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 << "Error: Fit size is not multiple of HOG cell size (" << p_cell_size << ")" << std::endl;
100 std::exit(EXIT_FAILURE);
102 double tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
103 if (fabs(tmp - fit_size_x) > p_floating_error)
104 p_scale_factor_x = fit_size_x / tmp;
105 tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
106 if (fabs(tmp - fit_size_y) > p_floating_error)
107 p_scale_factor_y = fit_size_y / tmp;
108 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
109 << p_scale_factor_y << std::endl;
111 p_pose.scale_x(p_scale_factor_x);
112 p_pose.scale_y(p_scale_factor_y);
113 if (fabs(p_scale_factor_x - 1) > p_floating_error && fabs(p_scale_factor_y - 1) > p_floating_error) {
114 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
115 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
116 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
118 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y,
120 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
125 // compute win size + fit to fhog cell size
126 p_windows_size.width = int(round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size);
127 p_windows_size.height = int(round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size);
130 if (m_use_color) p_num_of_feats += 3;
131 if (m_use_cnfeat) p_num_of_feats += 10;
132 p_roi_width = p_windows_size.width / p_cell_size;
133 p_roi_height = p_windows_size.height / p_cell_size;
137 for (int i = -p_num_scales / 2; i <= p_num_scales / 2; ++i)
138 p_scales.push_back(std::pow(p_scale_step, i));
140 p_scales.push_back(1.);
143 if (p_windows_size.height / p_cell_size * (p_windows_size.width / p_cell_size / 2 + 1) > 1024) {
144 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
145 "the window dimensions so its size is less or equal to "
146 << 1024 * p_cell_size * p_cell_size * 2 + 1
147 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
148 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
149 std::exit(EXIT_FAILURE);
152 if (m_use_linearkernel) {
153 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
154 std::exit(EXIT_FAILURE);
156 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
157 p_rot_labels_data = DynMem(
158 ((uint(p_windows_size.width) / p_cell_size) * (uint(p_windows_size.height) / p_cell_size)) * sizeof(float));
159 p_rot_labels = cv::Mat(p_windows_size.height / int(p_cell_size), p_windows_size.width / int(p_cell_size), CV_32FC1,
160 p_rot_labels_data.hostMem());
162 p_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.height / p_cell_size)) / 2 + 1,
166 #if defined(CUFFT) || defined(FFTW)
167 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1,
168 uint(p_num_of_feats));
169 p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1, 1);
170 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size) / 2 + 1,
173 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)),
174 uint(p_num_of_feats));
175 p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)), 1);
176 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size), p_num_of_feats);
179 int max = m_use_big_batch ? 2 : p_num_scales;
180 for (int i = 0; i < max; ++i) {
181 if (m_use_big_batch && i == 1)
182 p_threadctxs.emplace_back(p_windows_size, p_cell_size, p_num_of_feats * p_num_scales, p_num_scales);
184 p_threadctxs.emplace_back(p_windows_size, p_cell_size, p_num_of_feats, 1);
187 p_current_scale = 1.;
189 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
190 double max_size_ratio =
191 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
192 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
193 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
194 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
196 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
197 std::cout << "init: win size. " << p_windows_size.width << " " << p_windows_size.height << std::endl;
198 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
200 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
202 fft.init(uint(p_windows_size.width / p_cell_size), uint(p_windows_size.height / p_cell_size), uint(p_num_of_feats),
203 uint(p_num_scales), m_use_big_batch);
204 fft.set_window(cosine_window_function(p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size));
206 // window weights, i.e. labels
208 gaussian_shaped_labels(p_output_sigma, p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size), p_yf,
209 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_threadctxs.front().stream);
212 // obtain a sub-window for training initial model
213 p_threadctxs.front().patch_feats.clear();
214 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
215 p_threadctxs.front());
216 fft.forward_window(p_threadctxs.front().patch_feats, p_model_xf, p_threadctxs.front().fw_all,
217 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr, p_threadctxs.front().stream);
218 DEBUG_PRINTM(p_model_xf);
219 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
220 p_threadctxs.front().model_xf = p_model_xf;
221 p_threadctxs.front().model_xf.set_stream(p_threadctxs.front().stream);
222 p_yf.set_stream(p_threadctxs.front().stream);
223 p_model_xf.set_stream(p_threadctxs.front().stream);
224 p_xf.set_stream(p_threadctxs.front().stream);
227 if (m_use_linearkernel) {
228 ComplexMat xfconj = p_model_xf.conj();
