine_window_function(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
167 scale_vars[0].flag = Tracker_flags::TRACKER_INIT;
168 #if defined(FFTW) || defined(CUFFT)
169 p_model_xf.create(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
170 p_yf.create(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
172 p_model_xf.create(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, p_num_of_feats);
173 p_yf.create(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, 1);
175 scale_vars[0].p_model_xf_ptr = & p_model_xf;
176 scale_vars[0].p_yf_ptr = & p_yf;
177 //window weights, i.e. labels
178 gaussian_shaped_labels(scale_vars[0], p_output_sigma, p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size);
179 fft.forward(scale_vars[0]);
182 //obtain a sub-window for training initial model
183 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], scale_vars[0]);
184 fft.forward_window(scale_vars[0]);
185 DEBUG_PRINTM(p_model_xf);
186 scale_vars[0].flag = Tracker_flags::AUTO_CORRELATION;
189 if (m_use_linearkernel) {
190 ComplexMat xfconj = p_model_xf.conj();
191 p_model_alphaf_num = xfconj.mul(p_yf);
192 p_model_alphaf_den = (p_model_xf * xfconj);
194 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
195 gaussian_correlation(scale_vars[0], p_model_xf, p_model_xf, p_kernel_sigma, true);
196 DEBUG_PRINTM(scale_vars[0].kf);
197 p_model_alphaf_num = p_yf * scale_vars[0].kf;
198 DEBUG_PRINTM(p_model_alphaf_num);
199 p_model_alphaf_den = scale_vars[0].kf * (scale_vars[0].kf + p_lambda);
200 DEBUG_PRINTM(p_model_alphaf_den);
202 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
203 DEBUG_PRINTM(p_model_alphaf);
204 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
207 void KCF_Tracker::init_scale_vars()
212 if (p_windows_size[1]/p_cell_size*(p_windows_size[0]/p_cell_size/2+1) > 1024) {
213 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
214 "the window dimensions so its size is less or equal to " << 1024*p_cell_size*p_cell_size*2+1 <<
215 " pixels . Currently the size of the window is: " << p_windows_size[0] << "x" << p_windows_size[1] <<
216 " which is " << p_windows_size[0]*p_windows_size[1] << " pixels. " << std::endl;
217 std::exit(EXIT_FAILURE);
220 if (m_use_linearkernel){
221 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
222 std::exit(EXIT_FAILURE);
224 cudaSetDeviceFlags(cudaDeviceMapHost);
226 for (int i = 0;i<p_num_scales;++i) {
227 alloc_size = p_windows_size[0]/p_cell_size*p_windows_size[1]/p_cell_size*sizeof(cufftReal);
228 CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].data_i_1ch, alloc_size, cudaHostAllocMapped));
229 CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].data_i_1ch_d, (void*)scale_vars[i].data_i_1ch, 0));
230 alloc_size = p_windows_size[0]/p_cell_size*p_windows_size[1]/p_cell_size*p_num_of_feats*sizeof(cufftReal);
231 CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].data_i_features, alloc_size, cudaHostAllocMapped));
232 CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].data_i_features_d, (void*)scale_vars[i].data_i_features, 0));
235 scale_vars[i].ifft2_res = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC(p_num_of_feats), scale_vars[i].data_i_features);
236 scale_vars[i].response = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC1, scale_vars[i].data_i_1ch);
238 scale_vars[i].zf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
239 scale_vars[i].kzf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
240 scale_vars[i].kf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
243 scale_vars[i].xf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
246 alloc_size = p_num_of_feats;
251 CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].xf_sqr_norm, alloc_size*sizeof(float), cudaHostAllocMapped));
252 CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].xf_sqr_norm_d, (void*)scale_vars[i].xf_sqr_norm, 0));
254 CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].yf_sqr_norm, sizeof(float), cudaHostAllocMapped));
255 CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].yf_sqr_norm_d, (void*)scale_vars[i].yf_sqr_norm, 0));
257 CudaSafeCall(cudaMalloc((void**)&scale_vars[i].gauss_corr_res, (p_windows_size[0]/p_cell_size)*(p_windows_size[1]/p_cell_size)*alloc_size*sizeof(float)));
261 alloc_size = p_num_of_feats;
265 for (int i = 0;i<p_num_scales;++i) {
266 scale_vars[i].xf_sqr_norm = (float*) malloc(alloc_size*sizeof(float));
267 scale_vars[i].yf_sqr_norm = (float*) malloc(sizeof(float));
269 scale_vars[i].patch_feats.reserve(p_num_of_feats);
271 scale_vars[i].ifft2_res = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC(p_num_of_feats));
272 scale_vars[i].response = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC1);
274 scale_vars[i].zf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
275 scale_vars[i].kzf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
276 scale_vars[i].kf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
277 //We use scale_vars[0] for updating the tracker, so we only allocate memory for its xf only.
