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; \
32 T clamp(const T& n, const T& lower, const T& upper)
34 return std::max(lower, std::min(n, upper));
38 void clamp2(T& n, const T& lower, const T& upper)
40 n = std::max(lower, std::min(n, upper));
43 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
44 double output_sigma_factor, int cell_size)
45 : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
46 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size)
50 KCF_Tracker::KCF_Tracker() : fft(*new FFT()) {}
52 KCF_Tracker::~KCF_Tracker()
57 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
59 // check boundary, enforce min size
60 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
62 if (x2 > img.cols - 1) x2 = img.cols - 1;
64 if (y2 > img.rows - 1) y2 = img.rows - 1;
66 if (x2 - x1 < 2 * p_cell_size) {
67 double diff = (2 * p_cell_size - x2 + x1) / 2.;
68 if (x1 - diff >= 0 && x2 + diff < img.cols) {
71 } else if (x1 - 2 * diff >= 0) {
77 if (y2 - y1 < 2 * p_cell_size) {
78 double diff = (2 * p_cell_size - y2 + y1) / 2.;
79 if (y1 - diff >= 0 && y2 + diff < img.rows) {
82 } else if (y1 - 2 * diff >= 0) {
91 p_pose.cx = x1 + p_pose.w / 2.;
92 p_pose.cy = y1 + p_pose.h / 2.;
94 cv::Mat input_gray, input_rgb = img.clone();
95 if (img.channels() == 3) {
96 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
97 input_gray.convertTo(input_gray, CV_32FC1);
99 img.convertTo(input_gray, CV_32FC1);
101 // don't need too large image
102 if (p_pose.w * p_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
103 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
104 p_resize_image = true;
105 p_pose.scale(p_downscale_factor);
106 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
107 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
108 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
109 if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
110 std::cerr << "Error: Fit size is not multiple of HOG cell size (" << p_cell_size << ")" << std::endl;
111 std::exit(EXIT_FAILURE);
113 p_scale_factor_x = (double)fit_size_x / round(p_pose.w * (1. + p_padding));
114 p_scale_factor_y = (double)fit_size_y / round(p_pose.h * (1. + p_padding));
115 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
116 << p_scale_factor_y << std::endl;
118 p_pose.scale_x(p_scale_factor_x);
119 p_pose.scale_y(p_scale_factor_y);
120 if (fabs(p_scale_factor_x - 1) > p_floating_error || fabs(p_scale_factor_y - 1) > p_floating_error) {
121 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
122 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
123 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
125 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
126 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
131 // compute win size + fit to fhog cell size
132 p_windows_size.width = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
133 p_windows_size.height = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
134 p_roi.width = p_windows_size.width / p_cell_size;
135 p_roi.height = p_windows_size.height / p_cell_size;
138 if (m_use_color) p_num_of_feats += 3;
139 if (m_use_cnfeat) p_num_of_feats += 10;
143 for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
144 p_scales.push_back(std::pow(p_scale_step, i));
146 p_scales.push_back(1.);
149 if (p_roi.height * (p_roi.width / 2 + 1) > 1024) {
150 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
151 "the window dimensions so its size is less or equal to "
152 << 1024 * p_cell_size * p_cell_size * 2 + 1
153 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
154 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
155 std::exit(EXIT_FAILURE);
158 if (m_use_linearkernel) {
159 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
160 std::exit(EXIT_FAILURE);
162 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
163 p_rot_labels_data = DynMem(p_roi.width * p_roi.height * sizeof(float));
164 p_rot_labels = cv::Mat(p_roi, CV_32FC1, p_rot_labels_data.hostMem());
166 p_xf.create(p_roi.height, p_roi.height / 2 + 1, p_num_of_feats);
169 #if defined(CUFFT) || defined(FFTW)
170 uint width = p_roi.width / 2 + 1;
172 uint width = p_roi.width;
174 p_model_xf.create(p_roi.height, width, p_num_of_feats);
175 p_yf.create(p_roi.height, width, 1);
176 p_xf.create(p_roi.height, width, p_num_of_feats);
178 int max = BIG_BATCH_MODE ? 2 : p_num_scales;
179 for (int i = 0; i < max; ++i) {
180 if (BIG_BATCH_MODE && i == 1)
181 p_threadctxs.emplace_back(p_roi, p_num_of_feats * p_num_scales, 1, p_num_scales);
183 p_threadctxs.emplace_back(p_roi, p_num_of_feats, p_scales[i], 1);
186 p_current_scale = 1.;
188 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
189 double max_size_ratio =
190 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
