5 #include "threadctx.hpp"
15 #include "fft_cufft.h"
18 #include "fft_opencv.h"
29 std::ios::fmtflags flags;
30 std::streamsize precision;
33 IOSave( std::ios& userStream )
34 : stream( userStream )
35 , flags( userStream.flags() )
36 , precision( userStream.precision() )
37 , fill( userStream.fill() )
42 stream.flags( flags );
43 stream.precision( precision );
54 std::string indent() { return std::string(indentLvl * 4, ' '); }
61 FTrace(DbgTracer &dt, const char *fn, const char *format, ...) : t(dt), funcName(fn)
67 if (-1 == vasprintf(&arg, format, vl))
68 throw std::runtime_error("vasprintf error");
71 std::cerr << t.indent() << funcName << "(" << arg << ") {" << std::endl;
78 std::cerr << t.indent() << "}" << std::endl;
83 void traceVal(const char *name, const T& obj, int line)
87 std::cerr << indent() << name /*<< " @" << line */ << " " << print(obj) << std::endl;
90 template <typename T> struct Printer {
92 Printer(const T &_obj) : obj(_obj) {}
95 template <typename T> Printer<T> print(const T& obj) { return Printer<T>(obj); }
96 Printer<cv::Mat> print(const MatScales& obj) { return Printer<cv::Mat>(obj); }
97 Printer<cv::Mat> print(const MatFeats& obj) { return Printer<cv::Mat>(obj); }
98 Printer<cv::Mat> print(const MatScaleFeats& obj) { return Printer<cv::Mat>(obj); }
101 template <typename T>
102 std::ostream &operator<<(std::ostream &os, const DbgTracer::Printer<T> &p) {
106 std::ostream &operator<<(std::ostream &os, const DbgTracer::Printer<cv::Mat> &p) {
108 os << std::setprecision(3);
109 os << p.obj.size << " " << p.obj.channels() << "ch " << static_cast<const void*>(p.obj.data);
111 constexpr size_t num = 10;
112 for (size_t i = 0; i < std::min(num, p.obj.total()); ++i)
113 os << *p.obj.ptr<float>(i) << ", ";
114 os << (num < p.obj.total() ? "... ]" : "]");
118 std::ostream &operator<<(std::ostream &os, const cufftComplex &p) {
124 std::ostream &operator<<(std::ostream &os, const DbgTracer::Printer<ComplexMat> &p) {
126 os << std::setprecision(3);
127 os << "<cplx> " << p.obj.size() << " " << p.obj.channels() << "ch " << p.obj.get_p_data();
129 constexpr int num = 10;
130 for (int i = 0; i < std::min(num, p.obj.size().area()); ++i)
131 os << p.obj.get_p_data()[i] << ", ";
132 os << (num < p.obj.size().area() ? "... ]" : "]");
136 DbgTracer __dbgTracer;
138 #define TRACE(...) const DbgTracer::FTrace __tracer(__dbgTracer, __PRETTY_FUNCTION__, ##__VA_ARGS__)
140 #define DEBUG_PRINT(obj) __dbgTracer.traceVal(#obj, (obj), __LINE__)
141 #define DEBUG_PRINTM(obj) DEBUG_PRINT(obj)
144 template <typename T>
145 T clamp(const T& n, const T& lower, const T& upper)
147 return std::max(lower, std::min(n, upper));
150 template <typename T>
151 void clamp2(T& n, const T& lower, const T& upper)
153 n = std::max(lower, std::min(n, upper));
156 class Kcf_Tracker_Private {
158 std::vector<ThreadCtx> threadctxs;
161 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
162 double output_sigma_factor, int cell_size)
163 : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
164 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size), d(*new Kcf_Tracker_Private)
168 KCF_Tracker::KCF_Tracker() : fft(*new FFT()), d(*new Kcf_Tracker_Private) {}
170 KCF_Tracker::~KCF_Tracker()
176 void KCF_Tracker::train(cv::Mat input_gray, cv::Mat input_rgb, double interp_factor)
178 // obtain a sub-window for training
179 // TODO: Move Mats outside from here
180 MatScaleFeats patch_feats(1, p_num_of_feats, p_roi);
181 MatScaleFeats temp(1, p_num_of_feats, p_roi);
182 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy,
183 p_windows_size.width, p_windows_size.height,
184 p_current_scale).copyTo(patch_feats.features(0));
186 fft.forward_window(patch_feats, p_xf, temp);
187 p_model_xf = p_model_xf * (1. - interp_factor) + p_xf * interp_factor;
188 DEBUG_PRINTM(p_model_xf);
190 ComplexMat alphaf_num, alphaf_den;
192 if (m_use_linearkernel) {
193 ComplexMat xfconj = p_xf.conj();
194 alphaf_num = xfconj.mul(p_yf);
195 alphaf_den = (p_xf * xfconj);
197 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
198 const uint num_scales = BIG_BATCH_MODE ? p_num_scales : 1;
199 cv::Size sz(Fft::freq_size(p_roi));
200 ComplexMat kf(sz.height, sz.width, num_scales);
201 (*gaussian_correlation)(*this, kf, p_model_xf, p_model_xf, p_kernel_sigma, true);
203 p_model_alphaf_num = p_yf * kf;
204 DEBUG_PRINTM(p_model_alphaf_num);
205 p_model_alphaf_den = kf * (kf + p_lambda);
206 DEBUG_PRINTM(p_model_alphaf_den);
