10 #include "fft_cufft.h"
13 #include "fft_opencv.h"
21 #define DEBUG_PRINT(obj) \
22 if (m_debug || m_visual_debug) { \
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;
139 for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
140 p_scales.push_back(std::pow(p_scale_step, i));
142 p_scales.push_back(1.);
147 for (int i = p_angle_min; i <= p_angle_max; i += p_angle_step)
148 p_angles.push_back(i);
150 p_angles.push_back(0);
155 if (p_roi.height * (p_roi.width / 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"
160 << p_windows_size.height << " which is " << p_windows_size.width * p_windows_size.height
161 << " pixels. " << std::endl;
162 std::exit(EXIT_FAILURE);
165 if (m_use_linearkernel) {
166 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
167 std::exit(EXIT_FAILURE);
170 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
172 p_rot_labels_data = DynMem(p_roi.width * p_roi.height * sizeof(float));
173 p_rot_labels = cv::Mat(p_roi, CV_32FC1, p_rot_labels_data.hostMem());
176 #if defined(CUFFT) || defined(FFTW)
177 uint width = p_roi.width / 2 + 1;
179 uint width = p_roi.width;
181 p_model_xf.create(p_roi.height, width, p_num_of_feats);
182 p_yf.create(p_roi.height, width, 1);
183 p_xf.create(p_roi.height, width, p_num_of_feats);
185 int max1 = BIG_BATCH_MODE ? 2 : p_num_scales;
186 int max2 = BIG_BATCH_MODE ? 1 : p_num_angles;
187 for (int i = 0; i < max1; ++i) {
188 for (int j = 0; j < max2; ++j) {
189 if (BIG_BATCH_MODE && i == 1)
190 p_threadctxs.emplace_back(p_roi, p_num_of_feats * p_num_scales * p_num_angles, 1, 0, p_num_scales,
193 p_threadctxs.emplace_back(p_roi, p_num_of_feats, p_scales[i], p_angles[j]);
197 p_current_scale = 1.;
199 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
200 double max_size_ratio =
201 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
202 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
203 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
204 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
206 std::cout << "init: img size " << img.cols << "x" << img.rows << std::endl;
207 std::cout << "init: win size " << p_windows_size.width << "x" << p_windows_size.height << std::endl;
208 std::cout << "init: FFT size " << p_roi.width << "x" << p_roi.height << std::endl;
209 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
211 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / p_cell_size;
213 fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales);
214 fft.set_window(cosine_window_function(p_roi.width, p_roi.height));
216 // window weights, i.e. labels
217 fft.forward(gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height), p_yf,
218 m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_threadctxs.front().stream);
221 // obtain a sub-window for training initial model
222 int size_x_scaled = floor(p_windows_size.width);
223 int size_y_scaled = floor(p_windows_size.height);
225 cv::Mat patch_gray = get_subwindow(input_gray, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
226 geometric_transformations(patch_gray, p_windows_size.width, p_windows_size.height, 0, false);
229 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
230 patch_rgb = get_subwindow(input_rgb, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
231 geometric_transformations(patch_rgb, p_windows_size.width, p_windows_size.height, 0, false);
234 std::vector<cv::Mat> patch_feats = get_features(patch_rgb, patch_gray);
235 fft.forward_window(patch_feats, p_model_xf, p_threadctxs.front().fw_all,
236 m_use_cuda ? p_threadctxs.front().data_features.deviceMem() : nullptr,
237 p_threadctxs.front().stream);
238 DEBUG_PRINTM(p_model_xf);
240 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
241 p_threadctxs.front().model_xf = p_model_xf;
242 p_threadctxs.front().model_xf.set_stream(p_threadctxs.front().stream);
243 p_yf.set_stream(p_threadctxs.front().stream);
244 p_model_xf.set_stream(p_threadctxs.front().stream);
245 p_xf.set_stream(p_threadctxs.front().stream);
248 if (m_use_linearkernel) {
249 ComplexMat xfconj = p_model_xf.conj();
250 p_model_alphaf_num = xfconj.mul(p_yf);
251 p_model_alphaf_den = (p_model_xf * xfconj);
253 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
254 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
255 gaussian_correlation(p_threadctxs.front(), p_threadctxs.front().model_xf, p_threadctxs.front().model_xf,
256 p_kernel_sigma, true);
258 gaussian_correlation(p_threadctxs.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
260 DEBUG_PRINTM(p_threadctxs.front().kf);
261 p_model_alphaf_num = p_yf * p_threadctxs.front().kf;
262 DEBUG_PRINTM(p_model_alphaf_num);
263 p_model_alphaf_den = p_threadctxs.front().kf * (p_threadctxs.front().kf + float(p_lambda));
264 DEBUG_PRINTM(p_model_alphaf_den);
