10 #include "fft_cufft.h"
13 #include "fft_opencv.h"
21 #define DEBUG_PRINT(obj) \
23 std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl; \
25 #define DEBUG_PRINTM(obj) \
27 std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl \
28 << (obj) << std::endl; \
31 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
32 double output_sigma_factor, int cell_size)
33 : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
34 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size)
38 KCF_Tracker::KCF_Tracker() : fft(*new FFT()) {}
40 KCF_Tracker::~KCF_Tracker()
45 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
47 // check boundary, enforce min size
48 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
50 if (x2 > img.cols - 1) x2 = img.cols - 1;
52 if (y2 > img.rows - 1) y2 = img.rows - 1;
54 if (x2 - x1 < 2 * p_cell_size) {
55 double diff = (2 * p_cell_size - x2 + x1) / 2.;
56 if (x1 - diff >= 0 && x2 + diff < img.cols) {
59 } else if (x1 - 2 * diff >= 0) {
65 if (y2 - y1 < 2 * p_cell_size) {
66 double diff = (2 * p_cell_size - y2 + y1) / 2.;
67 if (y1 - diff >= 0 && y2 + diff < img.rows) {
70 } else if (y1 - 2 * diff >= 0) {
79 p_pose.cx = x1 + p_pose.w / 2.;
80 p_pose.cy = y1 + p_pose.h / 2.;
82 cv::Mat input_gray, input_rgb = img.clone();
83 if (img.channels() == 3) {
84 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
85 input_gray.convertTo(input_gray, CV_32FC1);
87 img.convertTo(input_gray, CV_32FC1);
89 // don't need too large image
90 if (p_pose.w * p_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
91 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
92 p_resize_image = true;
93 p_pose.scale(p_downscale_factor);
94 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
95 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
96 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
97 if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
98 std::cerr << "Error: Fit size is not multiple of HOG cell size (" << p_cell_size << ")" << std::endl;
99 std::exit(EXIT_FAILURE);
101 p_scale_factor_x = (double)fit_size_x / round(p_pose.w * (1. + p_padding));
102 p_scale_factor_y = (double)fit_size_y / round(p_pose.h * (1. + p_padding));
103 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
104 << 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.width = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
121 p_windows_size.height = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
122 p_roi.width = p_windows_size.width / p_cell_size;
123 p_roi.height = p_windows_size.height / p_cell_size;
126 if (m_use_color) p_num_of_feats += 3;
127 if (m_use_cnfeat) p_num_of_feats += 10;
131 for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
132 p_scales.push_back(std::pow(p_scale_step, i));
134 p_scales.push_back(1.);
137 if (p_roi.height * (p_roi.width / 2 + 1) > 1024) {
138 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
139 "the window dimensions so its size is less or equal to "
140 << 1024 * p_cell_size * p_cell_size * 2 + 1
141 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
142 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
143 std::exit(EXIT_FAILURE);
146 if (m_use_linearkernel) {
147 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
148 std::exit(EXIT_FAILURE);
150 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
151 p_rot_labels_data = DynMem(p_roi.width * p_roi.height * sizeof(float));
152 p_rot_labels = cv::Mat(p_roi, CV_32FC1, p_rot_labels_data.hostMem());
154 p_xf.create(p_roi.height, p_roi.height / 2 + 1, p_num_of_feats);
157 #if defined(CUFFT) || defined(FFTW)
158 uint width = p_roi.width / 2 + 1;
160 uint width = p_roi.width;
162 p_model_xf.create(p_roi.height, width, p_num_of_feats);
163 p_yf.create(p_roi.height, width, 1);
164 p_xf.create(p_roi.height, width, p_num_of_feats);
166 int max = m_use_big_batch ? 2 : p_num_scales;
167 for (int i = 0; i < max; ++i) {
168 if (m_use_big_batch && i == 1)
169 p_threadctxs.emplace_back(p_roi, p_num_of_feats * p_num_scales, 1, p_num_scales);
171 p_threadctxs.emplace_back(p_roi, p_num_of_feats, p_scales[i], 1);
174 p_current_scale = 1.;
176 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
177 double max_size_ratio =
178 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
179 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
180 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
181 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
183 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
184 std::cout << "init: win size. " << p_windows_size.width << " " << p_windows_size.height << std::endl;
185 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
187 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
189 fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales, m_use_big_batch);
190 fft.set_window(cosine_window_function(p_roi.width, p_roi.height));
192 // window weights, i.e. labels
194 gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height), p_yf,
195 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_threadctxs.front().stream);