208 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
209 DEBUG_PRINTM(p_model_alphaf);
210 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
213 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
215 __dbgTracer.debug = m_debug;
218 // check boundary, enforce min size
219 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
221 if (x2 > img.cols - 1) x2 = img.cols - 1;
223 if (y2 > img.rows - 1) y2 = img.rows - 1;
225 if (x2 - x1 < 2 * p_cell_size) {
226 double diff = (2 * p_cell_size - x2 + x1) / 2.;
227 if (x1 - diff >= 0 && x2 + diff < img.cols) {
230 } else if (x1 - 2 * diff >= 0) {
236 if (y2 - y1 < 2 * p_cell_size) {
237 double diff = (2 * p_cell_size - y2 + y1) / 2.;
238 if (y1 - diff >= 0 && y2 + diff < img.rows) {
241 } else if (y1 - 2 * diff >= 0) {
250 p_pose.cx = x1 + p_pose.w / 2.;
251 p_pose.cy = y1 + p_pose.h / 2.;
253 cv::Mat input_gray, input_rgb = img.clone();
254 if (img.channels() == 3) {
255 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
256 input_gray.convertTo(input_gray, CV_32FC1);
258 img.convertTo(input_gray, CV_32FC1);
260 // don't need too large image
261 if (p_pose.w * p_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
262 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
263 p_resize_image = true;
264 p_pose.scale(p_downscale_factor);
265 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
266 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
267 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
268 if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
269 std::cerr << "Error: Fit size is not multiple of HOG cell size (" << p_cell_size << ")" << std::endl;
270 std::exit(EXIT_FAILURE);
272 p_scale_factor_x = (double)fit_size_x / round(p_pose.w * (1. + p_padding));
273 p_scale_factor_y = (double)fit_size_y / round(p_pose.h * (1. + p_padding));
274 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
275 << p_scale_factor_y << std::endl;
277 p_pose.scale_x(p_scale_factor_x);
278 p_pose.scale_y(p_scale_factor_y);
279 if (fabs(p_scale_factor_x - 1) > p_floating_error || fabs(p_scale_factor_y - 1) > p_floating_error) {
280 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
281 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
282 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
284 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
285 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
290 // compute win size + fit to fhog cell size
291 p_windows_size.width = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
292 p_windows_size.height = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
293 p_roi.width = p_windows_size.width / p_cell_size;
294 p_roi.height = p_windows_size.height / p_cell_size;
298 for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
299 p_scales.push_back(std::pow(p_scale_step, i));
301 p_scales.push_back(1.);
304 if (p_roi.height * (p_roi.width / 2 + 1) > 1024) {
305 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
306 "the window dimensions so its size is less or equal to "
307 << 1024 * p_cell_size * p_cell_size * 2 + 1
308 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
309 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
310 std::exit(EXIT_FAILURE);
313 if (m_use_linearkernel) {
314 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
315 std::exit(EXIT_FAILURE);
317 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
319 p_xf.create(p_roi.height, p_roi.height / 2 + 1, p_num_of_feats);
322 #if defined(CUFFT) || defined(FFTW)
323 uint width = p_roi.width / 2 + 1;
325 uint width = p_roi.width;
327 p_model_xf.create(p_roi.height, width, p_num_of_feats);
328 p_yf.create(p_roi.height, width, 1);
329 p_xf.create(p_roi.height, width, p_num_of_feats);
332 for (auto scale: p_scales)
333 d.threadctxs.emplace_back(p_roi, p_num_of_feats, scale);
335 d.threadctxs.emplace_back(p_roi, p_num_of_feats, p_num_scales);
338 gaussian_correlation.reset(
339 new GaussianCorrelation(IF_BIG_BATCH(p_num_scales, 1), p_roi));
341 p_current_scale = 1.;
343 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
344 double max_size_ratio =
345 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
346 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
347 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
348 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
350 std::cout << "init: img size " << img.cols << "x" << img.rows << std::endl;