191 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
192 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
193 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
195 std::cout << "init: img size " << img.cols << "x" << img.rows << std::endl;
196 std::cout << "init: win size " << p_windows_size.width << "x" << p_windows_size.height << std::endl;
197 std::cout << "init: FFT size " << p_roi.width << "x" << p_roi.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 / p_cell_size;
202 fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales);
203 fft.set_window(cosine_window_function(p_roi.width, p_roi.height));
205 // window weights, i.e. labels
207 gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height), p_yf,
208 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr);
211 // obtain a sub-window for training initial model
212 std::vector<cv::Mat> patch_feats = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy,
213 p_windows_size.width, p_windows_size.height);
214 fft.forward_window(patch_feats, p_model_xf, p_threadctxs.front().fw_all,
215 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr);
216 DEBUG_PRINTM(p_model_xf);
217 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
218 p_threadctxs.front().model_xf = p_model_xf;
221 if (m_use_linearkernel) {
222 ComplexMat xfconj = p_model_xf.conj();
223 p_model_alphaf_num = xfconj.mul(p_yf);
224 p_model_alphaf_den = (p_model_xf * xfconj);
226 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
227 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
228 gaussian_correlation(p_threadctxs.front(), p_threadctxs.front().model_xf, p_threadctxs.front().model_xf,
229 p_kernel_sigma, true);
231 gaussian_correlation(p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
233 DEBUG_PRINTM(p_threadctxs.front().kf);
234 p_model_alphaf_num = p_yf * p_threadctxs.front().kf;
235 DEBUG_PRINTM(p_model_alphaf_num);
236 p_model_alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
237 DEBUG_PRINTM(p_model_alphaf_den);
239 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
240 DEBUG_PRINTM(p_model_alphaf);
241 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
243 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
244 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
245 it->model_xf = p_model_xf;
246 it->model_alphaf = p_model_alphaf;
251 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
253 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
256 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
258 if (p_resize_image) {
260 tmp.scale(p_downscale_factor);
263 } else if (p_fit_to_pw2) {
265 tmp.scale_x(p_scale_factor_x);
266 tmp.scale_y(p_scale_factor_y);
275 BBox_c KCF_Tracker::getBBox()
278 tmp.w *= p_current_scale;
279 tmp.h *= p_current_scale;
281 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
283 tmp.scale_x(1 / p_scale_factor_x);
284 tmp.scale_y(1 / p_scale_factor_y);
290 double KCF_Tracker::getFilterResponse() const
292 return this->max_response;
295 void KCF_Tracker::track(cv::Mat &img)
297 if (m_debug) std::cout << "NEW FRAME" << '\n';
298 cv::Mat input_gray, input_rgb = img.clone();
299 if (img.channels() == 3) {
300 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
301 input_gray.convertTo(input_gray, CV_32FC1);
303 img.convertTo(input_gray, CV_32FC1);
305 // don't need too large image
306 if (p_resize_image) {
307 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
308 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
309 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
310 fabs(p_scale_factor_y - 1) > p_floating_error) {
311 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
312 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
313 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
315 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
316 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
321 ThreadCtx *max = nullptr;
322 cv::Point2i *max_response_pt = nullptr;
323 cv::Mat *max_response_map = nullptr;
326 for (auto &it : p_threadctxs)
327 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
328 scale_track(it, input_rgb, input_gray);
330 for (auto const &it : p_threadctxs)
334 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
335 NORMAL_OMP_PARALLEL_FOR
336 for (uint i = 0; i < p_threadctxs.size(); ++i)
337 scale_track(p_threadctxs[i], input_rgb, input_gray);
341 for (auto &it : p_threadctxs) {
342 if (it.max_response > max_response) {
343 max_response = it.max_response;
344 max_response_pt = &it.max_loc;
345 max_response_map = &it.response;
350 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
351 for (uint j = 0; j < p_scales.size(); ++j) {
352 if (p_threadctxs[0].max_responses[j] > max_response) {
353 max_response = p_threadctxs[0].max_responses[j];
