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 p_model_xf.create(p_roi.height, p_roi.width / 2 + 1, p_num_of_feats);
159 p_yf.create(p_roi.height, p_roi.width / 2 + 1, 1);
160 p_xf.create(p_roi.height, p_roi.width / 2 + 1, p_num_of_feats);
162 p_model_xf.create(p_roi.height, p_roi.width, p_num_of_feats);
163 p_yf.create(p_roi.height, p_roi.width, 1);
164 p_xf.create(p_roi.height, p_roi.width, p_num_of_feats);
167 int max = m_use_big_batch ? 2 : p_num_scales;
168 for (int i = 0; i < max; ++i) {
169 if (m_use_big_batch && i == 1)
170 p_threadctxs.emplace_back(p_roi, p_num_of_feats * p_num_scales, 1, p_num_scales);
172 p_threadctxs.emplace_back(p_roi, p_num_of_feats, p_scales[i], 1);
175 p_current_scale = 1.;
177 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
178 double max_size_ratio =
179 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
180 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
181 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
182 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
184 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
185 std::cout << "init: win size. " << p_windows_size.width << " " << p_windows_size.height << std::endl;
186 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
188 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
190 fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales, m_use_big_batch);
191 fft.set_window(cosine_window_function(p_roi.width, p_roi.height));
193 // window weights, i.e. labels
195 gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height), p_yf,
196 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_threadctxs.front().stream);
199 // obtain a sub-window for training initial model
200 p_threadctxs.front().patch_feats.clear();
201 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size.width, p_windows_size.height,
202 p_threadctxs.front());
203 fft.forward_window(p_threadctxs.front().patch_feats, p_model_xf, p_threadctxs.front().fw_all,
204 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr,
205 p_threadctxs.front().stream);
206 DEBUG_PRINTM(p_model_xf);
207 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
208 p_threadctxs.front().model_xf = p_model_xf;
209 p_threadctxs.front().model_xf.set_stream(p_threadctxs.front().stream);
210 p_yf.set_stream(p_threadctxs.front().stream);
211 p_model_xf.set_stream(p_threadctxs.front().stream);
212 p_xf.set_stream(p_threadctxs.front().stream);
215 if (m_use_linearkernel) {
216 ComplexMat xfconj = p_model_xf.conj();
217 p_model_alphaf_num = xfconj.mul(p_yf);
218 p_model_alphaf_den = (p_model_xf * xfconj);
220 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
221 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
222 gaussian_correlation(p_threadctxs.front(), p_threadctxs.front().model_xf, p_threadctxs.front().model_xf,
223 p_kernel_sigma, true);
225 gaussian_correlation(p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
227 DEBUG_PRINTM(p_threadctxs.front().kf);
228 p_model_alphaf_num = p_yf * p_threadctxs.front().kf;
229 DEBUG_PRINTM(p_model_alphaf_num);
230 p_model_alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
231 DEBUG_PRINTM(p_model_alphaf_den);
233 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
234 DEBUG_PRINTM(p_model_alphaf);
235 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
237 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
238 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
239 it->model_xf = p_model_xf;
240 it->model_xf.set_stream(it->stream);
241 it->model_alphaf = p_model_alphaf;
242 it->model_alphaf.set_stream(it->stream);
247 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
249 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
252 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
254 if (p_resize_image) {
256 tmp.scale(p_downscale_factor);
259 } else if (p_fit_to_pw2) {
261 tmp.scale_x(p_scale_factor_x);
262 tmp.scale_y(p_scale_factor_y);
271 BBox_c KCF_Tracker::getBBox()
274 tmp.w *= p_current_scale;
275 tmp.h *= p_current_scale;
277 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
279 tmp.scale_x(1 / p_scale_factor_x);
280 tmp.scale_y(1 / p_scale_factor_y);
286 double KCF_Tracker::getFilterResponse() const
288 return this->max_response;
291 void KCF_Tracker::track(cv::Mat &img)
293 if (m_debug) std::cout << "NEW FRAME" << '\n';
294 cv::Mat input_gray, input_rgb = img.clone();
295 if (img.channels() == 3) {
296 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
297 input_gray.convertTo(input_gray, CV_32FC1);
299 img.convertTo(input_gray, CV_32FC1);
301 // don't need too large image
302 if (p_resize_image) {
303 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
304 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
305 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
306 fabs(p_scale_factor_y - 1) > p_floating_error) {
307 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
308 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
309 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
311 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
312 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
317 ThreadCtx *max = nullptr;
318 cv::Point2i *max_response_pt = nullptr;
319 cv::Mat *max_response_map = nullptr;
322 for (auto &it : p_threadctxs)
323 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
324 scale_track(it, input_rgb, input_gray);
326 for (auto const &it : p_threadctxs)
330 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
331 NORMAL_OMP_PARALLEL_FOR
332 for (uint i = 0; i < p_threadctxs.size(); ++i)
333 scale_track(p_threadctxs[i], input_rgb, input_gray);
337 for (auto &it : p_threadctxs) {