279 scale_vars[i].xf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
281 scale_vars[i].zf = ComplexMat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, p_num_of_feats);
283 scale_vars[i].xf = ComplexMat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, p_num_of_feats);
289 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat & img, int fit_size_x, int fit_size_y)
291 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
294 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
296 if (p_resize_image) {
298 tmp.scale(p_downscale_factor);
301 } else if (p_fit_to_pw2) {
303 tmp.scale_x(p_scale_factor_x);
304 tmp.scale_y(p_scale_factor_y);
313 BBox_c KCF_Tracker::getBBox()
316 tmp.w *= p_current_scale;
317 tmp.h *= p_current_scale;
320 tmp.scale(1/p_downscale_factor);
322 tmp.scale_x(1/p_scale_factor_x);
323 tmp.scale_y(1/p_scale_factor_y);
329 void KCF_Tracker::track(cv::Mat &img)
332 std::cout << "NEW FRAME" << '\n';
333 cv::Mat input_gray, input_rgb = img.clone();
334 if (img.channels() == 3){
335 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
336 input_gray.convertTo(input_gray, CV_32FC1);
338 img.convertTo(input_gray, CV_32FC1);
340 // don't need too large image
341 if (p_resize_image) {
342 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
343 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
344 } else if (p_fit_to_pw2 && p_scale_factor_x != 1 && p_scale_factor_y != 1) {
345 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
346 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
347 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
349 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
350 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
354 double max_response = -1.;
356 cv::Point2i *max_response_pt = nullptr;
357 cv::Mat *max_response_map = nullptr;
359 if(m_use_multithreading) {
360 std::vector<std::future<void>> async_res(p_scales.size());
361 for (size_t i = 0; i < scale_vars.size(); ++i) {
362 async_res[i] = std::async(std::launch::async,
363 [this, &input_gray, &input_rgb, i]() -> void
364 {return scale_track(this->scale_vars[i], input_rgb, input_gray, this->p_scales[i]);});
366 for (size_t i = 0; i < p_scales.size(); ++i) {
368 if (this->scale_vars[i].max_response > max_response) {
369 max_response = this->scale_vars[i].max_response;
370 max_response_pt = & this->scale_vars[i].max_loc;
371 max_response_map = & this->scale_vars[i].response;
376 #pragma omp parallel for schedule(dynamic)
377 for (size_t i = 0; i < scale_vars.size(); ++i) {
378 scale_track(this->scale_vars[i], input_rgb, input_gray, this->p_scales[i]);
381 if (this->scale_vars[i].max_response > max_response) {
382 max_response = this->scale_vars[i].max_response;
383 max_response_pt = & this->scale_vars[i].max_loc;
384 max_response_map = & this->scale_vars[i].response;
391 DEBUG_PRINTM(*max_response_map);
392 DEBUG_PRINT(*max_response_pt);
394 //sub pixel quadratic interpolation from neighbours
395 if (max_response_pt->y > max_response_map->rows / 2) //wrap around to negative half-space of vertical axis
396 max_response_pt->y = max_response_pt->y - max_response_map->rows;
397 if (max_response_pt->x > max_response_map->cols / 2) //same for horizontal axis
398 max_response_pt->x = max_response_pt->x - max_response_map->cols;
400 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
401 DEBUG_PRINT(new_location);
403 if (m_use_subpixel_localization)
404 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
405 DEBUG_PRINT(new_location);
407 p_pose.cx += p_current_scale*p_cell_size*new_location.x;
408 p_pose.cy += p_current_scale*p_cell_size*new_location.y;
410 if (p_pose.cx < 0) p_pose.cx = 0;
411 if (p_pose.cx > (img.cols*p_scale_factor_x)-1) p_pose.cx = (img.cols*p_scale_factor_x)-1;
412 if (p_pose.cy < 0) p_pose.cy = 0;
413 if (p_pose.cy > (img.rows*p_scale_factor_y)-1) p_pose.cy = (img.rows*p_scale_factor_y)-1;
415 if (p_pose.cx < 0) p_pose.cx = 0;
416 if (p_pose.cx > img.cols-1) p_pose.cx = img.cols-1;
417 if (p_pose.cy < 0) p_pose.cy = 0;
418 if (p_pose.cy > img.rows-1) p_pose.cy = img.rows-1;
421 //sub grid scale interpolation
422 double new_scale = p_scales[scale_index];
423 if (m_use_subgrid_scale)
424 new_scale = sub_grid_scale(scale_index);
426 p_current_scale *= new_scale;
428 if (p_current_scale < p_min_max_scale[0])
429 p_current_scale = p_min_max_scale[0];
430 if (p_current_scale > p_min_max_scale[1])
431 p_current_scale = p_min_max_scale[1];
432 //obtain a subwindow for training at newly estimated target position
433 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], scale_vars[0], p_current_scale);
434 scale_vars[0].flag = Tracker_flags::TRACKER_UPDATE;