229 p_model_alphaf_num = xfconj.mul(p_yf);
230 p_model_alphaf_den = (p_model_xf * xfconj);
232 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
233 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
234 gaussian_correlation(p_threadctxs.front(), p_threadctxs.front().model_xf, p_threadctxs.front().model_xf,
235 p_kernel_sigma, true);
237 gaussian_correlation(p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
239 DEBUG_PRINTM(p_threadctxs.front().kf);
240 p_model_alphaf_num = p_yf * p_threadctxs.front().kf;
241 DEBUG_PRINTM(p_model_alphaf_num);
242 p_model_alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
243 DEBUG_PRINTM(p_model_alphaf_den);
245 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
246 DEBUG_PRINTM(p_model_alphaf);
247 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
249 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
250 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
251 it->model_xf = p_model_xf;
252 it->model_xf.set_stream(it->stream);
253 it->model_alphaf = p_model_alphaf;
254 it->model_alphaf.set_stream(it->stream);
259 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
261 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
264 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
266 if (p_resize_image) {
268 tmp.scale(p_downscale_factor);
271 } else if (p_fit_to_pw2) {
273 tmp.scale_x(p_scale_factor_x);
274 tmp.scale_y(p_scale_factor_y);
283 BBox_c KCF_Tracker::getBBox()
286 tmp.w *= p_current_scale;
287 tmp.h *= p_current_scale;
289 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
291 tmp.scale_x(1 / p_scale_factor_x);
292 tmp.scale_y(1 / p_scale_factor_y);
298 void KCF_Tracker::track(cv::Mat &img)
300 if (m_debug) std::cout << "NEW FRAME" << '\n';
301 cv::Mat input_gray, input_rgb = img.clone();
302 if (img.channels() == 3) {
303 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
304 input_gray.convertTo(input_gray, CV_32FC1);
306 img.convertTo(input_gray, CV_32FC1);
308 // don't need too large image
309 if (p_resize_image) {
310 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
311 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
312 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
313 fabs(p_scale_factor_y - 1) > p_floating_error) {
314 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
315 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
316 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
318 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
319 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
323 double max_response = -1.;
324 uint scale_index = 0;
325 cv::Point2i *max_response_pt = nullptr;
326 cv::Mat *max_response_map = nullptr;
328 if (m_use_multithreading) {
329 std::vector<std::future<void>> async_res(p_scales.size());
330 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
331 uint index = uint(std::distance(p_threadctxs.begin(), it));
332 async_res[index] = std::async(std::launch::async, [this, &input_gray, &input_rgb, index, it]() -> void {
333 return scale_track(*it, input_rgb, input_gray, this->p_scales[index]);
336 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
337 uint index = uint(std::distance(p_threadctxs.begin(), it));
338 async_res[index].wait();
339 if (it->max_response > max_response) {
340 max_response = it->max_response;
341 max_response_pt = &it->max_loc;
342 max_response_map = &it->response;
347 uint start = m_use_big_batch ? 1 : 0;
348 uint end = m_use_big_batch ? 2 : uint(p_num_scales);
349 NORMAL_OMP_PARALLEL_FOR
350 for (uint i = start; i < end; ++i) {
351 auto it = p_threadctxs.begin();
353 scale_track(*it, input_rgb, input_gray, this->p_scales[i]);
355 if (m_use_big_batch) {
356 for (size_t j = 0; j < p_scales.size(); ++j) {
357 if (it->max_responses[j] > max_response) {
358 max_response = it->max_responses[j];
359 max_response_pt = &it->max_locs[j];
360 max_response_map = &it->response_maps[j];
367 if (it->max_response > max_response) {
368 max_response = it->max_response;
369 max_response_pt = &it->max_loc;
370 max_response_map = &it->response;
378 DEBUG_PRINTM(*max_response_map);
379 DEBUG_PRINT(*max_response_pt);
381 // sub pixel quadratic interpolation from neighbours
382 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
383 max_response_pt->y = max_response_pt->y - max_response_map->rows;
384 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
385 max_response_pt->x = max_response_pt->x - max_response_map->cols;
387 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
388 DEBUG_PRINT(new_location);
390 if (m_use_subpixel_localization)
391 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
392 DEBUG_PRINT(new_location);
394 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
395 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
397 if (p_pose.cx < 0) p_pose.cx = 0;
398 if (p_pose.cx > (img.cols * p_scale_factor_x) - 1) p_pose.cx = (img.cols * p_scale_factor_x) - 1;
399 if (p_pose.cy < 0) p_pose.cy = 0;
400 if (p_pose.cy > (img.rows * p_scale_factor_y) - 1) p_pose.cy = (img.rows * p_scale_factor_y) - 1;
402 if (p_pose.cx < 0) p_pose.cx = 0;
403 if (p_pose.cx > img.cols - 1) p_pose.cx = img.cols - 1;
404 if (p_pose.cy < 0) p_pose.cy = 0;
405 if (p_pose.cy > img.rows - 1) p_pose.cy = img.rows - 1;
408 // sub grid scale interpolation