354 max_response_pt = &p_threadctxs[0].max_locs[j];
355 max_response_map = &p_threadctxs[0].response_maps[j];
356 max = &p_threadctxs[0];
361 DEBUG_PRINTM(*max_response_map);
362 DEBUG_PRINT(*max_response_pt);
364 // sub pixel quadratic interpolation from neighbours
365 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
366 max_response_pt->y = max_response_pt->y - max_response_map->rows;
367 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
368 max_response_pt->x = max_response_pt->x - max_response_map->cols;
370 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
371 DEBUG_PRINT(new_location);
373 if (m_use_subpixel_localization)
374 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
375 DEBUG_PRINT(new_location);
377 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
378 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
380 clamp2(p_pose.cx, 0.0, (img.cols * p_scale_factor_x) - 1);
381 clamp2(p_pose.cy, 0.0, (img.rows * p_scale_factor_y) - 1);
383 clamp2(p_pose.cx, 0.0, img.cols - 1.0);
384 clamp2(p_pose.cy, 0.0, img.rows - 1.0);
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;
395 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
397 ThreadCtx &ctx = p_threadctxs.front();
398 // obtain a subwindow for training at newly estimated target position
399 std::vector<cv::Mat> patch_feats = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy,
400 p_windows_size.width, p_windows_size.height,
402 fft.forward_window(patch_feats, p_xf, ctx.fw_all,
403 m_use_cuda ? ctx.data_features.deviceMem() : nullptr);
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(ctx, 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 * ctx.kf;
421 alphaf_den = ctx.kf * (ctx.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_alphaf = p_model_alphaf;
436 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray)
438 std::vector<cv::Mat> patch_feats;
439 if (BIG_BATCH_MODE) {
440 BIG_BATCH_OMP_PARALLEL_FOR
441 for (uint i = 0; i < p_num_scales; ++i) {
442 patch_feats = get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy,
443 this->p_windows_size.width, this->p_windows_size.height,
444 this->p_current_scale * this->p_scales[i]);
447 patch_feats = get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy,
448 this->p_windows_size.width, this->p_windows_size.height,
449 this->p_current_scale * vars.scale);
452 fft.forward_window(patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr);
453 DEBUG_PRINTM(vars.zf);
455 if (m_use_linearkernel) {
456 vars.kzf = BIG_BATCH_MODE ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
457 : (p_model_alphaf * vars.zf).sum_over_channels();
458 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr);
460 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
461 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
462 vars.kzf = vars.model_alphaf * vars.kzf;
464 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
465 DEBUG_PRINTM(this->p_model_alphaf);
466 DEBUG_PRINTM(vars.kzf);
467 vars.kzf = BIG_BATCH_MODE ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
469 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr);
472 DEBUG_PRINTM(vars.response);
474 /* target location is at the maximum response. we must take into
475 account the fact that, if the target doesn't move, the peak
476 will appear at the top-left corner, not at the center (this is
477 discussed in the paper). the responses wrap around cyclically. */
479 cv::split(vars.response, vars.response_maps);
481 for (size_t i = 0; i < p_scales.size(); ++i) {
482 double min_val, max_val;
483 cv::Point2i min_loc, max_loc;
484 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
485 DEBUG_PRINT(max_loc);
486 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
487 vars.max_responses[i] = max_val * weight;
488 vars.max_locs[i] = max_loc;
493 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
495 DEBUG_PRINT(vars.max_loc);
497 double weight = vars.scale < 1. ? vars.scale : 1. / vars.scale;
498 vars.max_response = vars.max_val * weight;
503 // ****************************************************************************
505 std::vector<cv::Mat> KCF_Tracker::get_features(cv::Mat & input_rgb, cv::Mat & input_gray, int cx, int cy, int size_x, int size_y, double scale)
507 int size_x_scaled = floor(size_x * scale);
508 int size_y_scaled = floor(size_y * scale);
510 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
511 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
513 // resize to default size
515 // if we downsample use INTER_AREA interpolation
516 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
518 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
521 // get hog(Histogram of Oriented Gradients) features
522 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
