11 #include "fft_cufft.h"
14 #include "fft_opencv.h"
23 #define DEBUG_PRINT(obj) \
25 std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl; \
27 #define DEBUG_PRINTM(obj) \
29 std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl \
30 << (obj) << std::endl; \
33 #define DEBUG_PRINT(obj) if (m_debug || m_visual_debug) {std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl;}
34 #define DEBUG_PRINTM(obj) if (m_debug) {std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl << (obj) << std::endl;}
35 >>>>>>> Addded visual debug mode and also modified the rotation tracking implementation.
37 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
38 double output_sigma_factor, int cell_size)
39 : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
40 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size)
44 KCF_Tracker::KCF_Tracker() : fft(*new FFT()) {}
46 KCF_Tracker::~KCF_Tracker()
51 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
53 // check boundary, enforce min size
54 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
56 if (x2 > img.cols - 1) x2 = img.cols - 1;
58 if (y2 > img.rows - 1) y2 = img.rows - 1;
60 if (x2 - x1 < 2 * p_cell_size) {
61 double diff = (2 * p_cell_size - x2 + x1) / 2.;
62 if (x1 - diff >= 0 && x2 + diff < img.cols) {
65 } else if (x1 - 2 * diff >= 0) {
71 if (y2 - y1 < 2 * p_cell_size) {
72 double diff = (2 * p_cell_size - y2 + y1) / 2.;
73 if (y1 - diff >= 0 && y2 + diff < img.rows) {
76 } else if (y1 - 2 * diff >= 0) {
85 p_pose.cx = x1 + p_pose.w / 2.;
86 p_pose.cy = y1 + p_pose.h / 2.;
88 cv::Mat input_gray, input_rgb = img.clone();
89 if (img.channels() == 3) {
90 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
91 input_gray.convertTo(input_gray, CV_32FC1);
93 img.convertTo(input_gray, CV_32FC1);
95 // don't need too large image
96 if (p_pose.w * p_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
97 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
98 p_resize_image = true;
99 p_pose.scale(p_downscale_factor);
100 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
101 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
102 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
103 if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
104 std::cerr << "Fit size does not fit to hog cell size. The dimensions have to be divisible by HOG cell "
106 << p_cell_size << std::endl;
108 std::exit(EXIT_FAILURE);
110 double tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
111 if (fabs(tmp - fit_size_x) > p_floating_error) p_scale_factor_x = fit_size_x / tmp;
112 tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
113 if (fabs(tmp - fit_size_y) > p_floating_error) p_scale_factor_y = fit_size_y / tmp;
114 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
115 << p_scale_factor_y << std::endl;
117 p_pose.scale_x(p_scale_factor_x);
118 p_pose.scale_y(p_scale_factor_y);
119 if (fabs(p_scale_factor_x - 1) > p_floating_error && fabs(p_scale_factor_y - 1) > p_floating_error) {
120 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
121 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
122 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
124 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y,
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 = int(round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size);
133 p_windows_size.height = int(round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size);
136 if (m_use_color) p_num_of_feats += 3;
137 if (m_use_cnfeat) p_num_of_feats += 10;
138 p_roi_width = p_windows_size.width / p_cell_size;
139 p_roi_height = p_windows_size.height / p_cell_size;
143 for (int i = -p_num_scales / 2; i <= p_num_scales / 2; ++i)
144 p_scales.push_back(std::pow(p_scale_step, i));
146 p_scales.push_back(1.);
149 for (int i = p_angle_min; i <=p_angle_max ; i += p_angle_step)
150 p_angles.push_back(i);
152 p_angles.push_back(0);
155 if (p_windows_size.height / p_cell_size * (p_windows_size.width / p_cell_size / 2 + 1) > 1024) {
156 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
157 "the window dimensions so its size is less or equal to "
158 << 1024 * p_cell_size * p_cell_size * 2 + 1
159 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
160 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
161 std::exit(EXIT_FAILURE);
164 if (m_use_linearkernel) {
165 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
166 std::exit(EXIT_FAILURE);
168 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
169 p_rot_labels_data = DynMem(
170 ((uint(p_windows_size.width) / p_cell_size) * (uint(p_windows_size.height) / p_cell_size)) * sizeof(float));
171 p_rot_labels = cv::Mat(p_windows_size.height / int(p_cell_size), p_windows_size.width / int(p_cell_size), CV_32FC1,
172 p_rot_labels_data.hostMem());