351 std::cout << "init: win size " << p_windows_size.width << "x" << p_windows_size.height << std::endl;
352 std::cout << "init: FFT size " << p_roi.width << "x" << p_roi.height << std::endl;
353 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
355 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / p_cell_size;
357 fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales);
358 fft.set_window(MatDynMem(cosine_window_function(p_roi.width, p_roi.height)));
360 // window weights, i.e. labels
361 MatScales gsl(1, p_roi);
362 gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height).copyTo(gsl.plane(0));
363 fft.forward(gsl, p_yf);
366 // train initial model
367 train(input_gray, input_rgb, 1.0);
370 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
372 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
375 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
377 if (p_resize_image) {
379 tmp.scale(p_downscale_factor);
382 } else if (p_fit_to_pw2) {
384 tmp.scale_x(p_scale_factor_x);
385 tmp.scale_y(p_scale_factor_y);
394 BBox_c KCF_Tracker::getBBox()
397 tmp.w *= p_current_scale;
398 tmp.h *= p_current_scale;
400 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
402 tmp.scale_x(1 / p_scale_factor_x);
403 tmp.scale_y(1 / p_scale_factor_y);
409 double KCF_Tracker::getFilterResponse() const
411 return this->max_response;
414 void KCF_Tracker::resizeImgs(cv::Mat &input_rgb, cv::Mat &input_gray)
416 if (p_resize_image) {
417 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
418 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
419 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
420 fabs(p_scale_factor_y - 1) > p_floating_error) {
421 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
422 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
423 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
425 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
426 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
431 void KCF_Tracker::findMaxReponse(uint &max_idx, cv::Point2f &new_location) const
435 for (uint j = 0; j < d.threadctxs.size(); ++j) {
436 if (d.threadctxs[j].max.response > max) {
437 max = d.threadctxs[j].max.response;
442 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
443 for (uint j = 0; j < p_scales.size(); ++j) {
444 if (d.threadctxs[0].max[j].response > max) {
445 max = d.threadctxs[0].max[j].response;
450 cv::Point2i &max_response_pt = IF_BIG_BATCH(d.threadctxs[0].max[max_idx].loc, d.threadctxs[max_idx].max.loc);
451 cv::Mat max_response_map = IF_BIG_BATCH(d.threadctxs[0].response.plane(max_idx), d.threadctxs[max_idx].response.plane(0));
453 DEBUG_PRINTM(max_response_map);
454 DEBUG_PRINT(max_response_pt);
456 // sub pixel quadratic interpolation from neighbours
457 if (max_response_pt.y > max_response_map.rows / 2) // wrap around to negative half-space of vertical axis
458 max_response_pt.y = max_response_pt.y - max_response_map.rows;
459 if (max_response_pt.x > max_response_map.cols / 2) // same for horizontal axis
460 max_response_pt.x = max_response_pt.x - max_response_map.cols;
463 if (m_use_subpixel_localization) {
464 new_location = sub_pixel_peak(max_response_pt, max_response_map);
466 new_location = max_response_pt;
468 DEBUG_PRINT(new_location);
471 void KCF_Tracker::track(cv::Mat &img)
473 __dbgTracer.debug = m_debug;
476 cv::Mat input_gray, input_rgb = img.clone();
477 if (img.channels() == 3) {
478 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
479 input_gray.convertTo(input_gray, CV_32FC1);
481 img.convertTo(input_gray, CV_32FC1);
483 // don't need too large image
484 resizeImgs(input_rgb, input_gray);
487 for (auto &it : d.threadctxs)
488 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
489 it.track(*this, input_rgb, input_gray);
491 for (auto const &it : d.threadctxs)
495 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
496 NORMAL_OMP_PARALLEL_FOR
497 for (uint i = 0; i < d.threadctxs.size(); ++i)
498 d.threadctxs[i].track(*this, input_rgb, input_gray);
501 cv::Point2f new_location;
503 findMaxReponse(max_idx, new_location);
505 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
506 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
508 clamp2(p_pose.cx, 0.0, (img.cols * p_scale_factor_x) - 1);
509 clamp2(p_pose.cy, 0.0, (img.rows * p_scale_factor_y) - 1);
511 clamp2(p_pose.cx, 0.0, img.cols - 1.0);
512 clamp2(p_pose.cy, 0.0, img.rows - 1.0);
515 // sub grid scale interpolation
516 if (m_use_subgrid_scale) {