409 double new_scale = p_scales[scale_index];
410 if (m_use_subgrid_scale)
411 new_scale = sub_grid_scale(scale_index);
413 p_current_scale *= new_scale;
415 if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
416 if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
418 // obtain a subwindow for training at newly estimated target position
419 p_threadctxs.front().patch_feats.clear();
420 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
421 p_threadctxs.front(), p_current_scale);
422 fft.forward_window(p_threadctxs.front().patch_feats, p_xf, p_threadctxs.front().fw_all,
423 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr, p_threadctxs.front().stream);
425 // subsequent frames, interpolate model
426 p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
428 ComplexMat alphaf_num, alphaf_den;
430 if (m_use_linearkernel) {
431 ComplexMat xfconj = p_xf.conj();
432 alphaf_num = xfconj.mul(p_yf);
433 alphaf_den = (p_xf * xfconj);
435 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
436 gaussian_correlation(p_threadctxs.front(), p_xf, p_xf, p_kernel_sigma,
438 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
439 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
440 alphaf_num = p_yf * p_threadctxs.front().kf;
441 alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
444 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
445 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
446 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
448 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
449 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
450 it->model_xf = p_model_xf;
451 it->model_xf.set_stream(it->stream);
452 it->model_alphaf = p_model_alphaf;
453 it->model_alphaf.set_stream(it->stream);
458 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray, double scale)
460 if (m_use_big_batch) {
461 vars.patch_feats.clear();
462 BIG_BATCH_OMP_PARALLEL_FOR
463 for (uint i = 0; i < uint(p_num_scales); ++i) {
464 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
465 this->p_windows_size.height, vars, this->p_current_scale * this->p_scales[i]);
468 vars.patch_feats.clear();
469 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
470 this->p_windows_size.height, vars, this->p_current_scale *scale);
473 fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
475 DEBUG_PRINTM(vars.zf);
477 if (m_use_linearkernel) {
478 vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
479 : (p_model_alphaf * vars.zf).sum_over_channels();
480 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
482 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
483 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
484 vars.kzf = vars.model_alphaf * vars.kzf;
486 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
487 DEBUG_PRINTM(this->p_model_alphaf);
488 DEBUG_PRINTM(vars.kzf);
489 vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
491 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
494 DEBUG_PRINTM(vars.response);
496 /* target location is at the maximum response. we must take into
497 account the fact that, if the target doesn't move, the peak
498 will appear at the top-left corner, not at the center (this is
499 discussed in the paper). the responses wrap around cyclically. */
500 if (m_use_big_batch) {
501 cv::split(vars.response, vars.response_maps);
503 for (size_t i = 0; i < p_scales.size(); ++i) {
504 double min_val, max_val;
505 cv::Point2i min_loc, max_loc;
506 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
507 DEBUG_PRINT(max_loc);
508 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
509 vars.max_responses[i] = max_val * weight;
510 vars.max_locs[i] = max_loc;
515 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
517 DEBUG_PRINT(vars.max_loc);
519 double weight = scale < 1. ? scale : 1. / scale;
520 vars.max_response = vars.max_val * weight;
525 // ****************************************************************************
527 void KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y,
528 ThreadCtx &vars, double scale)
530 int size_x_scaled = int(floor(size_x * scale));
531 int size_y_scaled = int(floor(size_y * scale));
533 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
534 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
536 // resize to default size
538 // if we downsample use INTER_AREA interpolation
539 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
541 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
544 // get hog(Histogram of Oriented Gradients) features
545 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
547 // get color rgb features (simple r,g,b channels)
548 std::vector<cv::Mat> color_feat;
549 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
550 // resize to default size
552 // if we downsample use INTER_AREA interpolation
553 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
556 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
561 if (m_use_color && input_rgb.channels() == 3) {
562 // use rgb color space
563 cv::Mat patch_rgb_norm;
564 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
565 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
566 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
567 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
568 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
569 cv::split(patch_rgb_norm, rgb);