174 p_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.height / p_cell_size)) / 2 + 1,
178 #if defined(CUFFT) || defined(FFTW)
179 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1,
180 uint(p_num_of_feats));
181 p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1, 1);
182 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size) / 2 + 1,
185 p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)),
186 uint(p_num_of_feats));
187 p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)), 1);
188 p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size), p_num_of_feats);
191 int max = m_use_big_batch ? 2 : p_num_scales;
192 for (int i = 0; i < max; ++i) {
193 if (m_use_big_batch && i == 1) {
194 p_threadctxs.emplace_back(
195 new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats * p_num_scales, p_num_scales));
197 p_threadctxs.emplace_back(new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats, 1));
201 p_current_scale = 1.;
203 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
204 double max_size_ratio =
205 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
206 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
207 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
208 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
210 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
211 std::cout << "init: win size. " << p_windows_size.width << " " << p_windows_size.height << std::endl;
212 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
214 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
216 fft.init(uint(p_windows_size.width / p_cell_size), uint(p_windows_size.height / p_cell_size), uint(p_num_of_feats),
217 uint(p_num_scales), m_use_big_batch);
218 fft.set_window(cosine_window_function(p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size));
220 // window weights, i.e. labels
222 gaussian_shaped_labels(p_output_sigma, p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size), p_yf,
223 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_threadctxs.front()->stream);
226 // obtain a sub-window for training initial model
227 p_threadctxs.front()->patch_feats.clear();
228 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
229 *p_threadctxs.front());
230 fft.forward_window(p_threadctxs.front()->patch_feats, p_model_xf, p_threadctxs.front()->fw_all,
231 m_use_cuda ? p_threadctxs.front()->data_features.deviceMem() : nullptr, p_threadctxs.front()->stream);
232 DEBUG_PRINTM(p_model_xf);
233 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
234 p_threadctxs.front()->model_xf = p_model_xf;
235 p_threadctxs.front()->model_xf.set_stream(p_threadctxs.front()->stream);
236 p_yf.set_stream(p_threadctxs.front()->stream);
237 p_model_xf.set_stream(p_threadctxs.front()->stream);
238 p_xf.set_stream(p_threadctxs.front()->stream);
241 if (m_use_linearkernel) {
242 ComplexMat xfconj = p_model_xf.conj();
243 p_model_alphaf_num = xfconj.mul(p_yf);
244 p_model_alphaf_den = (p_model_xf * xfconj);
246 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
247 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
248 gaussian_correlation(*p_threadctxs.front(), p_threadctxs.front()->model_xf, p_threadctxs.front()->model_xf,
249 p_kernel_sigma, true);
251 gaussian_correlation(*p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
253 DEBUG_PRINTM(p_threadctxs.front()->kf);
254 p_model_alphaf_num = p_yf * p_threadctxs.front()->kf;
255 DEBUG_PRINTM(p_model_alphaf_num);
256 p_model_alphaf_den = p_threadctxs.front()->kf * (p_threadctxs.front()->kf + float(p_lambda));
257 DEBUG_PRINTM(p_model_alphaf_den);
259 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
260 DEBUG_PRINTM(p_model_alphaf);
261 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
263 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
264 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
265 (*it)->model_xf = p_model_xf;
266 (*it)->model_xf.set_stream((*it)->stream);
267 (*it)->model_alphaf = p_model_alphaf;
268 (*it)->model_alphaf.set_stream((*it)->stream);
273 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
275 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
278 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
280 if (p_resize_image) {
282 tmp.scale(p_downscale_factor);
285 } else if (p_fit_to_pw2) {
287 tmp.scale_x(p_scale_factor_x);
288 tmp.scale_y(p_scale_factor_y);
297 BBox_c KCF_Tracker::getBBox()
300 tmp.w *= p_current_scale;
301 tmp.h *= p_current_scale;
302 tmp.a = p_current_angle;
304 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
306 tmp.scale_x(1 / p_scale_factor_x);
307 tmp.scale_y(1 / p_scale_factor_y);
313 void KCF_Tracker::track(cv::Mat &img)
315 if (m_debug || m_visual_debug) std::cout << "\nNEW FRAME" << std::endl;
316 cv::Mat input_gray, input_rgb = img.clone();