198 // obtain a sub-window for training initial model
199 p_threadctxs.front().patch_feats.clear();
200 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size.width, p_windows_size.height,
201 p_threadctxs.front());
202 fft.forward_window(p_threadctxs.front().patch_feats, p_model_xf, p_threadctxs.front().fw_all,
203 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr,
204 p_threadctxs.front().stream);
205 DEBUG_PRINTM(p_model_xf);
206 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
207 p_threadctxs.front().model_xf = p_model_xf;
208 p_threadctxs.front().model_xf.set_stream(p_threadctxs.front().stream);
209 p_yf.set_stream(p_threadctxs.front().stream);
210 p_model_xf.set_stream(p_threadctxs.front().stream);
211 p_xf.set_stream(p_threadctxs.front().stream);
214 if (m_use_linearkernel) {
215 ComplexMat xfconj = p_model_xf.conj();
216 p_model_alphaf_num = xfconj.mul(p_yf);
217 p_model_alphaf_den = (p_model_xf * xfconj);
219 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
220 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
221 gaussian_correlation(p_threadctxs.front(), p_threadctxs.front().model_xf, p_threadctxs.front().model_xf,
222 p_kernel_sigma, true);
224 gaussian_correlation(p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
226 DEBUG_PRINTM(p_threadctxs.front().kf);
227 p_model_alphaf_num = p_yf * p_threadctxs.front().kf;
228 DEBUG_PRINTM(p_model_alphaf_num);
229 p_model_alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
230 DEBUG_PRINTM(p_model_alphaf_den);
232 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
233 DEBUG_PRINTM(p_model_alphaf);
234 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
236 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
237 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
238 it->model_xf = p_model_xf;
239 it->model_xf.set_stream(it->stream);
240 it->model_alphaf = p_model_alphaf;
241 it->model_alphaf.set_stream(it->stream);
246 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
248 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
251 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
253 if (p_resize_image) {
255 tmp.scale(p_downscale_factor);
258 } else if (p_fit_to_pw2) {
260 tmp.scale_x(p_scale_factor_x);
261 tmp.scale_y(p_scale_factor_y);
270 BBox_c KCF_Tracker::getBBox()
273 tmp.w *= p_current_scale;
274 tmp.h *= p_current_scale;
276 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
278 tmp.scale_x(1 / p_scale_factor_x);
279 tmp.scale_y(1 / p_scale_factor_y);
285 double KCF_Tracker::getFilterResponse() const
287 return this->max_response;
290 void KCF_Tracker::track(cv::Mat &img)
292 if (m_debug) std::cout << "NEW FRAME" << '\n';
293 cv::Mat input_gray, input_rgb = img.clone();
294 if (img.channels() == 3) {
295 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
296 input_gray.convertTo(input_gray, CV_32FC1);
298 img.convertTo(input_gray, CV_32FC1);
300 // don't need too large image
301 if (p_resize_image) {
302 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
303 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
304 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
305 fabs(p_scale_factor_y - 1) > p_floating_error) {
306 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
307 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
308 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
310 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
311 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
316 ThreadCtx *max = nullptr;
317 cv::Point2i *max_response_pt = nullptr;
318 cv::Mat *max_response_map = nullptr;
321 for (auto &it : p_threadctxs)
322 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
323 scale_track(it, input_rgb, input_gray);
325 for (auto const &it : p_threadctxs)
329 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
330 NORMAL_OMP_PARALLEL_FOR
331 for (uint i = 0; i < p_threadctxs.size(); ++i)
332 scale_track(p_threadctxs[i], input_rgb, input_gray);
336 for (auto &it : p_threadctxs) {
337 if (it.max_response > max_response) {
338 max_response = it.max_response;
339 max_response_pt = &it.max_loc;
340 max_response_map = &it.response;
345 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
346 for (uint j = 0; j < p_scales.size(); ++j) {
347 if (p_threadctxs[0].max_responses[j] > max_response) {
348 max_response = p_threadctxs[0].max_responses[j];
349 max_response_pt = &p_threadctxs[0].max_locs[j];
350 max_response_map = &p_threadctxs[0].response_maps[j];
351 max = &p_threadctxs[0];
356 DEBUG_PRINTM(*max_response_map);
357 DEBUG_PRINT(*max_response_pt);
359 // sub pixel quadratic interpolation from neighbours
360 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
361 max_response_pt->y = max_response_pt->y - max_response_map->rows;
362 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
363 max_response_pt->x = max_response_pt->x - max_response_map->cols;
365 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
366 DEBUG_PRINT(new_location);
368 if (m_use_subpixel_localization)
369 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
370 DEBUG_PRINT(new_location);
