5 #include "threadctx.hpp"
11 #include "fft_cufft.h"
14 #include "fft_opencv.h"
22 static bool kcf_debug = false;
24 #define DEBUG_PRINT(obj) \
26 std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl; \
28 #define DEBUG_PRINTM(obj) \
30 std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl \
31 << (obj) << std::endl; \
35 T clamp(const T& n, const T& lower, const T& upper)
37 return std::max(lower, std::min(n, upper));
41 void clamp2(T& n, const T& lower, const T& upper)
43 n = std::max(lower, std::min(n, upper));
46 class Kcf_Tracker_Private {
48 std::vector<ThreadCtx> threadctxs;
51 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
52 double output_sigma_factor, int cell_size)
53 : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
54 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size), d(*new Kcf_Tracker_Private)
58 KCF_Tracker::KCF_Tracker() : fft(*new FFT()), d(*new Kcf_Tracker_Private) {}
60 KCF_Tracker::~KCF_Tracker()
66 void KCF_Tracker::train(cv::Mat input_gray, cv::Mat input_rgb, double interp_factor)
68 // obtain a sub-window for training
69 int sizes[3] = {p_num_of_feats, p_roi.height, p_roi.width};
70 MatDynMem patch_feats(3, sizes, CV_32FC1);
71 MatDynMem temp(3, sizes, CV_32FC1);
72 get_features(patch_feats, input_rgb, input_gray, p_pose.cx, p_pose.cy,
73 p_windows_size.width, p_windows_size.height,
75 fft.forward_window(patch_feats, p_xf, temp);
76 p_model_xf = p_model_xf * (1. - interp_factor) + p_xf * interp_factor;
77 DEBUG_PRINTM(p_model_xf);
79 ComplexMat alphaf_num, alphaf_den;
81 if (m_use_linearkernel) {
82 ComplexMat xfconj = p_xf.conj();
83 alphaf_num = xfconj.mul(p_yf);
84 alphaf_den = (p_xf * xfconj);
86 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
87 const uint num_scales = BIG_BATCH_MODE ? p_num_scales : 1;
88 cv::Size sz(Fft::freq_size(p_roi));
89 ComplexMat kf(sz.height, sz.width, num_scales);
90 (*gaussian_correlation)(*this, kf, p_model_xf, p_model_xf, p_kernel_sigma, true);
92 p_model_alphaf_num = p_yf * kf;
93 DEBUG_PRINTM(p_model_alphaf_num);
94 p_model_alphaf_den = kf * (kf + p_lambda);
95 DEBUG_PRINTM(p_model_alphaf_den);
97 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
98 DEBUG_PRINTM(p_model_alphaf);
99 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
102 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
105 // check boundary, enforce min size
106 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
108 if (x2 > img.cols - 1) x2 = img.cols - 1;
110 if (y2 > img.rows - 1) y2 = img.rows - 1;
112 if (x2 - x1 < 2 * p_cell_size) {
113 double diff = (2 * p_cell_size - x2 + x1) / 2.;
114 if (x1 - diff >= 0 && x2 + diff < img.cols) {
117 } else if (x1 - 2 * diff >= 0) {
123 if (y2 - y1 < 2 * p_cell_size) {
124 double diff = (2 * p_cell_size - y2 + y1) / 2.;
125 if (y1 - diff >= 0 && y2 + diff < img.rows) {
128 } else if (y1 - 2 * diff >= 0) {
137 p_pose.cx = x1 + p_pose.w / 2.;
138 p_pose.cy = y1 + p_pose.h / 2.;
140 cv::Mat input_gray, input_rgb = img.clone();
141 if (img.channels() == 3) {
142 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
143 input_gray.convertTo(input_gray, CV_32FC1);
145 img.convertTo(input_gray, CV_32FC1);
147 // don't need too large image
148 if (p_pose.w * p_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
149 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
150 p_resize_image = true;
151 p_pose.scale(p_downscale_factor);
152 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
153 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
154 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
155 if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
156 std::cerr << "Error: Fit size is not multiple of HOG cell size (" << p_cell_size << ")" << std::endl;
157 std::exit(EXIT_FAILURE);
159 p_scale_factor_x = (double)fit_size_x / round(p_pose.w * (1. + p_padding));
160 p_scale_factor_y = (double)fit_size_y / round(p_pose.h * (1. + p_padding));
161 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
162 << p_scale_factor_y << std::endl;
164 p_pose.scale_x(p_scale_factor_x);
165 p_pose.scale_y(p_scale_factor_y);