317 if (img.channels() == 3) {
318 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
319 input_gray.convertTo(input_gray, CV_32FC1);
321 img.convertTo(input_gray, CV_32FC1);
323 // don't need too large image
324 if (p_resize_image) {
325 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
326 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
327 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
328 fabs(p_scale_factor_y - 1) > p_floating_error) {
329 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
330 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
331 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
333 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
334 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
338 double max_response = -1.;
340 cv::Point2i *max_response_pt = nullptr;
341 cv::Mat *max_response_map = nullptr;
343 if (m_use_multithreading) {
344 std::vector<std::future<void>> async_res(p_scales.size());
345 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
346 uint index = uint(std::distance(p_threadctxs.begin(), it));
347 async_res[index] = std::async(std::launch::async, [this, &input_gray, &input_rgb, index, it]() -> void {
348 return scale_track(*(*it), input_rgb, input_gray, this->p_scales[index]);
351 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
352 uint index = uint(std::distance(p_threadctxs.begin(), it));
353 async_res[index].wait();
354 if ((*it)->max_response > max_response) {
355 max_response = (*it)->max_response;
356 max_response_pt = &(*it)->max_loc;
357 max_response_map = &(*it)->response;
358 scale_index = int(index);
362 uint start = m_use_big_batch ? 1 : 0;
363 uint end = m_use_big_batch ? 2 : uint(p_num_scales);
364 NORMAL_OMP_PARALLEL_FOR
365 for (uint i = start; i < end; ++i) {
366 auto it = p_threadctxs.begin();
368 scale_track(*(*it), input_rgb, input_gray, this->p_scales[i]);
370 if (m_use_big_batch) {
371 for (size_t j = 0; j < p_scales.size(); ++j) {
372 if ((*it)->max_responses[j] > max_response) {
373 max_response = (*it)->max_responses[j];
374 max_response_pt = &(*it)->max_locs[j];
375 max_response_map = &(*it)->response_maps[j];
376 scale_index = int(j);
382 if ((*it)->max_response > max_response) {
383 max_response = (*it)->max_response;
384 max_response_pt = &(*it)->max_loc;
385 max_response_map = &(*it)->response;
386 scale_index = int(i);
393 DEBUG_PRINTM(*max_response_map);
394 DEBUG_PRINT(*max_response_pt);
396 // sub pixel quadratic interpolation from neighbours
397 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
398 max_response_pt->y = max_response_pt->y - max_response_map->rows;
399 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
400 max_response_pt->x = max_response_pt->x - max_response_map->cols;
402 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
403 DEBUG_PRINT(new_location);
405 if (m_use_subpixel_localization) new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
406 DEBUG_PRINT(new_location);
408 if (m_visual_debug) std::cout << "Old p_pose, cx: " << p_pose.cx << " cy: " << p_pose.cy << std::endl;
410 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
411 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
413 if (m_visual_debug) {
414 std::cout << "New p_pose, cx: " << p_pose.cx << " cy: " << p_pose.cy << std::endl;
419 if (p_pose.cx < 0) p_pose.cx = 0;
420 if (p_pose.cx > (img.cols * p_scale_factor_x) - 1) p_pose.cx = (img.cols * p_scale_factor_x) - 1;
421 if (p_pose.cy < 0) p_pose.cy = 0;
422 if (p_pose.cy > (img.rows * p_scale_factor_y) - 1) p_pose.cy = (img.rows * p_scale_factor_y) - 1;
424 if (p_pose.cx < 0) p_pose.cx = 0;
425 if (p_pose.cx > img.cols - 1) p_pose.cx = img.cols - 1;
426 if (p_pose.cy < 0) p_pose.cy = 0;
427 if (p_pose.cy > img.rows - 1) p_pose.cy = img.rows - 1;
430 // sub grid scale interpolation
431 double new_scale = p_scales[uint(scale_index)];
432 if (m_use_subgrid_scale) new_scale = sub_grid_scale(scale_index);
434 p_current_scale *= new_scale;
436 if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
437 if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
439 // TODO Missing angle_index
440 // int tmp_angle = p_current_angle + p_angles[angle_index];
441 // p_current_angle = tmp_angle < 0 ? -std::abs(tmp_angle)%360 : tmp_angle%360;
443 // obtain a subwindow for training at newly estimated target position
444 p_threadctxs.front()->patch_feats.clear();
445 get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
446 *p_threadctxs.front(), p_current_scale, p_current_angle);
447 fft.forward_window(p_threadctxs.front()->patch_feats, p_xf, p_threadctxs.front()->fw_all,
448 m_use_cuda ? p_threadctxs.front()->data_features.deviceMem() : nullptr,
449 p_threadctxs.front()->stream);
451 // subsequent frames, interpolate model
452 p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