524 // get color rgb features (simple r,g,b channels)
525 std::vector<cv::Mat> color_feat;
526 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
527 // resize to default size
529 // if we downsample use INTER_AREA interpolation
530 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_AREA);
532 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_LINEAR);
536 if (m_use_color && input_rgb.channels() == 3) {
537 // use rgb color space
538 cv::Mat patch_rgb_norm;
539 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
540 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
541 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
542 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
543 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
544 cv::split(patch_rgb_norm, rgb);
545 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
548 if (m_use_cnfeat && input_rgb.channels() == 3) {
549 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
550 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
553 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
557 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
559 cv::Mat labels(dim2, dim1, CV_32FC1);
560 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
561 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
563 double sigma_s = sigma * sigma;
565 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
566 float *row_ptr = labels.ptr<float>(j);
568 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
569 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
573 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
575 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
576 tmp.copyTo(p_rot_labels);
578 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
581 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
582 // sanity check, 1 at top left corner
583 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
589 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
591 cv::Mat rot_patch(patch.size(), CV_32FC1);
592 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
594 // circular rotate x-axis
596 // move part that does not rotate over the edge
597 cv::Range orig_range(-x_rot, patch.cols);
598 cv::Range rot_range(0, patch.cols - (-x_rot));
599 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
602 orig_range = cv::Range(0, -x_rot);
603 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
604 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
605 } else if (x_rot > 0) {
606 // move part that does not rotate over the edge
607 cv::Range orig_range(0, patch.cols - x_rot);
608 cv::Range rot_range(x_rot, patch.cols);
609 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
612 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
613 rot_range = cv::Range(0, x_rot);
614 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
615 } else { // zero rotation
616 // move part that does not rotate over the edge
617 cv::Range orig_range(0, patch.cols);
618 cv::Range rot_range(0, patch.cols);
619 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
622 // circular rotate y-axis
624 // move part that does not rotate over the edge
625 cv::Range orig_range(-y_rot, patch.rows);
626 cv::Range rot_range(0, patch.rows - (-y_rot));
627 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
630 orig_range = cv::Range(0, -y_rot);
631 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
632 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
633 } else if (y_rot > 0) {
634 // move part that does not rotate over the edge
635 cv::Range orig_range(0, patch.rows - y_rot);
636 cv::Range rot_range(y_rot, patch.rows);
637 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
640 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
641 rot_range = cv::Range(0, y_rot);
642 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
643 } else { // zero rotation
644 // move part that does not rotate over the edge
645 cv::Range orig_range(0, patch.rows);
646 cv::Range rot_range(0, patch.rows);
647 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
653 // hann window actually (Power-of-cosine windows)
654 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
656 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
657 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
658 for (int i = 0; i < dim1; ++i)
659 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
660 N_inv = 1. / (static_cast<double>(dim2) - 1.);
661 for (int i = 0; i < dim2; ++i)
662 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
663 cv::Mat ret = m2 * m1;
667 // Returns sub-window of image input centered at [cx, cy] coordinates),
668 // with size [width, height]. If any pixels are outside of the image,
669 // they will replicate the values at the borders.