435 fft.forward_window(scale_vars[0]);
437 //subsequent frames, interpolate model
438 p_model_xf = p_model_xf * (1. - p_interp_factor) + scale_vars[0].xf * p_interp_factor;
440 ComplexMat alphaf_num, alphaf_den;
442 if (m_use_linearkernel) {
443 ComplexMat xfconj = scale_vars[0].xf.conj();
444 alphaf_num = xfconj.mul(p_yf);
445 alphaf_den = (scale_vars[0].xf * xfconj);
447 scale_vars[0].flag = Tracker_flags::AUTO_CORRELATION;
448 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
449 gaussian_correlation(scale_vars[0], scale_vars[0].xf, scale_vars[0].xf, p_kernel_sigma, true);
450 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
451 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
452 alphaf_num = p_yf * scale_vars[0].kf;
453 alphaf_den = scale_vars[0].kf * (scale_vars[0].kf + p_lambda);
456 p_model_alphaf_num = p_model_alphaf_num * (1. - p_interp_factor) + alphaf_num * p_interp_factor;
457 p_model_alphaf_den = p_model_alphaf_den * (1. - p_interp_factor) + alphaf_den * p_interp_factor;
458 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
461 // ****************************************************************************
463 void KCF_Tracker::scale_track(Scale_vars & vars, cv::Mat & input_rgb, cv::Mat & input_gray, double scale)
465 get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size[0], this->p_windows_size[1],
466 vars, this->p_current_scale * scale);
468 vars.flag = Tracker_flags::SCALE_RESPONSE;
469 fft.forward_window(vars);
470 DEBUG_PRINTM(vars.zf);
472 if (m_use_linearkernel) {
473 vars.kzf = (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels();
474 vars.flag = Tracker_flags::RESPONSE;
477 vars.flag = Tracker_flags::CROSS_CORRELATION;
478 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
479 DEBUG_PRINTM(this->p_model_alphaf);
480 DEBUG_PRINTM(vars.kzf);
481 DEBUG_PRINTM(this->p_model_alphaf * vars.kzf);
482 vars.flag = Tracker_flags::RESPONSE;
483 vars.kzf = this->p_model_alphaf * vars.kzf;
484 //TODO Add support for fft.inverse(vars) for CUFFT
488 DEBUG_PRINTM(vars.response);
490 /* target location is at the maximum response. we must take into
491 account the fact that, if the target doesn't move, the peak
492 will appear at the top-left corner, not at the center (this is
493 discussed in the paper). the responses wrap around cyclically. */
496 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
498 DEBUG_PRINT(vars.max_loc);
500 double weight = scale < 1. ? scale : 1./scale;
501 vars.max_response = vars.max_val*weight;
504 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)
506 int size_x_scaled = floor(size_x*scale);
507 int size_y_scaled = floor(size_y*scale);
509 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
510 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
512 //resize to default size
514 //if we downsample use INTER_AREA interpolation
515 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
517 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
520 // get hog(Histogram of Oriented Gradients) features
521 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
523 //get color rgb features (simple r,g,b channels)
524 std::vector<cv::Mat> color_feat;
525 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
526 //resize to default size
528 //if we downsample use INTER_AREA interpolation
529 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_AREA);
531 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_LINEAR);
535 if (m_use_color && input_rgb.channels() == 3) {
536 //use rgb color space
537 cv::Mat patch_rgb_norm;
538 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
539 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
540 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
541 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
542 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
543 cv::split(patch_rgb_norm, rgb);
544 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
547 if (m_use_cnfeat && input_rgb.channels() == 3) {
548 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
549 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
552 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
556 void KCF_Tracker::gaussian_shaped_labels(Scale_vars & vars, double sigma, int dim1, int dim2)
558 cv::Mat labels(dim2, dim1, CV_32FC1);
559 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
560 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
562 double sigma_s = sigma*sigma;
564 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j){
565 float * row_ptr = labels.ptr<float>(j);
567 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i){