266 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
267 DEBUG_PRINTM(p_model_alphaf);
268 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
270 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
271 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
272 it->model_xf = p_model_xf;
273 it->model_xf.set_stream(it->stream);
274 it->model_alphaf = p_model_alphaf;
275 it->model_alphaf.set_stream(it->stream);
280 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
282 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
285 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
287 if (p_resize_image) {
289 tmp.scale(p_downscale_factor);
292 } else if (p_fit_to_pw2) {
294 tmp.scale_x(p_scale_factor_x);
295 tmp.scale_y(p_scale_factor_y);
304 BBox_c KCF_Tracker::getBBox()
307 tmp.w *= p_current_scale;
308 tmp.h *= p_current_scale;
309 tmp.a = p_current_angle;
311 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
313 tmp.scale_x(1 / p_scale_factor_x);
314 tmp.scale_y(1 / p_scale_factor_y);
320 double KCF_Tracker::getFilterResponse() const
322 return this->max_response;
325 void KCF_Tracker::track(cv::Mat &img)
327 if (m_debug || m_visual_debug) std::cout << "\nNEW FRAME" << std::endl;
328 cv::Mat input_gray, input_rgb = img.clone();
329 if (img.channels() == 3) {
330 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
331 input_gray.convertTo(input_gray, CV_32FC1);
333 img.convertTo(input_gray, CV_32FC1);
335 // don't need too large image
336 if (p_resize_image) {
337 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
338 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
339 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
340 fabs(p_scale_factor_y - 1) > p_floating_error) {
341 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
342 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
343 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
345 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
346 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
350 ThreadCtx *max = nullptr;
351 cv::Point2i *max_response_pt = nullptr;
352 cv::Mat *max_response_map = nullptr;
355 for (auto &it : p_threadctxs)
356 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
357 scale_track(it, input_rgb, input_gray);
359 for (auto const &it : p_threadctxs)
362 NORMAL_OMP_PARALLEL_FOR
363 for (uint i = BIG_BATCH_MODE ? 1 : 0; i < p_threadctxs.size(); ++i)
364 scale_track(p_threadctxs[i], input_rgb, input_gray);
368 for (auto &it : p_threadctxs) {
369 if (it.max_response > max_response) {
370 max_response = it.max_response;
371 max_response_pt = &it.max_loc;
372 max_response_map = &it.response;
377 for (uint j = 0; j < p_num_scales; ++j) {
378 for (uint k = 0; k < p_num_angles; ++k) {
379 if (p_threadctxs.back().max_responses[j + k] > max_response) {
380 max_response = p_threadctxs.back().max_responses[j + k];
381 max_response_pt = &p_threadctxs.back().max_locs[j + k];
382 max_response_map = &p_threadctxs.back().response_maps[j + k];
386 max = &p_threadctxs.back();
388 if (m_visual_debug) {
389 cv::Mat all_responses(cv::Size(p_num_angles* p_debug_image_size, p_num_scales * p_debug_image_size),
390 p_debug_scale_responses[0].type(), cv::Scalar::all(0));
391 cv::Mat all_subwindows(cv::Size(p_num_angles* p_debug_image_size, p_num_scales* p_debug_image_size),
392 p_debug_subwindows[0].type(), cv::Scalar::all(0));
393 for (size_t i = 0; i < p_num_scales; ++i) {
394 for (size_t j = 0; j < p_num_angles; ++j) {
395 cv::Mat in_roi(all_responses, cv::Rect(j * p_debug_image_size, i * p_debug_image_size,
396 p_debug_image_size, p_debug_image_size));
397 p_debug_scale_responses[5 * i + j].copyTo(in_roi);
398 in_roi = all_subwindows(
399 cv::Rect(j * p_debug_image_size, i * p_debug_image_size, p_debug_image_size, p_debug_image_size));
400 p_debug_subwindows[5 * i + j].copyTo(in_roi);
403 cv::namedWindow("All subwindows", CV_WINDOW_AUTOSIZE);
404 cv::imshow("All subwindows", all_subwindows);
405 cv::namedWindow("All responses", CV_WINDOW_AUTOSIZE);
406 cv::imshow("All responses", all_responses);
408 p_debug_scale_responses.clear();
409 p_debug_subwindows.clear();
412 DEBUG_PRINTM(*max_response_map);
413 DEBUG_PRINT(*max_response_pt);
415 // sub pixel quadratic interpolation from neighbours
416 if (max_response_pt->y > max_response_map->rows / 2) // wrap around to negative half-space of vertical axis
417 max_response_pt->y = max_response_pt->y - max_response_map->rows;
418 if (max_response_pt->x > max_response_map->cols / 2) // same for horizontal axis
419 max_response_pt->x = max_response_pt->x - max_response_map->cols;
421 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
422 DEBUG_PRINT(new_location);
424 if (m_use_subpixel_localization)
425 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
426 DEBUG_PRINT(new_location);