372 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
373 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
375 if (p_pose.cx < 0) p_pose.cx = 0;
376 if (p_pose.cx > (img.cols * p_scale_factor_x) - 1) p_pose.cx = (img.cols * p_scale_factor_x) - 1;
377 if (p_pose.cy < 0) p_pose.cy = 0;
378 if (p_pose.cy > (img.rows * p_scale_factor_y) - 1) p_pose.cy = (img.rows * p_scale_factor_y) - 1;
380 if (p_pose.cx < 0) p_pose.cx = 0;
381 if (p_pose.cx > img.cols - 1) p_pose.cx = img.cols - 1;
382 if (p_pose.cy < 0) p_pose.cy = 0;
383 if (p_pose.cy > img.rows - 1) p_pose.cy = img.rows - 1;
386 // sub grid scale interpolation
387 if (m_use_subgrid_scale) {
388 auto it = std::find_if(p_threadctxs.begin(), p_threadctxs.end(), [max](ThreadCtx &ctx) { return &ctx == max; });
389 p_current_scale *= sub_grid_scale(std::distance(p_threadctxs.begin(), it));
391 p_current_scale *= max->scale;
395 if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
396 if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
398 // obtain a subwindow for training at newly estimated target position
399 p_threadctxs.front().patch_feats.clear();
400 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size.width, p_windows_size.height,
401 p_threadctxs.front(), p_current_scale);
402 fft.forward_window(p_threadctxs.front().patch_feats, p_xf, p_threadctxs.front().fw_all,
403 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr, p_threadctxs.front().stream);
405 // subsequent frames, interpolate model
406 p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
408 ComplexMat alphaf_num, alphaf_den;
410 if (m_use_linearkernel) {
411 ComplexMat xfconj = p_xf.conj();
412 alphaf_num = xfconj.mul(p_yf);
413 alphaf_den = (p_xf * xfconj);
415 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
416 gaussian_correlation(p_threadctxs.front(), p_xf, p_xf, p_kernel_sigma,
418 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
419 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
420 alphaf_num = p_yf * p_threadctxs.front().kf;
421 alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
424 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
425 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
426 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
428 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
429 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
430 it->model_xf = p_model_xf;
431 it->model_xf.set_stream(it->stream);
432 it->model_alphaf = p_model_alphaf;
433 it->model_alphaf.set_stream(it->stream);
438 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray)
440 if (m_use_big_batch) {
441 vars.patch_feats.clear();
442 BIG_BATCH_OMP_PARALLEL_FOR
443 for (uint i = 0; i < p_num_scales; ++i) {
444 get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size.width,
445 this->p_windows_size.height, vars, this->p_current_scale * this->p_scales[i]);
448 get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size.width,
449 this->p_windows_size.height, vars, this->p_current_scale * vars.scale);
452 fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
454 DEBUG_PRINTM(vars.zf);
456 if (m_use_linearkernel) {
457 vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
458 : (p_model_alphaf * vars.zf).sum_over_channels();
459 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
461 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
462 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
463 vars.kzf = vars.model_alphaf * vars.kzf;
465 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
466 DEBUG_PRINTM(this->p_model_alphaf);
467 DEBUG_PRINTM(vars.kzf);
468 vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
470 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
473 DEBUG_PRINTM(vars.response);
475 /* target location is at the maximum response. we must take into
476 account the fact that, if the target doesn't move, the peak
477 will appear at the top-left corner, not at the center (this is
478 discussed in the paper). the responses wrap around cyclically. */
479 if (m_use_big_batch) {
480 cv::split(vars.response, vars.response_maps);
482 for (size_t i = 0; i < p_scales.size(); ++i) {
483 double min_val, max_val;
484 cv::Point2i min_loc, max_loc;
485 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
486 DEBUG_PRINT(max_loc);
487 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
488 vars.max_responses[i] = max_val * weight;
489 vars.max_locs[i] = max_loc;
494 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
496 DEBUG_PRINT(vars.max_loc);
498 double weight = vars.scale < 1. ? vars.scale : 1. / vars.scale;
499 vars.max_response = vars.max_val * weight;
504 // ****************************************************************************
506 void KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y,
507 ThreadCtx &vars, double scale)
509 int size_x_scaled = floor(size_x * scale);
510 int size_y_scaled = floor(size_y * scale);
512 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
513 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
515 // resize to default size