338 if (it.max_response > max_response) {
339 max_response = it.max_response;
340 max_response_pt = &it.max_loc;
341 max_response_map = &it.response;
346 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
347 for (uint j = 0; j < p_scales.size(); ++j) {
348 if (p_threadctxs[0].max_responses[j] > max_response) {
349 max_response = p_threadctxs[0].max_responses[j];
350 max_response_pt = &p_threadctxs[0].max_locs[j];
351 max_response_map = &p_threadctxs[0].response_maps[j];
352 max = &p_threadctxs[0];
357 DEBUG_PRINTM(*max_response_map);
358 DEBUG_PRINT(*max_response_pt);
360 // sub pixel quadratic interpolation from neighbours
361 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
362 max_response_pt->y = max_response_pt->y - max_response_map->rows;
363 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
364 max_response_pt->x = max_response_pt->x - max_response_map->cols;
366 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
367 DEBUG_PRINT(new_location);
369 if (m_use_subpixel_localization)
370 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
371 DEBUG_PRINT(new_location);
373 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
374 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
376 if (p_pose.cx < 0) p_pose.cx = 0;
377 if (p_pose.cx > (img.cols * p_scale_factor_x) - 1) p_pose.cx = (img.cols * p_scale_factor_x) - 1;
378 if (p_pose.cy < 0) p_pose.cy = 0;
379 if (p_pose.cy > (img.rows * p_scale_factor_y) - 1) p_pose.cy = (img.rows * p_scale_factor_y) - 1;
381 if (p_pose.cx < 0) p_pose.cx = 0;
382 if (p_pose.cx > img.cols - 1) p_pose.cx = img.cols - 1;
383 if (p_pose.cy < 0) p_pose.cy = 0;
384 if (p_pose.cy > img.rows - 1) p_pose.cy = img.rows - 1;
387 // sub grid scale interpolation
388 if (m_use_subgrid_scale) {
389 auto it = std::find_if(p_threadctxs.begin(), p_threadctxs.end(), [max](ThreadCtx &ctx) { return &ctx == max; });
390 p_current_scale *= sub_grid_scale(std::distance(p_threadctxs.begin(), it));
392 p_current_scale *= max->scale;
396 if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
397 if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
399 // obtain a subwindow for training at newly estimated target position
400 p_threadctxs.front().patch_feats.clear();
401 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size.width, p_windows_size.height,
402 p_threadctxs.front(), p_current_scale);
403 fft.forward_window(p_threadctxs.front().patch_feats, p_xf, p_threadctxs.front().fw_all,
404 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr, p_threadctxs.front().stream);
406 // subsequent frames, interpolate model
407 p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
409 ComplexMat alphaf_num, alphaf_den;
411 if (m_use_linearkernel) {
412 ComplexMat xfconj = p_xf.conj();
413 alphaf_num = xfconj.mul(p_yf);
414 alphaf_den = (p_xf * xfconj);
416 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
417 gaussian_correlation(p_threadctxs.front(), p_xf, p_xf, p_kernel_sigma,
419 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
420 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
421 alphaf_num = p_yf * p_threadctxs.front().kf;
422 alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
425 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
426 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
427 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
429 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
430 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
431 it->model_xf = p_model_xf;
432 it->model_xf.set_stream(it->stream);
433 it->model_alphaf = p_model_alphaf;
434 it->model_alphaf.set_stream(it->stream);
439 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray)
441 if (m_use_big_batch) {
442 vars.patch_feats.clear();
443 BIG_BATCH_OMP_PARALLEL_FOR
444 for (uint i = 0; i < p_num_scales; ++i) {
445 get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size.width,
446 this->p_windows_size.height, vars, this->p_current_scale * this->p_scales[i]);
449 vars.patch_feats.clear();
450 get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size.width,
451 this->p_windows_size.height, vars, this->p_current_scale * vars.scale);
454 fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
456 DEBUG_PRINTM(vars.zf);
458 if (m_use_linearkernel) {
459 vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
460 : (p_model_alphaf * vars.zf).sum_over_channels();
461 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
463 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
464 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
465 vars.kzf = vars.model_alphaf * vars.kzf;
467 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
468 DEBUG_PRINTM(this->p_model_alphaf);
469 DEBUG_PRINTM(vars.kzf);
470 vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
472 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
475 DEBUG_PRINTM(vars.response);
477 /* target location is at the maximum response. we must take into
478 account the fact that, if the target doesn't move, the peak
479 will appear at the top-left corner, not at the center (this is
480 discussed in the paper). the responses wrap around cyclically. */
481 if (m_use_big_batch) {
482 cv::split(vars.response, vars.response_maps);
484 for (size_t i = 0; i < p_scales.size(); ++i) {
485 double min_val, max_val;
486 cv::Point2i min_loc, max_loc;
487 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
488 DEBUG_PRINT(max_loc);
489 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
490 vars.max_responses[i] = max_val * weight;
491 vars.max_locs[i] = max_loc;
496 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