428 if (m_visual_debug) std::cout << "Old p_pose, cx: " << p_pose.cx << " cy: " << p_pose.cy << std::endl;
430 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
431 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
433 if (m_visual_debug) std::cout << "New p_pose, cx: " << p_pose.cx << " cy: " << p_pose.cy << std::endl;
436 clamp2(p_pose.cx, 0.0, (img.cols * p_scale_factor_x) - 1);
437 clamp2(p_pose.cy, 0.0, (img.rows * p_scale_factor_y) - 1);
439 clamp2(p_pose.cx, 0.0, img.cols - 1.0);
440 clamp2(p_pose.cy, 0.0, img.rows - 1.0);
443 // sub grid scale interpolation
444 if (m_use_subgrid_scale) {
445 auto it = std::find_if(p_threadctxs.begin(), p_threadctxs.end(), [max](ThreadCtx &ctx) { return &ctx == max; });
446 p_current_scale *= sub_grid_scale(std::distance(p_threadctxs.begin(), it));
448 p_current_scale *= max->scale;
451 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
453 if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
454 if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
456 p_current_angle = (p_current_angle + max->angle) < 0
457 ? -std::abs(p_current_angle + max->angle) % 360
458 : (p_current_angle + max->angle) % 360;
460 // obtain a subwindow for training at newly estimated target position
461 int size_x_scaled = floor(p_windows_size.width * p_current_scale);
462 int size_y_scaled = floor(p_windows_size.height * p_current_scale);
464 cv::Mat patch_gray = get_subwindow(input_gray, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
465 geometric_transformations(patch_gray, p_windows_size.width, p_windows_size.height, p_current_angle, false);
467 cv::Mat patch_rgb = cv::Mat::zeros(size_y_scaled, size_x_scaled, CV_32F);
468 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
469 patch_rgb = get_subwindow(input_rgb, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
470 geometric_transformations(patch_rgb, p_windows_size.width, p_windows_size.height, p_current_angle, false);
473 ThreadCtx &ctx = p_threadctxs.front();
474 std::vector<cv::Mat> patch_feats = get_features(patch_rgb, patch_gray);
475 fft.forward_window(patch_feats, p_xf, ctx.fw_all, m_use_cuda ? ctx.data_features.deviceMem() : nullptr, ctx.stream);
477 // subsequent frames, interpolate model
478 p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
480 ComplexMat alphaf_num, alphaf_den;
482 if (m_use_linearkernel) {
483 ComplexMat xfconj = p_xf.conj();
484 alphaf_num = xfconj.mul(p_yf);
485 alphaf_den = (p_xf * xfconj);
487 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
488 gaussian_correlation(ctx, p_xf, p_xf, p_kernel_sigma,
490 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
491 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
492 alphaf_num = p_yf * ctx.kf;
493 alphaf_den = ctx.kf * (ctx.kf + float(p_lambda));
496 p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
497 p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
498 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
500 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
501 for (auto it = p_threadctxs.begin(); it != p_threadctxs.end(); ++it) {
502 it->model_xf = p_model_xf;
503 it->model_xf.set_stream(it->stream);
504 it->model_alphaf = p_model_alphaf;
505 it->model_alphaf.set_stream(it->stream);
510 void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray)
512 std::vector<cv::Mat> patch_feats;
513 if (BIG_BATCH_MODE) {
514 BIG_BATCH_OMP_PARALLEL_FOR
515 for (uint i = 0; i < this->p_scales.size(); ++i) {
516 for (uint j = 0; j < this->p_angles.size(); ++j) {
517 int size_x_scaled = floor(this->p_windows_size.width * this->p_current_scale * this->p_scales[i]);
518 int size_y_scaled = floor(this->p_windows_size.height * this->p_current_scale * this->p_scales[i]);
521 get_subwindow(input_gray, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
522 geometric_transformations(patch_gray, p_windows_size.width, p_windows_size.height,
523 p_current_scale * this->p_scales[i], p_current_angle + this->p_angles[j]);
526 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
528 get_subwindow(input_rgb, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
529 geometric_transformations(patch_rgb, p_windows_size.width, p_windows_size.height,
530 p_current_scale * this->p_scales[i], p_current_angle + this->p_angles[j]);
532 std::vector<cv::Mat> tmp = get_features(patch_rgb, patch_gray);
533 BIG_BATCH_OMP_ORDERED
534 patch_feats.insert(patch_feats.end(), tmp.begin(), tmp.end());
538 int size_x_scaled = floor(this->p_windows_size.width * this->p_current_scale * vars.scale);
539 int size_y_scaled = floor(this->p_windows_size.height * this->p_current_scale * vars.scale);
541 cv::Mat patch_gray = get_subwindow(input_gray, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