454 ComplexMat alphaf_num, alphaf_den;
456 if (m_use_linearkernel) {
457 ComplexMat xfconj = p_xf.conj();
458 alphaf_num = xfconj.mul(p_yf);
459 alphaf_den = (p_xf * xfconj);
461 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
462 gaussian_correlation(*p_threadctxs.front(), p_xf, p_xf, p_kernel_sigma,
464 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
465 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
466 alphaf_num = p_yf * p_threadctxs.front()->kf;
467 alphaf_den = p_threadctxs.front()->kf * (p_threadctxs.front()->kf + float(p_lambda));
470 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
471 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
472 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
474 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
475 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
476 (*it)->model_xf = p_model_xf;
477 (*it)->model_xf.set_stream((*it)->stream);
478 (*it)->model_alphaf = p_model_alphaf;
479 (*it)->model_alphaf.set_stream((*it)->stream);
484 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray, double scale)
486 if (m_use_big_batch) {
487 vars.patch_feats.clear();
488 BIG_BATCH_OMP_PARALLEL_FOR
489 for (uint i = 0; i < uint(p_num_scales); ++i) {
490 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
491 this->p_windows_size.height, vars, this->p_current_scale * this->p_scales[i]);
494 vars.patch_feats.clear();
495 get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
496 this->p_windows_size.height, vars, this->p_current_scale *scale);
499 fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
501 DEBUG_PRINTM(vars.zf);
503 if (m_use_linearkernel) {
504 vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
505 : (p_model_alphaf * vars.zf).sum_over_channels();
506 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
508 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
509 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
510 vars.kzf = vars.model_alphaf * vars.kzf;
512 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
513 DEBUG_PRINTM(this->p_model_alphaf);
514 DEBUG_PRINTM(vars.kzf);
515 vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
517 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
520 DEBUG_PRINTM(vars.response);
522 /* target location is at the maximum response. we must take into
523 account the fact that, if the target doesn't move, the peak
524 will appear at the top-left corner, not at the center (this is
525 discussed in the paper). the responses wrap around cyclically. */
526 if (m_use_big_batch) {
527 cv::split(vars.response, vars.response_maps);
529 for (size_t i = 0; i < p_scales.size(); ++i) {
530 double min_val, max_val;
531 cv::Point2i min_loc, max_loc;
532 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
533 DEBUG_PRINT(max_loc);
534 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
535 vars.max_responses[i] = max_val * weight;
536 vars.max_locs[i] = max_loc;
541 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
543 DEBUG_PRINT(vars.max_loc);
545 double weight = scale < 1. ? scale : 1. / scale;
546 vars.max_response = vars.max_val * weight;
551 // ****************************************************************************
553 void KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y,
554 ThreadCtx &vars, double scale, int angle)
556 int size_x_scaled = int(floor(size_x * scale));
557 int size_y_scaled = int(floor(size_y * scale));
559 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled /*, angle*/);
560 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled /*, angle*/);
563 cv::Point2f center((patch_gray.cols - 1) / 2., (patch_gray.rows - 1) / 2.);
564 cv::Mat r = cv::getRotationMatrix2D(center, angle, 1.0);
566 cv::warpAffine(patch_gray, patch_gray, r, cv::Size(patch_gray.cols, patch_gray.rows), cv::INTER_LINEAR,
567 cv::BORDER_REPLICATE);
569 // resize to default size
571 // if we downsample use INTER_AREA interpolation
572 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
574 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
577 // get hog(Histogram of Oriented Gradients) features
578 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
580 // get color rgb features (simple r,g,b channels)
581 std::vector<cv::Mat> color_feat;
582 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
584 cv::Point2f center((patch_rgb.cols - 1) / 2., (patch_rgb.rows - 1) / 2.);
585 cv::Mat r = cv::getRotationMatrix2D(center, angle, 1.0);
587 cv::warpAffine(patch_rgb, patch_rgb, r, cv::Size(patch_rgb.cols, patch_rgb.rows), cv::INTER_LINEAR,
588 cv::BORDER_REPLICATE);
590 if (m_visual_debug) {
591 cv::Mat patch_rgb_copy = patch_rgb.clone();
592 cv::namedWindow("Patch RGB copy", CV_WINDOW_AUTOSIZE);
593 cv::putText(patch_rgb_copy, std::to_string(angle), cv::Point(0, patch_rgb_copy.rows - 1),