166 if (fabs(p_scale_factor_x - 1) > p_floating_error || fabs(p_scale_factor_y - 1) > p_floating_error) {
167 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
168 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
169 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
171 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
172 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
177 // compute win size + fit to fhog cell size
178 p_windows_size.width = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
179 p_windows_size.height = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
180 p_roi.width = p_windows_size.width / p_cell_size;
181 p_roi.height = p_windows_size.height / p_cell_size;
185 for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
186 p_scales.push_back(std::pow(p_scale_step, i));
188 p_scales.push_back(1.);
191 if (p_roi.height * (p_roi.width / 2 + 1) > 1024) {
192 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
193 "the window dimensions so its size is less or equal to "
194 << 1024 * p_cell_size * p_cell_size * 2 + 1
195 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
196 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
197 std::exit(EXIT_FAILURE);
200 if (m_use_linearkernel) {
201 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
202 std::exit(EXIT_FAILURE);
204 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
206 p_xf.create(p_roi.height, p_roi.height / 2 + 1, p_num_of_feats);
209 #if defined(CUFFT) || defined(FFTW)
210 uint width = p_roi.width / 2 + 1;
212 uint width = p_roi.width;
214 p_model_xf.create(p_roi.height, width, p_num_of_feats);
215 p_yf.create(p_roi.height, width, 1);
216 p_xf.create(p_roi.height, width, p_num_of_feats);
219 for (auto scale: p_scales)
220 d.threadctxs.emplace_back(p_roi, p_num_of_feats, 1, scale);
222 d.threadctxs.emplace_back(p_roi, p_num_of_feats * p_num_scales, p_num_scales);
225 gaussian_correlation.reset(new GaussianCorrelation(p_roi, IF_BIG_BATCH(p_num_scales, 1),
226 p_num_of_feats * IF_BIG_BATCH(p_num_scales, 1)));
228 p_current_scale = 1.;
230 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
231 double max_size_ratio =
232 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
233 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
234 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
235 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
237 std::cout << "init: img size " << img.cols << "x" << img.rows << std::endl;
238 std::cout << "init: win size " << p_windows_size.width << "x" << p_windows_size.height << std::endl;
239 std::cout << "init: FFT size " << p_roi.width << "x" << p_roi.height << std::endl;
240 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
242 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / p_cell_size;
244 fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales);
245 fft.set_window(MatDynMem(cosine_window_function(p_roi.width, p_roi.height)));
247 // window weights, i.e. labels
248 fft.forward(gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height), p_yf);
251 // train initial model
252 train(input_gray, input_rgb, 1.0);
255 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
257 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
260 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
262 if (p_resize_image) {
264 tmp.scale(p_downscale_factor);
267 } else if (p_fit_to_pw2) {
269 tmp.scale_x(p_scale_factor_x);
270 tmp.scale_y(p_scale_factor_y);
279 BBox_c KCF_Tracker::getBBox()
282 tmp.w *= p_current_scale;
283 tmp.h *= p_current_scale;
285 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
287 tmp.scale_x(1 / p_scale_factor_x);
288 tmp.scale_y(1 / p_scale_factor_y);
294 double KCF_Tracker::getFilterResponse() const
296 return this->max_response;
299 void KCF_Tracker::resizeImgs(cv::Mat &input_rgb, cv::Mat &input_gray)
301 if (p_resize_image) {
302 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
303 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
304 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
305 fabs(p_scale_factor_y - 1) > p_floating_error) {
306 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