542 geometric_transformations(patch_gray, p_windows_size.width, p_windows_size.height, p_current_scale * vars.scale);
545 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
546 patch_rgb = get_subwindow(input_rgb, this->p_pose.cx, this->p_pose.cy, size_x_scaled, size_y_scaled);
547 geometric_transformations(patch_rgb, p_windows_size.width, p_windows_size.height, p_current_scale * vars.scale,
548 p_current_angle + vars.angle);
550 patch_feats = get_features(patch_rgb, patch_gray);
553 fft.forward_window(patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
555 DEBUG_PRINTM(vars.zf);
557 if (m_use_linearkernel) {
558 vars.kzf = BIG_BATCH_MODE ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
559 : (p_model_alphaf * vars.zf).sum_over_channels();
560 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
562 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
563 gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
564 vars.kzf = vars.model_alphaf * vars.kzf;
566 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
567 DEBUG_PRINTM(this->p_model_alphaf);
568 DEBUG_PRINTM(vars.kzf);
569 vars.kzf = BIG_BATCH_MODE ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
571 fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
574 DEBUG_PRINTM(vars.response);
576 /* target location is at the maximum response. we must take into
577 account the fact that, if the target doesn't move, the peak
578 will appear at the top-left corner, not at the center (this is
579 discussed in the paper). the responses wrap around cyclically. */
581 cv::split(vars.response, vars.response_maps);
583 for (size_t i = 0; i < p_scales.size(); ++i) {
584 double min_val, max_val;
585 cv::Point2i min_loc, max_loc;
586 cv::minMaxLoc(vars.response_maps[i], &min_val, &max_val, &min_loc, &max_loc);
587 DEBUG_PRINT(max_loc);
588 double weight = p_scales[i] < 1. ? p_scales[i] : 1. / p_scales[i];
589 vars.max_responses[i] = max_val * weight;
590 vars.max_locs[i] = max_loc;
595 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
597 DEBUG_PRINT(vars.max_loc);
599 double weight = vars.scale < 1. ? vars.scale : 1. / vars.scale;
600 vars.max_response = vars.max_val * weight;
605 // ****************************************************************************
607 std::vector<cv::Mat> KCF_Tracker::get_features(cv::Mat &patch_rgb, cv::Mat &patch_gray)
609 // get hog(Histogram of Oriented Gradients) features
610 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
612 // get color rgb features (simple r,g,b channels)
613 std::vector<cv::Mat> color_feat;
615 if (m_use_color && patch_rgb.channels() == 3) {
616 // use rgb color space
617 cv::Mat patch_rgb_norm;
618 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
619 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
620 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
621 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
622 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
623 cv::split(patch_rgb_norm, rgb);
624 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
627 if (m_use_cnfeat && patch_rgb.channels() == 3) {
628 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
629 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
632 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
636 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
638 cv::Mat labels(dim2, dim1, CV_32FC1);
639 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
640 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
642 double sigma_s = sigma * sigma;
644 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
645 float *row_ptr = labels.ptr<float>(j);
647 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
648 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
652 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
654 cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
655 tmp.copyTo(p_rot_labels);
657 assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
660 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
661 // sanity check, 1 at top left corner
662 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
668 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
670 cv::Mat rot_patch(patch.size(), CV_32FC1);
671 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
673 // circular rotate x-axis
675 // move part that does not rotate over the edge
676 cv::Range orig_range(-x_rot, patch.cols);
677 cv::Range rot_range(0, patch.cols - (-x_rot));
678 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
681 orig_range = cv::Range(0, -x_rot);
682 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
683 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
684 } else if (x_rot > 0) {
685 // move part that does not rotate over the edge
686 cv::Range orig_range(0, patch.cols - x_rot);
687 cv::Range rot_range(x_rot, patch.cols);