517 // if we downsample use INTER_AREA interpolation
518 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
520 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
523 // get hog(Histogram of Oriented Gradients) features
524 vars.patch_feats = FHoG::extract(patch_gray, 2, p_cell_size, 9);
526 // get color rgb features (simple r,g,b channels)
527 std::vector<cv::Mat> color_feat;
528 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
529 // resize to default size
531 // if we downsample use INTER_AREA interpolation
532 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
535 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
540 if (m_use_color && input_rgb.channels() == 3) {
541 // use rgb color space
542 cv::Mat patch_rgb_norm;
543 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
544 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
545 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
546 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
547 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
548 cv::split(patch_rgb_norm, rgb);
549 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
552 if (m_use_cnfeat && input_rgb.channels() == 3) {
553 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
554 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
556 BIG_BATCH_OMP_ORDERED
557 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
561 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
563 cv::Mat labels(dim2, dim1, CV_32FC1);
564 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
565 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
567 double sigma_s = sigma * sigma;
569 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
570 float *row_ptr = labels.ptr<float>(j);
572 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
573 row_ptr[i] = float(std::exp(-0.5 * (y_s + x * x) / sigma_s)); //-1/2*e^((y^2+x^2)/sigma^2)
577 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
579 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
580 tmp.copyTo(p_rot_labels);
582 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
585 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
586 // sanity check, 1 at top left corner
587 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
593 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
595 cv::Mat rot_patch(patch.size(), CV_32FC1);
596 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
598 // circular rotate x-axis
600 // move part that does not rotate over the edge
601 cv::Range orig_range(-x_rot, patch.cols);
602 cv::Range rot_range(0, patch.cols - (-x_rot));
603 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
606 orig_range = cv::Range(0, -x_rot);
607 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
608 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
609 } else if (x_rot > 0) {
610 // move part that does not rotate over the edge
611 cv::Range orig_range(0, patch.cols - x_rot);
612 cv::Range rot_range(x_rot, patch.cols);
613 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
616 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
617 rot_range = cv::Range(0, x_rot);
618 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
619 } else { // zero rotation
620 // move part that does not rotate over the edge
621 cv::Range orig_range(0, patch.cols);
622 cv::Range rot_range(0, patch.cols);
623 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
626 // circular rotate y-axis
628 // move part that does not rotate over the edge
629 cv::Range orig_range(-y_rot, patch.rows);
630 cv::Range rot_range(0, patch.rows - (-y_rot));
631 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
634 orig_range = cv::Range(0, -y_rot);
635 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
636 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
637 } else if (y_rot > 0) {
638 // move part that does not rotate over the edge
639 cv::Range orig_range(0, patch.rows - y_rot);
640 cv::Range rot_range(y_rot, patch.rows);
641 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
644 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
645 rot_range = cv::Range(0, y_rot);
646 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
647 } else { // zero rotation
648 // move part that does not rotate over the edge
649 cv::Range orig_range(0, patch.rows);
650 cv::Range rot_range(0, patch.rows);
651 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
657 // hann window actually (Power-of-cosine windows)
658 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
660 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
661 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
662 for (int i = 0; i < dim1; ++i)
663 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
664 N_inv = 1. / (static_cast<double>(dim2) - 1.);
665 for (int i = 0; i < dim2; ++i)
666 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
667 cv::Mat ret = m2 * m1;
671 // Returns sub-window of image input centered at [cx, cy] coordinates),
672 // with size [width, height]. If any pixels are outside of the image,
673 // they will replicate the values at the borders.