594 cv::FONT_HERSHEY_COMPLEX_SMALL, 1, cv::Scalar(0, 255, 0), 2, cv::LINE_AA);
595 cv::imshow("Patch RGB copy", patch_rgb_copy);
598 // resize to default size
600 // if we downsample use INTER_AREA interpolation
601 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
604 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0.,
609 if (m_use_color && input_rgb.channels() == 3) {
610 // use rgb color space
611 cv::Mat patch_rgb_norm;
612 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
613 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
614 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
615 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
616 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
617 cv::split(patch_rgb_norm, rgb);
618 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
621 if (m_use_cnfeat && input_rgb.channels() == 3) {
622 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
623 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
625 BIG_BATCH_OMP_ORDERED
626 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
630 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
632 cv::Mat labels(dim2, dim1, CV_32FC1);
633 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
634 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
636 double sigma_s = sigma * sigma;
638 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
639 float *row_ptr = labels.ptr<float>(j);
641 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
642 row_ptr[i] = float(std::exp(-0.5 * (y_s + x * x) / sigma_s)); //-1/2*e^((y^2+x^2)/sigma^2)
646 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
648 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
649 tmp.copyTo(p_rot_labels);
651 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
654 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
655 // sanity check, 1 at top left corner
656 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
662 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
664 cv::Mat rot_patch(patch.size(), CV_32FC1);
665 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
667 // circular rotate x-axis
669 // move part that does not rotate over the edge
670 cv::Range orig_range(-x_rot, patch.cols);
671 cv::Range rot_range(0, patch.cols - (-x_rot));
672 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
675 orig_range = cv::Range(0, -x_rot);
676 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
677 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
678 } else if (x_rot > 0) {
679 // move part that does not rotate over the edge
680 cv::Range orig_range(0, patch.cols - x_rot);
681 cv::Range rot_range(x_rot, patch.cols);
682 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
685 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
686 rot_range = cv::Range(0, x_rot);
687 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
688 } else { // zero rotation
689 // move part that does not rotate over the edge
690 cv::Range orig_range(0, patch.cols);
691 cv::Range rot_range(0, patch.cols);
692 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
695 // circular rotate y-axis
697 // move part that does not rotate over the edge
698 cv::Range orig_range(-y_rot, patch.rows);
699 cv::Range rot_range(0, patch.rows - (-y_rot));
700 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
703 orig_range = cv::Range(0, -y_rot);
704 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
705 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
706 } else if (y_rot > 0) {
707 // move part that does not rotate over the edge
708 cv::Range orig_range(0, patch.rows - y_rot);
709 cv::Range rot_range(y_rot, patch.rows);
710 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
713 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
714 rot_range = cv::Range(0, y_rot);
715 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
716 } else { // zero rotation
717 // move part that does not rotate over the edge
718 cv::Range orig_range(0, patch.rows);
719 cv::Range rot_range(0, patch.rows);
720 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
726 // hann window actually (Power-of-cosine windows)
727 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
729 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
730 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
731 for (int i = 0; i < dim1; ++i)
732 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
733 N_inv = 1. / (static_cast<double>(dim2) - 1.);
734 for (int i = 0; i < dim2; ++i)
735 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
736 cv::Mat ret = m2 * m1;
740 // Returns sub-window of image input centered at [cx, cy] coordinates),
741 // with size [width, height]. If any pixels are outside of the image,
742 // they will replicate the values at the borders.