570 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
573 if (m_use_cnfeat && input_rgb.channels() == 3) {
574 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
575 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
577 BIG_BATCH_OMP_ORDERED
578 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
582 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
584 cv::Mat labels(dim2, dim1, CV_32FC1);
585 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
586 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
588 double sigma_s = sigma * sigma;
590 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
591 float *row_ptr = labels.ptr<float>(j);
593 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
594 row_ptr[i] = float(std::exp(-0.5 * (y_s + x * x) / sigma_s)); //-1/2*e^((y^2+x^2)/sigma^2)
598 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
600 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
601 tmp.copyTo(p_rot_labels);
603 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
606 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
607 // sanity check, 1 at top left corner
608 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
614 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
616 cv::Mat rot_patch(patch.size(), CV_32FC1);
617 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
619 // circular rotate x-axis
621 // move part that does not rotate over the edge
622 cv::Range orig_range(-x_rot, patch.cols);
623 cv::Range rot_range(0, patch.cols - (-x_rot));
624 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
627 orig_range = cv::Range(0, -x_rot);
628 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
629 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
630 } else if (x_rot > 0) {
631 // move part that does not rotate over the edge
632 cv::Range orig_range(0, patch.cols - x_rot);
633 cv::Range rot_range(x_rot, patch.cols);
634 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
637 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
638 rot_range = cv::Range(0, x_rot);
639 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
640 } else { // zero rotation
641 // move part that does not rotate over the edge
642 cv::Range orig_range(0, patch.cols);
643 cv::Range rot_range(0, patch.cols);
644 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
647 // circular rotate y-axis
649 // move part that does not rotate over the edge
650 cv::Range orig_range(-y_rot, patch.rows);
651 cv::Range rot_range(0, patch.rows - (-y_rot));
652 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
655 orig_range = cv::Range(0, -y_rot);
656 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
657 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
658 } else if (y_rot > 0) {
659 // move part that does not rotate over the edge
660 cv::Range orig_range(0, patch.rows - y_rot);
661 cv::Range rot_range(y_rot, patch.rows);
662 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
665 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
666 rot_range = cv::Range(0, y_rot);
667 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
668 } else { // zero rotation
669 // move part that does not rotate over the edge
670 cv::Range orig_range(0, patch.rows);
671 cv::Range rot_range(0, patch.rows);
672 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
678 // hann window actually (Power-of-cosine windows)
679 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
681 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
682 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
683 for (int i = 0; i < dim1; ++i)
684 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
685 N_inv = 1. / (static_cast<double>(dim2) - 1.);
686 for (int i = 0; i < dim2; ++i)
687 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
688 cv::Mat ret = m2 * m1;
692 // Returns sub-window of image input centered at [cx, cy] coordinates),
693 // with size [width, height]. If any pixels are outside of the image,
694 // they will replicate the values at the borders.
695 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
699 int x1 = cx - width / 2;
700 int y1 = cy - height / 2;
701 int x2 = cx + width / 2;
702 int y2 = cy + height / 2;
705 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
706 patch.create(height, width, input.type());
707 patch.setTo(double(0.f));
711 int top = 0, bottom = 0, left = 0, right = 0;
713 // fit to image coordinates, set border extensions;
722 if (x2 >= input.cols) {
723 right = x2 - input.cols + width % 2;
728 if (y2 >= input.rows) {
729 bottom = y2 - input.rows + height % 2;
734 if (x2 - x1 == 0 || y2 - y1 == 0)
735 patch = cv::Mat::zeros(height, width, CV_32FC1);
737 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
738 cv::BORDER_REPLICATE);
739 // imshow( "copyMakeBorder", patch);
744 assert(patch.cols == width && patch.rows == height);
749 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
750 double sigma, bool auto_correlation)
753 xf.sqr_norm(vars.xf_sqr_norm.deviceMem());
754 if (!auto_correlation) yf.sqr_norm(vars.yf_sqr_norm.deviceMem());
756 xf.sqr_norm(vars.xf_sqr_norm.hostMem());
757 if (auto_correlation) {
758 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
760 yf.sqr_norm(vars.yf_sqr_norm.hostMem());
763 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