517 p_current_scale *= sub_grid_scale(max_idx);
519 p_current_scale *= p_scales[max_idx];
522 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
524 // train at newly estimated target position
525 train(input_rgb, input_gray, p_interp_factor);
528 void ThreadCtx::track(const KCF_Tracker &kcf, cv::Mat &input_rgb, cv::Mat &input_gray)
532 BIG_BATCH_OMP_PARALLEL_FOR
533 for (uint i = 0; i < IF_BIG_BATCH(kcf.p_num_scales, 1); ++i)
535 kcf.get_features(input_rgb, input_gray, kcf.p_pose.cx, kcf.p_pose.cy,
536 kcf.p_windows_size.width, kcf.p_windows_size.height,
537 kcf.p_current_scale * IF_BIG_BATCH(kcf.p_scales[i], scale))
538 .copyTo(patch_feats.features(i));
539 DEBUG_PRINT(patch_feats.features(i));
542 DEBUG_PRINT(patch_feats);
543 kcf.fft.forward_window(patch_feats, zf, temp);
546 if (kcf.m_use_linearkernel) {
547 kzf = zf.mul(kcf.p_model_alphaf).sum_over_channels();
549 gaussian_correlation(kcf, kzf, zf, kcf.p_model_xf, kcf.p_kernel_sigma);
550 DEBUG_PRINTM(kcf.p_model_alphaf);
552 kzf = kzf.mul(kcf.p_model_alphaf);
555 kcf.fft.inverse(kzf, response);
557 DEBUG_PRINTM(response);
559 /* target location is at the maximum response. we must take into
560 account the fact that, if the target doesn't move, the peak
561 will appear at the top-left corner, not at the center (this is
562 discussed in the paper). the responses wrap around cyclically. */
563 double min_val, max_val;
564 cv::Point2i min_loc, max_loc;
566 for (size_t i = 0; i < kcf.p_scales.size(); ++i) {
567 cv::minMaxLoc(response.plane(i), &min_val, &max_val, &min_loc, &max_loc);
568 DEBUG_PRINT(max_loc);
569 double weight = kcf.p_scales[i] < 1. ? kcf.p_scales[i] : 1. / kcf.p_scales[i];
570 max[i].response = max_val * weight;
571 max[i].loc = max_loc;
574 cv::minMaxLoc(response.plane(0), &min_val, &max_val, &min_loc, &max_loc);
576 DEBUG_PRINT(max_loc);
578 double weight = scale < 1. ? scale : 1. / scale;
579 max.response = max_val * weight;
584 // ****************************************************************************
586 cv::Mat KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy,
587 int size_x, int size_y, double scale) const
589 int size_x_scaled = floor(size_x * scale);
590 int size_y_scaled = floor(size_y * scale);
592 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
593 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
595 // resize to default size
597 // if we downsample use INTER_AREA interpolation
598 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
600 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
603 // get hog(Histogram of Oriented Gradients) features
604 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
606 // get color rgb features (simple r,g,b channels)
607 std::vector<cv::Mat> color_feat;
608 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
609 // resize to default size
611 // if we downsample use INTER_AREA interpolation
612 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_AREA);
614 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_LINEAR);
618 if (m_use_color && input_rgb.channels() == 3) {
619 // use rgb color space
620 cv::Mat patch_rgb_norm;
621 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
622 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
623 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
624 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
625 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
626 cv::split(patch_rgb_norm, rgb);
627 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
630 if (m_use_cnfeat && input_rgb.channels() == 3) {
631 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
632 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
635 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
637 int size[] = {p_num_of_feats, p_roi.height, p_roi.width};
638 cv::Mat result(3, size, CV_32F);
639 for (uint i = 0; i < hog_feat.size(); ++i)
640 hog_feat[i].copyTo(cv::Mat(size[1], size[2], CV_32FC1, result.ptr(i)));
645 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
647 cv::Mat labels(dim2, dim1, CV_32FC1);
648 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
649 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
651 double sigma_s = sigma * sigma;
653 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
654 float *row_ptr = labels.ptr<float>(j);
656 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
657 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
661 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