743 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height/*, int angle*/)
747 int x1 = cx - width / 2;
748 int y1 = cy - height / 2;
749 int x2 = cx + width / 2;
750 int y2 = cy + height / 2;
753 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
754 patch.create(height, width, input.type());
755 patch.setTo(double(0.f));
759 int top = 0, bottom = 0, left = 0, right = 0;
761 // fit to image coordinates, set border extensions;
770 if (x2 >= input.cols) {
771 right = x2 - input.cols + width % 2;
776 if (y2 >= input.rows) {
777 bottom = y2 - input.rows + height % 2;
782 // cv::Point2f center(x1+width/2, y1+height/2);
783 // cv::Mat r = getRotationMatrix2D(center, angle, 1.0);
785 // cv::Mat input_clone = input.clone();
787 // cv::warpAffine(input_clone, input_clone, r, cv::Size(input_clone.cols, input_clone.rows), cv::INTER_LINEAR,
788 // cv::BORDER_CONSTANT);
790 if (m_visual_debug) {
791 input_clone = input.clone();
792 cv::rectangle(input_clone, cv::Point(x1, y1), cv::Point(x2, y2), cv::Scalar(0, 255, 0));
793 cv::line(input_clone, cv::Point(0, (input_clone.rows - 1) / 2),
794 cv::Point(input_clone.cols - 1, (input_clone.rows - 1) / 2), cv::Scalar(0, 0, 255));
795 cv::line(input_clone, cv::Point((input_clone.cols - 1) / 2, 0),
796 cv::Point((input_clone.cols - 1) / 2, input_clone.rows - 1), cv::Scalar(0, 0, 255));
798 cv::imshow("Patch before copyMakeBorder", input_clone);
801 if (x2 - x1 == 0 || y2 - y1 == 0)
802 patch = cv::Mat::zeros(height, width, CV_32FC1);
804 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
805 cv::BORDER_REPLICATE);
806 if (m_visual_debug) {
808 cv::copyMakeBorder(input_clone(cv::Range(y1, y2), cv::Range(x1, x2)), patch_dummy, top, bottom, left, right,
809 cv::BORDER_REPLICATE);
810 cv::imshow("Patch after copyMakeBorder", patch_dummy);
815 assert(patch.cols == width && patch.rows == height);
820 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
821 double sigma, bool auto_correlation)
824 xf.sqr_norm(vars.xf_sqr_norm.deviceMem());
825 if (!auto_correlation) yf.sqr_norm(vars.yf_sqr_norm.deviceMem());
827 xf.sqr_norm(vars.xf_sqr_norm.hostMem());
828 if (auto_correlation) {
829 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
831 yf.sqr_norm(vars.yf_sqr_norm.hostMem());
834 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
835 DEBUG_PRINTM(vars.xyf);
836 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
838 if (auto_correlation)
839 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(),
840 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
842 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.yf_sqr_norm.deviceMem(),
843 sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
845 // ifft2 and sum over 3rd dimension, we dont care about individual channels
846 DEBUG_PRINTM(vars.ifft2_res);
848 if (xf.channels() != p_num_scales * p_num_of_feats)
849 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
851 xy_sum.create(vars.ifft2_res.size(), CV_32FC(int(p_scales.size())));
853 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
854 float *row_ptr = vars.ifft2_res.ptr<float>(y);
855 float *row_ptr_sum = xy_sum.ptr<float>(y);
856 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
857 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
858 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
859 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
860 (row_ptr + x * vars.ifft2_res.channels() +
861 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
866 DEBUG_PRINTM(xy_sum);
868 std::vector<cv::Mat> scales;
869 cv::split(xy_sum, scales);
871 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
872 for (uint i = 0; i < uint(xf.n_scales); ++i) {
873 cv::Mat in_roi(vars.in_all, cv::Rect(0, int(i) * scales[0].rows, scales[0].cols, scales[0].rows));
875 -1. / (sigma * sigma) *