568 row_ptr[i] = std::exp(-0.5 * (y_s + x*x) / sigma_s);//-1/2*e^((y^2+x^2)/sigma^2)
572 //rotate so that 1 is at top-left corner (see KCF paper for explanation)
573 vars.rot_labels = circshift(labels, range_x[0], range_y[0]);
574 //sanity check, 1 at top left corner
575 assert(vars.rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
580 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
582 cv::Mat rot_patch(patch.size(), CV_32FC1);
583 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
585 //circular rotate x-axis
587 //move part that does not rotate over the edge
588 cv::Range orig_range(-x_rot, patch.cols);
589 cv::Range rot_range(0, patch.cols - (-x_rot));
590 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
593 orig_range = cv::Range(0, -x_rot);
594 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
595 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
596 }else if (x_rot > 0){
597 //move part that does not rotate over the edge
598 cv::Range orig_range(0, patch.cols - x_rot);
599 cv::Range rot_range(x_rot, patch.cols);
600 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
603 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
604 rot_range = cv::Range(0, x_rot);
605 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
606 }else { //zero rotation
607 //move part that does not rotate over the edge
608 cv::Range orig_range(0, patch.cols);
609 cv::Range rot_range(0, patch.cols);
610 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
613 //circular rotate y-axis
615 //move part that does not rotate over the edge
616 cv::Range orig_range(-y_rot, patch.rows);
617 cv::Range rot_range(0, patch.rows - (-y_rot));
618 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
621 orig_range = cv::Range(0, -y_rot);
622 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
623 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
624 }else if (y_rot > 0){
625 //move part that does not rotate over the edge
626 cv::Range orig_range(0, patch.rows - y_rot);
627 cv::Range rot_range(y_rot, patch.rows);
628 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
631 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
632 rot_range = cv::Range(0, y_rot);
633 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
634 }else { //zero rotation
635 //move part that does not rotate over the edge
636 cv::Range orig_range(0, patch.rows);
637 cv::Range rot_range(0, patch.rows);
638 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
644 //hann window actually (Power-of-cosine windows)
645 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
647 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
648 double N_inv = 1./(static_cast<double>(dim1)-1.);
649 for (int i = 0; i < dim1; ++i)
650 m1.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
651 N_inv = 1./(static_cast<double>(dim2)-1.);
652 for (int i = 0; i < dim2; ++i)
653 m2.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
658 // Returns sub-window of image input centered at [cx, cy] coordinates),
659 // with size [width, height]. If any pixels are outside of the image,
660 // they will replicate the values at the borders.
661 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat & input, int cx, int cy, int width, int height)
665 int x1 = cx - width/2;
666 int y1 = cy - height/2;
667 int x2 = cx + width/2;
668 int y2 = cy + height/2;
671 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
672 patch.create(height, width, input.type());
677 int top = 0, bottom = 0, left = 0, right = 0;
679 //fit to image coordinates, set border extensions;
688 if (x2 >= input.cols) {
689 right = x2 - input.cols + width % 2;
694 if (y2 >= input.rows) {
695 bottom = y2 - input.rows + height % 2;
700 if (x2 - x1 == 0 || y2 - y1 == 0)
701 patch = cv::Mat::zeros(height, width, CV_32FC1);
704 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right, cv::BORDER_REPLICATE);
705 // imshow( "copyMakeBorder", patch);
710 assert(patch.cols == width && patch.rows == height);
715 void KCF_Tracker::gaussian_correlation(struct Scale_vars & vars, const ComplexMat & xf, const ComplexMat & yf, double sigma, bool auto_correlation)
718 xf.sqr_norm(vars.xf_sqr_norm_d);
719 if (!auto_correlation)
720 yf.sqr_norm(vars.yf_sqr_norm_d);
722 xf.sqr_norm(vars.xf_sqr_norm);
723 if (auto_correlation){
724 vars.yf_sqr_norm[0] = vars.xf_sqr_norm[0];
726 yf.sqr_norm(vars.yf_sqr_norm);
729 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
730 DEBUG_PRINTM(vars.xyf);
734 cuda_gaussian_correlation(vars.data_i_features, vars.gauss_corr_res, vars.xf_sqr_norm_d, vars.xf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