688 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
691 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
692 rot_range = cv::Range(0, x_rot);
693 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
694 } else { // zero rotation
695 // move part that does not rotate over the edge
696 cv::Range orig_range(0, patch.cols);
697 cv::Range rot_range(0, patch.cols);
698 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
701 // circular rotate y-axis
703 // move part that does not rotate over the edge
704 cv::Range orig_range(-y_rot, patch.rows);
705 cv::Range rot_range(0, patch.rows - (-y_rot));
706 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
709 orig_range = cv::Range(0, -y_rot);
710 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
711 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
712 } else if (y_rot > 0) {
713 // move part that does not rotate over the edge
714 cv::Range orig_range(0, patch.rows - y_rot);
715 cv::Range rot_range(y_rot, patch.rows);
716 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
719 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
720 rot_range = cv::Range(0, y_rot);
721 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
722 } else { // zero rotation
723 // move part that does not rotate over the edge
724 cv::Range orig_range(0, patch.rows);
725 cv::Range rot_range(0, patch.rows);
726 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
732 // hann window actually (Power-of-cosine windows)
733 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
735 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
736 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
737 for (int i = 0; i < dim1; ++i)
738 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
739 N_inv = 1. / (static_cast<double>(dim2) - 1.);
740 for (int i = 0; i < dim2; ++i)
741 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
742 cv::Mat ret = m2 * m1;
746 // Returns sub-window of image input centered at [cx, cy] coordinates),
747 // with size [width, height]. If any pixels are outside of the image,
748 // they will replicate the values at the borders.
749 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
753 int x1 = cx - width / 2;
754 int y1 = cy - height / 2;
755 int x2 = cx + width / 2;
756 int y2 = cy + height / 2;
759 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
760 patch.create(height, width, input.type());
761 patch.setTo(double(0.f));
765 int top = 0, bottom = 0, left = 0, right = 0;
767 // fit to image coordinates, set border extensions;
776 if (x2 >= input.cols) {
777 right = x2 - input.cols + width % 2;
782 if (y2 >= input.rows) {
783 bottom = y2 - input.rows + height % 2;
788 if (x2 - x1 == 0 || y2 - y1 == 0)
789 patch = cv::Mat::zeros(height, width, CV_32FC1);
791 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
792 cv::BORDER_REPLICATE);
795 assert(patch.cols == width && patch.rows == height);
800 void KCF_Tracker::geometric_transformations(cv::Mat &patch, int size_x, int size_y, int angle, bool allow_debug)
803 cv::Point2f center((patch.cols - 1) / 2., (patch.rows - 1) / 2.);
804 cv::Mat r = cv::getRotationMatrix2D(center, angle, 1.0);
806 cv::warpAffine(patch, patch, r, cv::Size(patch.cols, patch.rows), cv::INTER_LINEAR, cv::BORDER_REPLICATE);
809 // resize to default size
810 if (patch.channels() != 3) {
811 if (patch.cols / size_x > 1.) {
812 // if we downsample use INTER_AREA interpolation
813 cv::resize(patch, patch, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
815 cv::resize(patch, patch, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
818 if (patch.cols / size_x > 1.) {
819 // if we downsample use INTER_AREA interpolation
820 cv::resize(patch, patch, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_AREA);
822 cv::resize(patch, patch, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_LINEAR);
824 if (m_visual_debug && allow_debug) {
825 cv::Mat input_clone = patch.clone();
826 cv::resize(input_clone, input_clone, cv::Size(p_debug_image_size, p_debug_image_size), 0., 0.,
829 std::string angle_string = std::to_string(p_current_angle + angle);
831 cv::putText(input_clone, angle_string, cv::Point(1, input_clone.rows - 5), cv::FONT_HERSHEY_COMPLEX_SMALL,
832 0.5, cv::Scalar(0, 255, 0), 1);
834 p_debug_subwindows.push_back(input_clone);
839 void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf, double sigma,
840 bool auto_correlation)
842 xf.sqr_norm(vars.xf_sqr_norm);
843 if (auto_correlation) {
844 vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
846 yf.sqr_norm(vars.yf_sqr_norm);
848 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
849 DEBUG_PRINTM(vars.xyf);
850 fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
852 cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(),