307 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
308 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
310 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
311 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
316 void KCF_Tracker::findMaxReponse(uint &max_idx, cv::Point2f &new_location) const
320 for (uint j = 0; j < d.threadctxs.size(); ++j) {
321 if (d.threadctxs[j].max.response > max) {
322 max = d.threadctxs[j].max.response;
327 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
328 for (uint j = 0; j < p_scales.size(); ++j) {
329 if (d.threadctxs[0].max[j].response > max) {
330 max = d.threadctxs[0].max[j].response;
335 cv::Point2i &max_response_pt = IF_BIG_BATCH(d.threadctxs[0].max[max_idx].loc, d.threadctxs[max_idx].max.loc);
336 cv::Mat max_response_map = IF_BIG_BATCH(d.threadctxs[0].response.plane(max_idx), d.threadctxs[max_idx].response);
338 DEBUG_PRINTM(max_response_map);
339 DEBUG_PRINT(max_response_pt);
341 // sub pixel quadratic interpolation from neighbours
342 if (max_response_pt.y > max_response_map.rows / 2) // wrap around to negative half-space of vertical axis
343 max_response_pt.y = max_response_pt.y - max_response_map.rows;
344 if (max_response_pt.x > max_response_map.cols / 2) // same for horizontal axis
345 max_response_pt.x = max_response_pt.x - max_response_map.cols;
348 if (m_use_subpixel_localization) {
349 new_location = sub_pixel_peak(max_response_pt, max_response_map);
351 new_location = max_response_pt;
353 DEBUG_PRINT(new_location);
356 void KCF_Tracker::track(cv::Mat &img)
359 if (m_debug) std::cout << "NEW FRAME" << '\n';
361 cv::Mat input_gray, input_rgb = img.clone();
362 if (img.channels() == 3) {
363 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
364 input_gray.convertTo(input_gray, CV_32FC1);
366 img.convertTo(input_gray, CV_32FC1);
368 // don't need too large image
369 resizeImgs(input_rgb, input_gray);
372 for (auto &it : d.threadctxs)
373 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
374 it.track(*this, input_rgb, input_gray);
376 for (auto const &it : d.threadctxs)
380 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
381 NORMAL_OMP_PARALLEL_FOR
382 for (uint i = 0; i < d.threadctxs.size(); ++i)
383 d.threadctxs[i].track(*this, input_rgb, input_gray);
386 cv::Point2f new_location;
388 findMaxReponse(max_idx, new_location);
390 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
391 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
393 clamp2(p_pose.cx, 0.0, (img.cols * p_scale_factor_x) - 1);
394 clamp2(p_pose.cy, 0.0, (img.rows * p_scale_factor_y) - 1);
396 clamp2(p_pose.cx, 0.0, img.cols - 1.0);
397 clamp2(p_pose.cy, 0.0, img.rows - 1.0);
400 // sub grid scale interpolation
401 if (m_use_subgrid_scale) {
402 p_current_scale *= sub_grid_scale(max_idx);
404 p_current_scale *= p_scales[max_idx];
407 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
409 // train at newly estimated target position
410 train(input_rgb, input_gray, p_interp_factor);
413 void ThreadCtx::track(const KCF_Tracker &kcf, cv::Mat &input_rgb, cv::Mat &input_gray)
415 // TODO: Move matrices to thread ctx
416 int sizes[3] = {kcf.p_num_of_feats, kcf.p_windows_size.height, kcf.p_windows_size.width};
417 MatDynMem patch_feats(3, sizes, CV_32FC1);
418 MatDynMem temp(3, sizes, CV_32FC1);
421 BIG_BATCH_OMP_PARALLEL_FOR
422 for (uint i = 0; i < kcf.p_num_scales; ++i)
425 kcf.get_features(patch_feats, input_rgb, input_gray, kcf.p_pose.cx, kcf.p_pose.cy,
426 kcf.p_windows_size.width, kcf.p_windows_size.height,
427 kcf.p_current_scale * IF_BIG_BATCH(kcf.p_scales[i], scale));
430 kcf.fft.forward_window(patch_feats, zf, temp);
433 if (kcf.m_use_linearkernel) {
434 kzf = zf.mul(kcf.p_model_alphaf).sum_over_channels();
436 gaussian_correlation(kcf, kzf, zf, kcf.p_model_xf, kcf.p_kernel_sigma);
437 DEBUG_PRINTM(kcf.p_model_alphaf);
439 kzf = kzf.mul(kcf.p_model_alphaf);
441 kcf.fft.inverse(kzf, response);
443 DEBUG_PRINTM(response);
445 /* target location is at the maximum response. we must take into
446 account the fact that, if the target doesn't move, the peak