764 DEBUG_PRINTM(vars.xyf);
765 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
767 if (auto_correlation)
768 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(),
769 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
771 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.yf_sqr_norm.deviceMem(),
772 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
774 // ifft2 and sum over 3rd dimension, we dont care about individual channels
775 DEBUG_PRINTM(vars.ifft2_res);
777 if (xf.channels() != p_num_scales * p_num_of_feats)
778 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
780 xy_sum.create(vars.ifft2_res.size(), CV_32FC(int(p_scales.size())));
782 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
783 float *row_ptr = vars.ifft2_res.ptr<float>(y);
784 float *row_ptr_sum = xy_sum.ptr<float>(y);
785 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
786 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
787 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
788 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
789 (row_ptr + x * vars.ifft2_res.channels() +
790 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
795 DEBUG_PRINTM(xy_sum);
797 std::vector<cv::Mat> scales;
798 cv::split(xy_sum, scales);
800 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
801 for (uint i = 0; i < uint(xf.n_scales); ++i) {
802 cv::Mat in_roi(vars.in_all, cv::Rect(0, int(i) * scales[0].rows, scales[0].cols, scales[0].rows));
804 -1. / (sigma * sigma) *
805 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
807 DEBUG_PRINTM(in_roi);
810 DEBUG_PRINTM(vars.in_all);
811 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
816 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
820 if (x < 0) x = response.cols + x;
821 if (y < 0) y = response.rows + y;
822 if (x >= response.cols) x = x - response.cols;
823 if (y >= response.rows) y = y - response.rows;
825 return response.at<float>(y, x);
828 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
830 // find neighbourhood of max_loc (response is circular)
834 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);
835 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
836 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);
839 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
840 cv::Mat A = (cv::Mat_<float>(9, 6) <<
841 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
842 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
843 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
844 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
845 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
846 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
847 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
848 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
849 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);
850 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
851 get_response_circular(p1, response),
852 get_response_circular(p2, response),
853 get_response_circular(p3, response),
854 get_response_circular(p4, response),
855 get_response_circular(p5, response),
856 get_response_circular(p6, response),
857 get_response_circular(p7, response),
858 get_response_circular(p8, response),
859 get_response_circular(max_loc, response));
862 cv::solve(A, fval, x, cv::DECOMP_SVD);
864 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);
866 cv::Point2f sub_peak(max_loc.x, max_loc.y);
867 if (b > 0 || b < 0) {
868 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
869 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
875 double KCF_Tracker::sub_grid_scale(int index)
878 if (index < 0 || index > int(p_scales.size()) - 1) {
879 // interpolate from all values
880 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
881 A.create(int(p_scales.size()), 3, CV_32FC1);
882 fval.create(int(p_scales.size()), 1, CV_32FC1);
883 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
884 uint i = uint(std::distance(p_threadctxs.begin(), it));
886 A.at<float>(j, 0) = float(p_scales[i] * p_scales[i]);
887 A.at<float>(j, 1) = float(p_scales[i]);
888 A.at<float>(j, 2) = 1;
890 m_use_big_batch ? float(p_threadctxs.back().max_responses[i]) : float(it->max_response);
893 // only from neighbours
894 if (index == 0 || index == int(p_scales.size()) - 1) return p_scales[uint(index)];
896 A = (cv::Mat_<float>(3, 3) << p_scales[uint(index) - 1] * p_scales[uint(index) - 1], p_scales[uint(index) - 1],
897 1, p_scales[uint(index)] * p_scales[uint(index)], p_scales[uint(index)], 1,
898 p_scales[uint(index) + 1] * p_scales[uint(index) + 1], p_scales[uint(index) + 1], 1);
899 auto it1 = p_threadctxs.begin();
900 std::advance(it1, index - 1);
901 auto it2 = p_threadctxs.begin();
902 std::advance(it2, index);
903 auto it3 = p_threadctxs.begin();
904 std::advance(it3, index + 1);
905 fval = (cv::Mat_<float>(3, 1) << (m_use_big_batch ? p_threadctxs.back().max_responses[uint(index) - 1]
906 : it1->max_response),
907 (m_use_big_batch ? p_threadctxs.back().max_responses[uint(index)] : it2->max_response),
908 (m_use_big_batch ? p_threadctxs.back().max_responses[uint(index) + 1] : it3->max_response));
912 cv::solve(A, fval, x, cv::DECOMP_SVD);
913 float a = x.at<float>(0), b = x.at<float>(1);
914 double scale = p_scales[uint(index)];
915 if (a > 0 || a < 0) scale = double(-b / (2 * a));