853 vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(), sigma, xf.n_channels,
854 xf.n_scales, p_roi.height, p_roi.width, vars.stream);
856 // ifft2 and sum over 3rd dimension, we dont care about individual channels
857 DEBUG_PRINTM(vars.ifft2_res);
859 if (xf.channels() != p_num_scales * p_num_of_feats)
860 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
862 xy_sum.create(vars.ifft2_res.size(), CV_32FC(p_scales.size()));
864 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
865 float *row_ptr = vars.ifft2_res.ptr<float>(y);
866 float *row_ptr_sum = xy_sum.ptr<float>(y);
867 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
868 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
869 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
870 row_ptr + x * vars.ifft2_res.channels() + sum_ch * (vars.ifft2_res.channels() / xy_sum.channels()),
871 (row_ptr + x * vars.ifft2_res.channels() +
872 (sum_ch + 1) * (vars.ifft2_res.channels() / xy_sum.channels())),
877 DEBUG_PRINTM(xy_sum);
879 std::vector<cv::Mat> scales;
880 cv::split(xy_sum, scales);
882 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
883 for (uint i = 0; i < xf.n_scales; ++i) {
884 cv::Mat in_roi(vars.in_all, cv::Rect(0, i * scales[0].rows, scales[0].cols, scales[0].rows));
886 -1. / (sigma * sigma) *
887 cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
889 DEBUG_PRINTM(in_roi);
892 DEBUG_PRINTM(vars.in_all);
893 fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
898 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
902 if (x < 0) x = response.cols + x;
903 if (y < 0) y = response.rows + y;
904 if (x >= response.cols) x = x - response.cols;
905 if (y >= response.rows) y = y - response.rows;
907 return response.at<float>(y, x);
910 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
912 // find neighbourhood of max_loc (response is circular)
916 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);
917 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
918 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);
921 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
922 cv::Mat A = (cv::Mat_<float>(9, 6) <<
923 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
924 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
925 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
926 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
927 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
928 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
929 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
930 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
931 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);
932 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
933 get_response_circular(p1, response),
934 get_response_circular(p2, response),
935 get_response_circular(p3, response),
936 get_response_circular(p4, response),
937 get_response_circular(p5, response),
938 get_response_circular(p6, response),
939 get_response_circular(p7, response),
940 get_response_circular(p8, response),
941 get_response_circular(max_loc, response));
944 cv::solve(A, fval, x, cv::DECOMP_SVD);
946 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);
948 cv::Point2f sub_peak(max_loc.x, max_loc.y);
949 if (b > 0 || b < 0) {
950 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
951 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
957 double KCF_Tracker::sub_grid_scale(uint index)
960 if (index >= p_scales.size()) {
961 // interpolate from all values
962 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
963 A.create(p_scales.size(), 3, CV_32FC1);
964 fval.create(p_scales.size(), 1, CV_32FC1);
965 for (size_t i = 0; i < p_scales.size(); ++i) {
966 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
967 A.at<float>(i, 1) = float(p_scales[i]);
968 A.at<float>(i, 2) = 1;
970 fval.at<float>(i) = p_threadctxs.back().max_responses[i];
972 fval.at<float>(i) = p_threadctxs[i].max_response;
976 // only from neighbours
977 if (index == 0 || index == p_scales.size() - 1)
978 return p_scales[index];
980 A = (cv::Mat_<float>(3, 3) <<
981 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
982 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
983 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
985 fval = (cv::Mat_<float>(3, 1) <<
986 p_threadctxs.back().max_responses[index - 1],
987 p_threadctxs.back().max_responses[index + 0],
988 p_threadctxs.back().max_responses[index + 1]);
990 fval = (cv::Mat_<float>(3, 1) <<
991 p_threadctxs[index - 1].max_response,
992 p_threadctxs[index + 0].max_response,
993 p_threadctxs[index + 1].max_response);
998 cv::solve(A, fval, x, cv::DECOMP_SVD);
999 float a = x.at<float>(0), b = x.at<float>(1);
1000 double scale = p_scales[index];
1002 scale = -b / (2 * a);