447 will appear at the top-left corner, not at the center (this is
448 discussed in the paper). the responses wrap around cyclically. */
449 double min_val, max_val;
450 cv::Point2i min_loc, max_loc;
452 for (size_t i = 0; i < kcf.p_scales.size(); ++i) {
453 cv::minMaxLoc(response.plane(i), &min_val, &max_val, &min_loc, &max_loc);
454 DEBUG_PRINT(max_loc);
455 double weight = kcf.p_scales[i] < 1. ? kcf.p_scales[i] : 1. / kcf.p_scales[i];
456 max[i].response = max_val * weight;
457 max[i].loc = max_loc;
460 cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
462 DEBUG_PRINT(max_loc);
464 double weight = scale < 1. ? scale : 1. / scale;
465 max.response = max_val * weight;
470 // ****************************************************************************
472 void KCF_Tracker::get_features(MatDynMem &result_3d, cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy,
473 int size_x, int size_y, double scale) const
475 assert(result_3d.size[0] == p_num_of_feats);
476 assert(result_3d.size[1] == size_y / p_cell_size);
477 assert(result_3d.size[2] == size_x / p_cell_size);
479 int size_x_scaled = floor(size_x * scale);
480 int size_y_scaled = floor(size_y * scale);
482 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
483 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
485 // resize to default size
487 // if we downsample use INTER_AREA interpolation
488 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
490 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
493 // get hog(Histogram of Oriented Gradients) features
494 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
496 // get color rgb features (simple r,g,b channels)
497 std::vector<cv::Mat> color_feat;
498 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
499 // resize to default size
501 // if we downsample use INTER_AREA interpolation
502 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_AREA);
504 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_LINEAR);
508 if (m_use_color && input_rgb.channels() == 3) {
509 // use rgb color space
510 cv::Mat patch_rgb_norm;
511 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
512 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
513 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
514 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
515 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
516 cv::split(patch_rgb_norm, rgb);
517 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
520 if (m_use_cnfeat && input_rgb.channels() == 3) {
521 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
522 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
525 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
527 for (uint i = 0; i < hog_feat.size(); ++i) {
528 cv::Mat result_plane(result_3d.dims - 1, result_3d.size + 1, result_3d.cv::Mat::type(), result_3d.ptr<void>(i));
529 hog_feat[i].copyTo(result_plane);
533 MatDynMem KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
535 MatDynMem labels(dim2, dim1, CV_32FC1);
536 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
537 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
539 double sigma_s = sigma * sigma;
541 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
542 float *row_ptr = labels.ptr<float>(j);
544 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
545 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
549 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
550 MatDynMem rot_labels = circshift(labels, range_x[0], range_y[0]);
551 // sanity check, 1 at top left corner
552 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
557 MatDynMem KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
559 MatDynMem rot_patch(patch.size(), CV_32FC1);
560 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
562 // circular rotate x-axis
564 // move part that does not rotate over the edge
565 cv::Range orig_range(-x_rot, patch.cols);
566 cv::Range rot_range(0, patch.cols - (-x_rot));
567 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
570 orig_range = cv::Range(0, -x_rot);
571 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
572 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