662 MatDynMem rot_labels = circshift(labels, range_x[0], range_y[0]);
663 // sanity check, 1 at top left corner
664 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
669 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
671 cv::Mat rot_patch(patch.size(), CV_32FC1);
672 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
674 // circular rotate x-axis
676 // move part that does not rotate over the edge
677 cv::Range orig_range(-x_rot, patch.cols);
678 cv::Range rot_range(0, patch.cols - (-x_rot));
679 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
682 orig_range = cv::Range(0, -x_rot);
683 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
684 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
685 } else if (x_rot > 0) {
686 // move part that does not rotate over the edge
687 cv::Range orig_range(0, patch.cols - x_rot);
688 cv::Range rot_range(x_rot, patch.cols);
689 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
692 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
693 rot_range = cv::Range(0, x_rot);
694 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
695 } else { // zero rotation
696 // move part that does not rotate over the edge
697 cv::Range orig_range(0, patch.cols);
698 cv::Range rot_range(0, patch.cols);
699 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
702 // circular rotate y-axis
704 // move part that does not rotate over the edge
705 cv::Range orig_range(-y_rot, patch.rows);
706 cv::Range rot_range(0, patch.rows - (-y_rot));
707 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
710 orig_range = cv::Range(0, -y_rot);
711 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
712 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
713 } else if (y_rot > 0) {
714 // move part that does not rotate over the edge
715 cv::Range orig_range(0, patch.rows - y_rot);
716 cv::Range rot_range(y_rot, patch.rows);
717 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
720 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
721 rot_range = cv::Range(0, y_rot);
722 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
723 } else { // zero rotation
724 // move part that does not rotate over the edge
725 cv::Range orig_range(0, patch.rows);
726 cv::Range rot_range(0, patch.rows);
727 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
733 // hann window actually (Power-of-cosine windows)
734 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
736 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
737 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
738 for (int i = 0; i < dim1; ++i)
739 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
740 N_inv = 1. / (static_cast<double>(dim2) - 1.);
741 for (int i = 0; i < dim2; ++i)
742 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
743 cv::Mat ret = m2 * m1;
747 // Returns sub-window of image input centered at [cx, cy] coordinates),
748 // with size [width, height]. If any pixels are outside of the image,
749 // they will replicate the values at the borders.
750 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height) const
754 int x1 = cx - width / 2;
755 int y1 = cy - height / 2;
756 int x2 = cx + width / 2;
757 int y2 = cy + height / 2;
760 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
761 patch.create(height, width, input.type());
762 patch.setTo(double(0.f));
766 int top = 0, bottom = 0, left = 0, right = 0;
768 // fit to image coordinates, set border extensions;
777 if (x2 >= input.cols) {
778 right = x2 - input.cols + width % 2;
783 if (y2 >= input.rows) {
784 bottom = y2 - input.rows + height % 2;
789 if (x2 - x1 == 0 || y2 - y1 == 0)
790 patch = cv::Mat::zeros(height, width, CV_32FC1);
792 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
793 cv::BORDER_REPLICATE);
794 // imshow( "copyMakeBorder", patch);
799 assert(patch.cols == width && patch.rows == height);
804 void KCF_Tracker::GaussianCorrelation::operator()(const KCF_Tracker &kcf, ComplexMat &result, const ComplexMat &xf,
805 const ComplexMat &yf, double sigma, bool auto_correlation)
808 xf.sqr_norm(xf_sqr_norm);
809 if (auto_correlation) {
810 yf_sqr_norm = xf_sqr_norm;
812 yf.sqr_norm(yf_sqr_norm);
814 xyf = auto_correlation ? xf.sqr_mag() : xf * yf.conj(); // xf.muln(yf.conj());
816 kcf.fft.inverse(xyf, ifft_res);
818 cuda_gaussian_correlation(ifft_res.deviceMem(), k.deviceMem(), xf_sqr_norm.deviceMem(),
819 auto_correlation ? xf_sqr_norm.deviceMem() : yf_sqr_norm.deviceMem(), sigma,