498 DEBUG_PRINT(vars.max_loc);
500 double weight = vars.scale < 1. ? vars.scale : 1. / vars.scale;
501 vars.max_response = vars.max_val * weight;
506 // ****************************************************************************
508 void KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y,
509 ThreadCtx &vars, double scale)
511 int size_x_scaled = floor(size_x * scale);
512 int size_y_scaled = floor(size_y * scale);
514 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
515 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
517 // resize to default size
519 // if we downsample use INTER_AREA interpolation
520 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
522 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
525 // get hog(Histogram of Oriented Gradients) features
526 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
528 // get color rgb features (simple r,g,b channels)
529 std::vector<cv::Mat> color_feat;
530 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
531 // resize to default size
533 // if we downsample use INTER_AREA interpolation
534 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
537 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
542 if (m_use_color && input_rgb.channels() == 3) {
543 // use rgb color space
544 cv::Mat patch_rgb_norm;
545 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
546 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
547 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
548 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
549 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
550 cv::split(patch_rgb_norm, rgb);
551 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
554 if (m_use_cnfeat && input_rgb.channels() == 3) {
555 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
556 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
558 BIG_BATCH_OMP_ORDERED
559 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
563 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
565 cv::Mat labels(dim2, dim1, CV_32FC1);
566 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
567 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
569 double sigma_s = sigma * sigma;
571 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
572 float *row_ptr = labels.ptr<float>(j);
574 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
575 row_ptr[i] = float(std::exp(-0.5 * (y_s + x * x) / sigma_s)); //-1/2*e^((y^2+x^2)/sigma^2)
579 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
581 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
582 tmp.copyTo(p_rot_labels);
584 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
587 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
588 // sanity check, 1 at top left corner
589 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
595 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
597 cv::Mat rot_patch(patch.size(), CV_32FC1);
598 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
600 // circular rotate x-axis
602 // move part that does not rotate over the edge
603 cv::Range orig_range(-x_rot, patch.cols);
604 cv::Range rot_range(0, patch.cols - (-x_rot));
605 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
608 orig_range = cv::Range(0, -x_rot);
609 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
610 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
611 } else if (x_rot > 0) {
612 // move part that does not rotate over the edge
613 cv::Range orig_range(0, patch.cols - x_rot);
614 cv::Range rot_range(x_rot, patch.cols);
615 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
618 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
619 rot_range = cv::Range(0, x_rot);
620 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
621 } else { // zero rotation
622 // move part that does not rotate over the edge
623 cv::Range orig_range(0, patch.cols);
624 cv::Range rot_range(0, patch.cols);
625 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
628 // circular rotate y-axis
630 // move part that does not rotate over the edge
631 cv::Range orig_range(-y_rot, patch.rows);
632 cv::Range rot_range(0, patch.rows - (-y_rot));
633 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
636 orig_range = cv::Range(0, -y_rot);
637 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
638 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
639 } else if (y_rot > 0) {
640 // move part that does not rotate over the edge
641 cv::Range orig_range(0, patch.rows - y_rot);
642 cv::Range rot_range(y_rot, patch.rows);
643 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
646 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
647 rot_range = cv::Range(0, y_rot);
648 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
649 } else { // zero rotation
650 // move part that does not rotate over the edge
651 cv::Range orig_range(0, patch.rows);
652 cv::Range rot_range(0, patch.rows);
653 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
659 // hann window actually (Power-of-cosine windows)
660 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
662 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
663 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
664 for (int i = 0; i < dim1; ++i)
665 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
666 N_inv = 1. / (static_cast<double>(dim2) - 1.);
667 for (int i = 0; i < dim2; ++i)
668 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
669 cv::Mat ret = m2 * m1;
673 // Returns sub-window of image input centered at [cx, cy] coordinates),
674 // with size [width, height]. If any pixels are outside of the image,
675 // they will replicate the values at the borders.