736 cuda_gaussian_correlation(vars.data_i_features, vars.gauss_corr_res, vars.xf_sqr_norm_d, vars.yf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
737 fft.forward_raw(vars, xf.n_scales==p_num_scales);
740 //ifft2 and sum over 3rd dimension, we dont care about individual channels
742 DEBUG_PRINTM(vars.ifft2_res);
744 if (xf.channels() != p_num_scales*p_num_of_feats)
745 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
747 xy_sum.create(vars.ifft2_res.size(), CV_32FC(p_scales.size()));
749 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
750 float * row_ptr = vars.ifft2_res.ptr<float>(y);
751 float * row_ptr_sum = xy_sum.ptr<float>(y);
752 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
753 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
754 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()),
755 (row_ptr + x*vars.ifft2_res.channels() + (sum_ch+1)*(vars.ifft2_res.channels()/xy_sum.channels())), 0.f);
759 DEBUG_PRINTM(xy_sum);
761 std::vector<cv::Mat> scales;
762 cv::split(xy_sum,scales);
763 vars.in_all = cv::Mat(scales[0].rows * xf.n_scales, scales[0].cols, CV_32F);
765 float numel_xf_inv = 1.f/(xf.cols * xf.rows * (xf.channels()/xf.n_scales));
766 for (int i = 0; i < xf.n_scales; ++i){
767 cv::Mat in_roi(vars.in_all, cv::Rect(0, i*scales[0].rows, scales[0].cols, scales[0].rows));
768 cv::exp(- 1.f / (sigma * sigma) * cv::max((vars.xf_sqr_norm[i] + vars.yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0), in_roi);
769 DEBUG_PRINTM(in_roi);
772 DEBUG_PRINTM(vars.in_all );
777 float get_response_circular(cv::Point2i & pt, cv::Mat & response)
782 x = response.cols + x;
784 y = response.rows + y;
785 if (x >= response.cols)
786 x = x - response.cols;
787 if (y >= response.rows)
788 y = y - response.rows;
790 return response.at<float>(y,x);
793 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point & max_loc, cv::Mat & response)
795 //find neighbourhood of max_loc (response is circular)
799 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);
800 cv::Point2i p4(max_loc.x-1, max_loc.y), p5(max_loc.x+1, max_loc.y);
801 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);
804 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
805 cv::Mat A = (cv::Mat_<float>(9, 6) <<
806 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
807 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
808 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
809 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
810 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
811 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
812 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
813 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
814 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);
815 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
816 get_response_circular(p1, response),
817 get_response_circular(p2, response),
818 get_response_circular(p3, response),
819 get_response_circular(p4, response),
820 get_response_circular(p5, response),
821 get_response_circular(p6, response),
822 get_response_circular(p7, response),
823 get_response_circular(p8, response),
824 get_response_circular(max_loc, response));
827 cv::solve(A, fval, x, cv::DECOMP_SVD);
829 double a = x.at<float>(0), b = x.at<float>(1), c = x.at<float>(2),
830 d = x.at<float>(3), e = x.at<float>(4);
832 cv::Point2f sub_peak(max_loc.x, max_loc.y);
833 if (b > 0 || b < 0) {
834 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
835 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
841 double KCF_Tracker::sub_grid_scale(int index)
844 if (index < 0 || index > (int)p_scales.size()-1) {
845 // interpolate from all values
846 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
847 A.create(p_scales.size(), 3, CV_32FC1);
848 fval.create(p_scales.size(), 1, CV_32FC1);
849 for (size_t i = 0; i < p_scales.size(); ++i) {
850 A.at<float>(i, 0) = p_scales[i] * p_scales[i];
851 A.at<float>(i, 1) = p_scales[i];
852 A.at<float>(i, 2) = 1;
853 fval.at<float>(i) = scale_vars[i].max_response;
856 //only from neighbours
857 if (index == 0 || index == (int)p_scales.size()-1)
858 return p_scales[index];
860 A = (cv::Mat_<float>(3, 3) <<
861 p_scales[index-1] * p_scales[index-1], p_scales[index-1], 1,
862 p_scales[index] * p_scales[index], p_scales[index], 1,
863 p_scales[index+1] * p_scales[index+1], p_scales[index+1], 1);
864 fval = (cv::Mat_<float>(3, 1) << scale_vars[index-1].max_response, scale_vars[index].max_response, scale_vars[index+1].max_response);
868 cv::solve(A, fval, x, cv::DECOMP_SVD);
869 double a = x.at<float>(0), b = x.at<float>(1);
870 double scale = p_scales[index];
872 scale = -b / (2 * a);