573 } else if (x_rot > 0) {
574 // move part that does not rotate over the edge
575 cv::Range orig_range(0, patch.cols - x_rot);
576 cv::Range rot_range(x_rot, patch.cols);
577 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
580 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
581 rot_range = cv::Range(0, x_rot);
582 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
583 } else { // zero rotation
584 // move part that does not rotate over the edge
585 cv::Range orig_range(0, patch.cols);
586 cv::Range rot_range(0, patch.cols);
587 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
590 // circular rotate y-axis
592 // move part that does not rotate over the edge
593 cv::Range orig_range(-y_rot, patch.rows);
594 cv::Range rot_range(0, patch.rows - (-y_rot));
595 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
598 orig_range = cv::Range(0, -y_rot);
599 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
600 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
601 } else if (y_rot > 0) {
602 // move part that does not rotate over the edge
603 cv::Range orig_range(0, patch.rows - y_rot);
604 cv::Range rot_range(y_rot, patch.rows);
605 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
608 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
609 rot_range = cv::Range(0, y_rot);
610 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
611 } else { // zero rotation
612 // move part that does not rotate over the edge
613 cv::Range orig_range(0, patch.rows);
614 cv::Range rot_range(0, patch.rows);
615 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
621 // hann window actually (Power-of-cosine windows)
622 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
624 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
625 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
626 for (int i = 0; i < dim1; ++i)
627 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
628 N_inv = 1. / (static_cast<double>(dim2) - 1.);
629 for (int i = 0; i < dim2; ++i)
630 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
631 cv::Mat ret = m2 * m1;
635 // Returns sub-window of image input centered at [cx, cy] coordinates),
636 // with size [width, height]. If any pixels are outside of the image,
637 // they will replicate the values at the borders.
638 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height) const
642 int x1 = cx - width / 2;
643 int y1 = cy - height / 2;
644 int x2 = cx + width / 2;
645 int y2 = cy + height / 2;
648 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
649 patch.create(height, width, input.type());
650 patch.setTo(double(0.f));
654 int top = 0, bottom = 0, left = 0, right = 0;
656 // fit to image coordinates, set border extensions;
665 if (x2 >= input.cols) {
666 right = x2 - input.cols + width % 2;
671 if (y2 >= input.rows) {
672 bottom = y2 - input.rows + height % 2;
677 if (x2 - x1 == 0 || y2 - y1 == 0)
678 patch = cv::Mat::zeros(height, width, CV_32FC1);
680 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
681 cv::BORDER_REPLICATE);
682 // imshow( "copyMakeBorder", patch);
687 assert(patch.cols == width && patch.rows == height);
692 void KCF_Tracker::GaussianCorrelation::operator()(const KCF_Tracker &kcf, ComplexMat &result, const ComplexMat &xf,
693 const ComplexMat &yf, double sigma, bool auto_correlation)
695 xf.sqr_norm(xf_sqr_norm);
696 if (auto_correlation) {
697 yf_sqr_norm = xf_sqr_norm;
699 yf.sqr_norm(yf_sqr_norm);
701 xyf = auto_correlation ? xf.sqr_mag() : xf.mul(yf.conj());
703 kcf.fft.inverse(xyf, ifft_res);
705 cuda_gaussian_correlation(ifft_res.deviceMem(), k.deviceMem(), xf_sqr_norm.deviceMem(),
706 auto_correlation ? xf_sqr_norm.deviceMem() : yf_sqr_norm.deviceMem(), sigma,
707 xf.n_channels, xf.n_scales, kcf.p_roi.height, kcf.p_roi.width);
709 // ifft2 and sum over 3rd dimension, we dont care about individual channels
710 //DEBUG_PRINTM(ifft_res);
712 if (xf.channels() != kcf.p_num_scales * kcf.p_num_of_feats)
713 xy_sum.create(ifft_res.size(), CV_32FC1);
715 xy_sum.create(ifft_res.size(), CV_32FC(kcf.p_scales.size()));
717 for (int y = 0; y < ifft_res.rows; ++y) {