676 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
680 int x1 = cx - width / 2;
681 int y1 = cy - height / 2;
682 int x2 = cx + width / 2;
683 int y2 = cy + height / 2;
686 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
687 patch.create(height, width, input.type());
688 patch.setTo(double(0.f));
692 int top = 0, bottom = 0, left = 0, right = 0;
694 // fit to image coordinates, set border extensions;
703 if (x2 >= input.cols) {
704 right = x2 - input.cols + width % 2;
709 if (y2 >= input.rows) {
710 bottom = y2 - input.rows + height % 2;
715 if (x2 - x1 == 0 || y2 - y1 == 0)
716 patch = cv::Mat::zeros(height, width, CV_32FC1);
718 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
719 cv::BORDER_REPLICATE);
720 // imshow( "copyMakeBorder", patch);
725 assert(patch.cols == width && patch.rows == height);
730 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
731 double sigma, bool auto_correlation)
734 xf.sqr_norm(vars.xf_sqr_norm.deviceMem());
735 if (!auto_correlation) yf.sqr_norm(vars.yf_sqr_norm.deviceMem());
737 xf.sqr_norm(vars.xf_sqr_norm.hostMem());
738 if (auto_correlation) {
739 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
741 yf.sqr_norm(vars.yf_sqr_norm.hostMem());
744 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
745 DEBUG_PRINTM(vars.xyf);
746 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
748 if (auto_correlation)
749 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(),
750 sigma, xf.n_channels, xf.n_scales, p_roi.height, p_roi.width, vars.stream);
752 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.yf_sqr_norm.deviceMem(),
753 sigma, xf.n_channels, xf.n_scales, p_roi.height, p_roi.width, vars.stream);
755 // ifft2 and sum over 3rd dimension, we dont care about individual channels
756 DEBUG_PRINTM(vars.ifft2_res);
758 if (xf.channels() != p_num_scales * p_num_of_feats)
759 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
761 xy_sum.create(vars.ifft2_res.size(), CV_32FC(p_scales.size()));
763 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
764 float *row_ptr = vars.ifft2_res.ptr<float>(y);
765 float *row_ptr_sum = xy_sum.ptr<float>(y);
766 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
767 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
768 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
769 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
770 (row_ptr + x * vars.ifft2_res.channels() +
771 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
776 DEBUG_PRINTM(xy_sum);
778 std::vector<cv::Mat> scales;
779 cv::split(xy_sum, scales);
781 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
782 for (uint i = 0; i < xf.n_scales; ++i) {
783 cv::Mat in_roi(vars.in_all, cv::Rect(0, i * scales[0].rows, scales[0].cols, scales[0].rows));
785 -1. / (sigma * sigma) *
786 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
788 DEBUG_PRINTM(in_roi);
791 DEBUG_PRINTM(vars.in_all);
792 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
797 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
801 if (x < 0) x = response.cols + x;
802 if (y < 0) y = response.rows + y;
803 if (x >= response.cols) x = x - response.cols;
804 if (y >= response.rows) y = y - response.rows;
806 return response.at<float>(y, x);
809 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
811 // find neighbourhood of max_loc (response is circular)
815 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);
816 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
817 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);
820 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
821 cv::Mat A = (cv::Mat_<float>(9, 6) <<
822 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
823 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
824 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
825 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
826 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
827 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
828 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
829 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
830 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);
831 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
832 get_response_circular(p1, response),
833 get_response_circular(p2, response),
834 get_response_circular(p3, response),
835 get_response_circular(p4, response),
836 get_response_circular(p5, response),
837 get_response_circular(p6, response),
838 get_response_circular(p7, response),
839 get_response_circular(p8, response),
840 get_response_circular(max_loc, response));
843 cv::solve(A, fval, x, cv::DECOMP_SVD);
845 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);
847 cv::Point2f sub_peak(max_loc.x, max_loc.y);
848 if (b > 0 || b < 0) {
849 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
850 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
856 double KCF_Tracker::sub_grid_scale(uint index)
859 if (index >= p_scales.size()) {
860 // interpolate from all values
861 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
862 A.create(p_scales.size(), 3, CV_32FC1);
863 fval.create(p_scales.size(), 1, CV_32FC1);
864 for (size_t i = 0; i < p_scales.size(); ++i) {
865 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
866 A.at<float>(i, 1) = float(p_scales[i]);
867 A.at<float>(i, 2) = 1;
868 fval.at<float>(i) = m_use_big_batch ? p_threadctxs.back().max_responses[i] : p_threadctxs[i].max_response;
871 // only from neighbours
872 if (index == 0 || index == p_scales.size() - 1)
873 return p_scales[index];
875 A = (cv::Mat_<float>(3, 3) <<
876 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
877 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
878 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
879 fval = (cv::Mat_<float>(3, 1) <<
880 (m_use_big_batch ? p_threadctxs.back().max_responses[index - 1] : p_threadctxs[index - 1].max_response),
881 (m_use_big_batch ? p_threadctxs.back().max_responses[index + 0] : p_threadctxs[index + 0].max_response),
882 (m_use_big_batch ? p_threadctxs.back().max_responses[index + 1] : p_threadctxs[index + 1].max_response));
886 cv::solve(A, fval, x, cv::DECOMP_SVD);
887 float a = x.at<float>(0), b = x.at<float>(1);
888 double scale = p_scales[index];
890 scale = -b / (2 * a);