876 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
878 DEBUG_PRINTM(in_roi);
881 DEBUG_PRINTM(vars.in_all);
882 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
887 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
891 if (x < 0) x = response.cols + x;
892 if (y < 0) y = response.rows + y;
893 if (x >= response.cols) x = x - response.cols;
894 if (y >= response.rows) y = y - response.rows;
896 return response.at<float>(y, x);
899 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
901 // find neighbourhood of max_loc (response is circular)
905 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);
906 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
907 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);
910 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
911 cv::Mat A = (cv::Mat_<float>(9, 6) <<
912 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
913 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
914 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
915 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
916 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
917 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
918 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
919 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
920 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);
921 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
922 get_response_circular(p1, response),
923 get_response_circular(p2, response),
924 get_response_circular(p3, response),
925 get_response_circular(p4, response),
926 get_response_circular(p5, response),
927 get_response_circular(p6, response),
928 get_response_circular(p7, response),
929 get_response_circular(p8, response),
930 get_response_circular(max_loc, response));
933 cv::solve(A, fval, x, cv::DECOMP_SVD);
935 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);
937 cv::Point2f sub_peak(max_loc.x, max_loc.y);
938 if (b > 0 || b < 0) {
939 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
940 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
946 double KCF_Tracker::sub_grid_scale(int index)
949 if (index < 0 || index > int(p_scales.size()) - 1) {
950 // interpolate from all values
951 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
952 A.create(int(p_scales.size()), 3, CV_32FC1);
953 fval.create(int(p_scales.size()), 1, CV_32FC1);
954 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
955 uint i = uint(std::distance(p_threadctxs.begin(), it));
957 A.at<float>(j, 0) = float(p_scales[i] * p_scales[i]);
958 A.at<float>(j, 1) = float(p_scales[i]);
959 A.at<float>(j, 2) = 1;
961 m_use_big_batch ? float(p_threadctxs.back()->max_responses[i]) : float((*it)->max_response);
964 // only from neighbours
965 if (index == 0 || index == int(p_scales.size()) - 1) return p_scales[uint(index)];
967 A = (cv::Mat_<float>(3, 3) << p_scales[uint(index) - 1] * p_scales[uint(index) - 1], p_scales[uint(index) - 1],
968 1, p_scales[uint(index)] * p_scales[uint(index)], p_scales[uint(index)], 1,
969 p_scales[uint(index) + 1] * p_scales[uint(index) + 1], p_scales[uint(index) + 1], 1);
970 auto it1 = p_threadctxs.begin();
971 std::advance(it1, index - 1);
972 auto it2 = p_threadctxs.begin();
973 std::advance(it2, index);
974 auto it3 = p_threadctxs.begin();
975 std::advance(it3, index + 1);
976 fval = (cv::Mat_<float>(3, 1) << (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index) - 1]
977 : (*it1)->max_response),
978 (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index)] : (*it2)->max_response),
979 (m_use_big_batch ? p_threadctxs.back()->max_responses[uint(index) + 1] : (*it3)->max_response));
983 cv::solve(A, fval, x, cv::DECOMP_SVD);
984 float a = x.at<float>(0), b = x.at<float>(1);
985 double scale = p_scales[uint(index)];
986 if (a > 0 || a < 0) scale = double(-b / (2 * a));