718 float *row_ptr = ifft_res.ptr<float>(y);
719 float *row_ptr_sum = xy_sum.ptr<float>(y);
720 for (int x = 0; x < ifft_res.cols; ++x) {
721 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
722 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
723 row_ptr + x * ifft_res.channels() + sum_ch * (ifft_res.channels() / xy_sum.channels()),
724 (row_ptr + x * ifft_res.channels() +
725 (sum_ch + 1) * (ifft_res.channels() / xy_sum.channels())),
730 DEBUG_PRINTM(xy_sum);
732 std::vector<cv::Mat> scales;
733 cv::split(xy_sum, scales);
735 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
736 for (uint i = 0; i < xf.n_scales; ++i) {
737 cv::Mat k_roi(k, cv::Rect(0, i * scales[0].rows, scales[0].cols, scales[0].rows));
738 cv::exp(-1. / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0),
743 kcf.fft.forward(k, result);
746 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
750 if (x < 0) x = response.cols + x;
751 if (y < 0) y = response.rows + y;
752 if (x >= response.cols) x = x - response.cols;
753 if (y >= response.rows) y = y - response.rows;
755 return response.at<float>(y, x);
758 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const
760 // find neighbourhood of max_loc (response is circular)
764 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);
765 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
766 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);
769 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
770 cv::Mat A = (cv::Mat_<float>(9, 6) <<
771 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
772 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
773 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
774 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
775 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
776 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
777 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
778 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
779 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);
780 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
781 get_response_circular(p1, response),
782 get_response_circular(p2, response),
783 get_response_circular(p3, response),
784 get_response_circular(p4, response),
785 get_response_circular(p5, response),
786 get_response_circular(p6, response),
787 get_response_circular(p7, response),
788 get_response_circular(p8, response),
789 get_response_circular(max_loc, response));
792 cv::solve(A, fval, x, cv::DECOMP_SVD);
794 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);
796 cv::Point2f sub_peak(max_loc.x, max_loc.y);
797 if (b > 0 || b < 0) {
798 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
799 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
805 double KCF_Tracker::sub_grid_scale(uint index)
808 if (index >= p_scales.size()) {
809 // interpolate from all values
810 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
811 A.create(p_scales.size(), 3, CV_32FC1);
812 fval.create(p_scales.size(), 1, CV_32FC1);
813 for (size_t i = 0; i < p_scales.size(); ++i) {
814 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
815 A.at<float>(i, 1) = float(p_scales[i]);
816 A.at<float>(i, 2) = 1;
817 fval.at<float>(i) = d.threadctxs.back().IF_BIG_BATCH(max[i].response, max.response);
820 // only from neighbours
821 if (index == 0 || index == p_scales.size() - 1)
822 return p_scales[index];
824 A = (cv::Mat_<float>(3, 3) <<
825 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
826 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
827 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
829 fval = (cv::Mat_<float>(3, 1) <<
830 d.threadctxs.back().max[index - 1].response,
831 d.threadctxs.back().max[index + 0].response,
832 d.threadctxs.back().max[index + 1].response);
834 fval = (cv::Mat_<float>(3, 1) <<
835 d.threadctxs[index - 1].max.response,
836 d.threadctxs[index + 0].max.response,
837 d.threadctxs[index + 1].max.response);
842 cv::solve(A, fval, x, cv::DECOMP_SVD);
843 float a = x.at<float>(0), b = x.at<float>(1);
844 double scale = p_scales[index];
846 scale = -b / (2 * a);