674 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
678 int x1 = cx - width / 2;
679 int y1 = cy - height / 2;
680 int x2 = cx + width / 2;
681 int y2 = cy + height / 2;
684 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
685 patch.create(height, width, input.type());
686 patch.setTo(double(0.f));
690 int top = 0, bottom = 0, left = 0, right = 0;
692 // fit to image coordinates, set border extensions;
701 if (x2 >= input.cols) {
702 right = x2 - input.cols + width % 2;
707 if (y2 >= input.rows) {
708 bottom = y2 - input.rows + height % 2;
713 if (x2 - x1 == 0 || y2 - y1 == 0)
714 patch = cv::Mat::zeros(height, width, CV_32FC1);
716 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
717 cv::BORDER_REPLICATE);
718 // imshow( "copyMakeBorder", patch);
723 assert(patch.cols == width && patch.rows == height);
728 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
729 double sigma, bool auto_correlation)
732 xf.sqr_norm(vars.xf_sqr_norm.deviceMem());
733 if (!auto_correlation) yf.sqr_norm(vars.yf_sqr_norm.deviceMem());
735 xf.sqr_norm(vars.xf_sqr_norm.hostMem());
736 if (auto_correlation) {
737 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
739 yf.sqr_norm(vars.yf_sqr_norm.hostMem());
742 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
743 DEBUG_PRINTM(vars.xyf);
744 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
746 if (auto_correlation)
747 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(),
748 sigma, xf.n_channels, xf.n_scales, p_roi.height, p_roi.width, vars.stream);
750 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.yf_sqr_norm.deviceMem(),
751 sigma, xf.n_channels, xf.n_scales, p_roi.height, p_roi.width, vars.stream);
753 // ifft2 and sum over 3rd dimension, we dont care about individual channels
754 DEBUG_PRINTM(vars.ifft2_res);
756 if (xf.channels() != p_num_scales * p_num_of_feats)
757 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
759 xy_sum.create(vars.ifft2_res.size(), CV_32FC(p_scales.size()));
761 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
762 float *row_ptr = vars.ifft2_res.ptr<float>(y);
763 float *row_ptr_sum = xy_sum.ptr<float>(y);
764 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
765 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
766 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
767 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
768 (row_ptr + x * vars.ifft2_res.channels() +
769 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
774 DEBUG_PRINTM(xy_sum);
776 std::vector<cv::Mat> scales;
777 cv::split(xy_sum, scales);
779 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
780 for (uint i = 0; i < xf.n_scales; ++i) {
781 cv::Mat in_roi(vars.in_all, cv::Rect(0, i * scales[0].rows, scales[0].cols, scales[0].rows));
783 -1. / (sigma * sigma) *
784 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
786 DEBUG_PRINTM(in_roi);
789 DEBUG_PRINTM(vars.in_all);
790 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
795 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
799 if (x < 0) x = response.cols + x;
800 if (y < 0) y = response.rows + y;
801 if (x >= response.cols) x = x - response.cols;
802 if (y >= response.rows) y = y - response.rows;
804 return response.at<float>(y, x);
807 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
809 // find neighbourhood of max_loc (response is circular)
813 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);
814 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
815 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);
818 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
819 cv::Mat A = (cv::Mat_<float>(9, 6) <<
820 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
821 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
822 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
823 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
824 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
825 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
826 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
827 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
828 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);
829 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
830 get_response_circular(p1, response),
831 get_response_circular(p2, response),
832 get_response_circular(p3, response),
833 get_response_circular(p4, response),
834 get_response_circular(p5, response),
835 get_response_circular(p6, response),
836 get_response_circular(p7, response),
837 get_response_circular(p8, response),
838 get_response_circular(max_loc, response));
841 cv::solve(A, fval, x, cv::DECOMP_SVD);
843 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);
845 cv::Point2f sub_peak(max_loc.x, max_loc.y);
846 if (b > 0 || b < 0) {
847 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
848 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
854 double KCF_Tracker::sub_grid_scale(uint index)
857 if (index >= p_scales.size()) {
858 // interpolate from all values
859 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
860 A.create(p_scales.size(), 3, CV_32FC1);
861 fval.create(p_scales.size(), 1, CV_32FC1);
862 for (size_t i = 0; i < p_scales.size(); ++i) {
863 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
864 A.at<float>(i, 1) = float(p_scales[i]);
865 A.at<float>(i, 2) = 1;
866 fval.at<float>(i) = m_use_big_batch ? p_threadctxs.back().max_responses[i] : p_threadctxs[i].max_response;
869 // only from neighbours
870 if (index == 0 || index == p_scales.size() - 1)
871 return p_scales[index];
873 A = (cv::Mat_<float>(3, 3) <<
874 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
875 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
876 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
877 fval = (cv::Mat_<float>(3, 1) <<
878 (m_use_big_batch ? p_threadctxs.back().max_responses[index - 1] : p_threadctxs[index - 1].max_response),
879 (m_use_big_batch ? p_threadctxs.back().max_responses[index + 0] : p_threadctxs[index + 0].max_response),
880 (m_use_big_batch ? p_threadctxs.back().max_responses[index + 1] : p_threadctxs[index + 1].max_response));
884 cv::solve(A, fval, x, cv::DECOMP_SVD);
885 float a = x.at<float>(0), b = x.at<float>(1);
886 double scale = p_scales[index];
888 scale = -b / (2 * a);