820 xf.n_channels, xf.n_scales, kcf.p_roi.height, kcf.p_roi.width);
822 // ifft2 and sum over 3rd dimension, we dont care about individual channels
823 //DEBUG_PRINTM(ifft_res);
825 if (xf.channels() != kcf.p_num_scales * kcf.p_num_of_feats)
826 xy_sum.create(ifft_res.size(), CV_32FC1);
828 xy_sum.create(ifft_res.size(), CV_32FC(kcf.p_scales.size()));
830 for (int y = 0; y < ifft_res.rows; ++y) {
831 float *row_ptr = ifft_res.ptr<float>(y);
832 float *row_ptr_sum = xy_sum.ptr<float>(y);
833 for (int x = 0; x < ifft_res.cols; ++x) {
834 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
835 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
836 row_ptr + x * ifft_res.channels() + sum_ch * (ifft_res.channels() / xy_sum.channels()),
837 (row_ptr + x * ifft_res.channels() +
838 (sum_ch + 1) * (ifft_res.channels() / xy_sum.channels())),
843 DEBUG_PRINTM(xy_sum);
845 std::vector<cv::Mat> scales;
846 cv::split(xy_sum, scales);
848 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
849 for (uint i = 0; i < xf.n_scales; ++i) {
850 cv::Mat k_roi = k.plane(i);
851 cv::exp(-1. / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0),
856 kcf.fft.forward(k, result);
859 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
863 if (x < 0) x = response.cols + x;
864 if (y < 0) y = response.rows + y;
865 if (x >= response.cols) x = x - response.cols;
866 if (y >= response.rows) y = y - response.rows;
868 return response.at<float>(y, x);
871 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const
873 // find neighbourhood of max_loc (response is circular)
877 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);
878 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
879 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);
882 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
883 cv::Mat A = (cv::Mat_<float>(9, 6) <<
884 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
885 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
886 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
887 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
888 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
889 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
890 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
891 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
892 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);
893 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
894 get_response_circular(p1, response),
895 get_response_circular(p2, response),
896 get_response_circular(p3, response),
897 get_response_circular(p4, response),
898 get_response_circular(p5, response),
899 get_response_circular(p6, response),
900 get_response_circular(p7, response),
901 get_response_circular(p8, response),
902 get_response_circular(max_loc, response));
905 cv::solve(A, fval, x, cv::DECOMP_SVD);
907 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);
909 cv::Point2f sub_peak(max_loc.x, max_loc.y);
910 if (b > 0 || b < 0) {
911 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
912 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
918 double KCF_Tracker::sub_grid_scale(uint index)
921 if (index >= p_scales.size()) {
922 // interpolate from all values
923 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
924 A.create(p_scales.size(), 3, CV_32FC1);
925 fval.create(p_scales.size(), 1, CV_32FC1);
926 for (size_t i = 0; i < p_scales.size(); ++i) {
927 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
928 A.at<float>(i, 1) = float(p_scales[i]);
929 A.at<float>(i, 2) = 1;
930 fval.at<float>(i) = d.threadctxs.back().IF_BIG_BATCH(max[i].response, max.response);
933 // only from neighbours
934 if (index == 0 || index == p_scales.size() - 1)
935 return p_scales[index];
937 A = (cv::Mat_<float>(3, 3) <<
938 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
939 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
940 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
942 fval = (cv::Mat_<float>(3, 1) <<
943 d.threadctxs.back().max[index - 1].response,
944 d.threadctxs.back().max[index + 0].response,
945 d.threadctxs.back().max[index + 1].response);
947 fval = (cv::Mat_<float>(3, 1) <<
948 d.threadctxs[index - 1].max.response,
949 d.threadctxs[index + 0].max.response,
950 d.threadctxs[index + 1].max.response);
955 cv::solve(A, fval, x, cv::DECOMP_SVD);
956 float a = x.at<float>(0), b = x.at<float>(1);
957 double scale = p_scales[index];
959 scale = -b / (2 * a);