670 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
674 int x1 = cx - width / 2;
675 int y1 = cy - height / 2;
676 int x2 = cx + width / 2;
677 int y2 = cy + height / 2;
680 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
681 patch.create(height, width, input.type());
682 patch.setTo(double(0.f));
686 int top = 0, bottom = 0, left = 0, right = 0;
688 // fit to image coordinates, set border extensions;
697 if (x2 >= input.cols) {
698 right = x2 - input.cols + width % 2;
703 if (y2 >= input.rows) {
704 bottom = y2 - input.rows + height % 2;
709 if (x2 - x1 == 0 || y2 - y1 == 0)
710 patch = cv::Mat::zeros(height, width, CV_32FC1);
712 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
713 cv::BORDER_REPLICATE);
714 // imshow( "copyMakeBorder", patch);
719 assert(patch.cols == width && patch.rows == height);
724 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
725 double sigma, bool auto_correlation)
727 xf.sqr_norm(vars.xf_sqr_norm);
728 if (auto_correlation) {
729 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
731 yf.sqr_norm(vars.yf_sqr_norm);
733 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
734 DEBUG_PRINTM(vars.xyf);
735 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr);
737 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(),
738 vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(), sigma, xf.n_channels,
739 xf.n_scales, p_roi.height, p_roi.width);
741 // ifft2 and sum over 3rd dimension, we dont care about individual channels
742 DEBUG_PRINTM(vars.ifft2_res);
744 if (xf.channels() != p_num_scales * p_num_of_feats)
745 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
747 xy_sum.create(vars.ifft2_res.size(), CV_32FC(p_scales.size()));
749 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
750 float *row_ptr = vars.ifft2_res.ptr<float>(y);
751 float *row_ptr_sum = xy_sum.ptr<float>(y);
752 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
753 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
754 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
755 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
756 (row_ptr + x * vars.ifft2_res.channels() +
757 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
762 DEBUG_PRINTM(xy_sum);
764 std::vector<cv::Mat> scales;
765 cv::split(xy_sum, scales);
767 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
768 for (uint i = 0; i < xf.n_scales; ++i) {
769 cv::Mat in_roi(vars.in_all, cv::Rect(0, i * scales[0].rows, scales[0].cols, scales[0].rows));
771 -1. / (sigma * sigma) *
772 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
774 DEBUG_PRINTM(in_roi);
777 DEBUG_PRINTM(vars.in_all);
778 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr);
782 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
786 if (x < 0) x = response.cols + x;
787 if (y < 0) y = response.rows + y;
788 if (x >= response.cols) x = x - response.cols;
789 if (y >= response.rows) y = y - response.rows;
791 return response.at<float>(y, x);
794 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
796 // find neighbourhood of max_loc (response is circular)
800 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);
801 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
802 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);
805 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
806 cv::Mat A = (cv::Mat_<float>(9, 6) <<
807 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
808 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
809 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
810 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
811 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
812 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
813 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
814 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
815 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);
816 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
817 get_response_circular(p1, response),
818 get_response_circular(p2, response),
819 get_response_circular(p3, response),
820 get_response_circular(p4, response),
821 get_response_circular(p5, response),
822 get_response_circular(p6, response),
823 get_response_circular(p7, response),
824 get_response_circular(p8, response),
825 get_response_circular(max_loc, response));
828 cv::solve(A, fval, x, cv::DECOMP_SVD);
830 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);
832 cv::Point2f sub_peak(max_loc.x, max_loc.y);
833 if (b > 0 || b < 0) {
834 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
835 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
841 double KCF_Tracker::sub_grid_scale(uint index)
844 if (index >= p_scales.size()) {
845 // interpolate from all values
846 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
847 A.create(p_scales.size(), 3, CV_32FC1);
848 fval.create(p_scales.size(), 1, CV_32FC1);
849 for (size_t i = 0; i < p_scales.size(); ++i) {
850 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
851 A.at<float>(i, 1) = float(p_scales[i]);
852 A.at<float>(i, 2) = 1;
854 fval.at<float>(i) = p_threadctxs.back().max_responses[i];
856 fval.at<float>(i) = p_threadctxs[i].max_response;
860 // only from neighbours
861 if (index == 0 || index == p_scales.size() - 1)
862 return p_scales[index];
864 A = (cv::Mat_<float>(3, 3) <<
865 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
866 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
867 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
869 fval = (cv::Mat_<float>(3, 1) <<
870 p_threadctxs.back().max_responses[index - 1],
871 p_threadctxs.back().max_responses[index + 0],
872 p_threadctxs.back().max_responses[index + 1]);
874 fval = (cv::Mat_<float>(3, 1) <<
875 p_threadctxs[index - 1].max_response,
876 p_threadctxs[index + 0].max_response,
877 p_threadctxs[index + 1].max_response);
882 cv::solve(A, fval, x, cv::DECOMP_SVD);
883 float a = x.at<float>(0), b = x.at<float>(1);
884 double scale = p_scales[index];
886 scale = -b / (2 * a);