#include <algorithm>
#include "threadctx.hpp"
#include "debug.h"
+#include <limits>
#ifdef FFTW
#include "fft_fftw.h"
n = std::max(lower, std::min(n, upper));
}
+#if CV_MAJOR_VERSION < 3
+template<typename _Tp> static inline
+cv::Size_<_Tp> operator / (const cv::Size_<_Tp>& a, _Tp b)
+{
+ return cv::Size_<_Tp>(a.width / b, a.height / b);
+}
+#endif
+
class Kcf_Tracker_Private {
friend KCF_Tracker;
std::vector<ThreadCtx> threadctxs;
// obtain a sub-window for training
// TODO: Move Mats outside from here
- MatScaleFeats patch_feats(1, p_num_of_feats, p_roi);
- DEBUG_PRINT(patch_feats);
- MatScaleFeats temp(1, p_num_of_feats, p_roi);
+ MatScaleFeats patch_feats(1, p_num_of_feats, feature_size);
+ MatScaleFeats temp(1, p_num_of_feats, feature_size);
get_features(input_rgb, input_gray, p_current_center.x, p_current_center.y,
p_windows_size.width, p_windows_size.height,
p_current_scale).copyTo(patch_feats.scale(0));
DEBUG_PRINT(patch_feats);
- fft.forward_window(patch_feats, p_xf, temp);
- DEBUG_PRINTM(p_xf);
- p_model_xf = p_model_xf * (1. - interp_factor) + p_xf * interp_factor;
- DEBUG_PRINTM(p_model_xf);
-
- ComplexMat alphaf_num, alphaf_den;
+ fft.forward_window(patch_feats, model->xf, temp);
+ DEBUG_PRINTM(model->xf);
+ model->model_xf = model->model_xf * (1. - interp_factor) + model->xf * interp_factor;
+ DEBUG_PRINTM(model->model_xf);
if (m_use_linearkernel) {
- ComplexMat xfconj = p_xf.conj();
- alphaf_num = xfconj.mul(p_yf);
- alphaf_den = (p_xf * xfconj);
+ ComplexMat xfconj = model->xf.conj();
+ model->model_alphaf_num = xfconj.mul(model->yf);
+ model->model_alphaf_den = (model->xf * xfconj);
} else {
// Kernel Ridge Regression, calculate alphas (in Fourier domain)
- cv::Size sz(Fft::freq_size(p_roi));
+ cv::Size sz(Fft::freq_size(feature_size));
ComplexMat kf(sz.height, sz.width, 1);
- (*gaussian_correlation)(kf, p_model_xf, p_model_xf, p_kernel_sigma, true, *this);
+ (*gaussian_correlation)(kf, model->model_xf, model->model_xf, p_kernel_sigma, true, *this);
DEBUG_PRINTM(kf);
- p_model_alphaf_num = p_yf * kf;
- p_model_alphaf_den = kf * (kf + p_lambda);
+ model->model_alphaf_num = model->yf * kf;
+ model->model_alphaf_den = kf * (kf + p_lambda);
}
- p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
- DEBUG_PRINTM(p_model_alphaf);
+ model->model_alphaf = model->model_alphaf_num / model->model_alphaf_den;
+ DEBUG_PRINTM(model->model_alphaf);
// p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
}
+static int round_pw2_down(int x)
+{
+ for (int i = 1; i < 32; i <<= 1)
+ x |= x >> i;
+ x++;
+ return x >> 1;
+}
+
+
void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
{
__dbgTracer.debug = m_debug;
img.convertTo(input_gray, CV_32FC1);
// don't need too large image
- if (p_init_pose.w * p_init_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
+ if (p_init_pose.w * p_init_pose.h > 100. * 100.) {
std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
p_resize_image = true;
p_init_pose.scale(p_downscale_factor);
cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
- } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
- if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
- std::cerr << "Error: Fit size is not multiple of HOG cell size (" << p_cell_size << ")" << std::endl;
- std::exit(EXIT_FAILURE);
- }
- p_fit_factor_x = (double)fit_size_x / round(p_init_pose.w * (1. + p_padding));
- p_fit_factor_y = (double)fit_size_y / round(p_init_pose.h * (1. + p_padding));
- std::cout << "resizing image horizontaly by factor of " << p_fit_factor_x << " and verticaly by factor of "
- << p_fit_factor_y << std::endl;
- p_fit_to_pw2 = true;
- p_init_pose.scale_x(p_fit_factor_x);
- p_init_pose.scale_y(p_fit_factor_y);
- if (fabs(p_fit_factor_x - 1) > p_floating_error || fabs(p_fit_factor_y - 1) > p_floating_error) {
- if (p_fit_factor_x < 1 && p_fit_factor_y < 1) {
- cv::resize(input_gray, input_gray, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_AREA);
- cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_AREA);
- } else {
- cv::resize(input_gray, input_gray, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_LINEAR);
- cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_LINEAR);
- }
- }
}
-
// compute win size + fit to fhog cell size
p_windows_size.width = round(p_init_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
p_windows_size.height = round(p_init_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
- p_roi.width = p_windows_size.width / p_cell_size;
- p_roi.height = p_windows_size.height / p_cell_size;
+
+ if (fit_size_x == 0 || fit_size_y == 0) {
+ // Round down to the next highest power of 2
+ fit_size = cv::Size(round_pw2_down(p_windows_size.width),
+ round_pw2_down(p_windows_size.height));
+ } else if (fit_size_x == -1 || fit_size_y == -1) {
+ fit_size = p_windows_size;
+ } else {
+ fit_size = cv::Size(fit_size_x, fit_size_y);
+ }
+
+ feature_size = fit_size / p_cell_size;
p_scales.clear();
for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
p_scales.push_back(std::pow(p_scale_step, i));
#ifdef CUFFT
- if (p_roi.height * (p_roi.width / 2 + 1) > 1024) {
- std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
- "the window dimensions so its size is less or equal to "
- << 1024 * p_cell_size * p_cell_size * 2 + 1
- << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
- << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
- std::exit(EXIT_FAILURE);
- }
-
if (m_use_linearkernel) {
std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
std::exit(EXIT_FAILURE);
}
-#else
- p_xf.create(p_roi.height, p_roi.height / 2 + 1, p_num_of_feats);
#endif
-#if defined(CUFFT) || defined(FFTW)
- uint width = p_roi.width / 2 + 1;
-#else
- uint width = p_roi.width;
-#endif
- p_model_xf.create(p_roi.height, width, p_num_of_feats);
- p_yf.create(p_roi.height, width, 1);
- p_xf.create(p_roi.height, width, p_num_of_feats);
+ model.reset(new Model(Fft::freq_size(feature_size), p_num_of_feats));
#ifndef BIG_BATCH
for (auto scale: p_scales)
- d.threadctxs.emplace_back(p_roi, p_num_of_feats, scale);
+ d.threadctxs.emplace_back(feature_size, p_num_of_feats, scale);
#else
- d.threadctxs.emplace_back(p_roi, p_num_of_feats, p_num_scales);
+ d.threadctxs.emplace_back(feature_size, p_num_of_feats, p_num_scales);
#endif
- gaussian_correlation.reset(new GaussianCorrelation(1, p_roi));
+ gaussian_correlation.reset(new GaussianCorrelation(1, p_num_of_feats, feature_size));
p_current_center = p_init_pose.center();
p_current_scale = 1.;
p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
- std::cout << "init: img size " << img.cols << "x" << img.rows << std::endl;
- std::cout << "init: win size " << p_windows_size.width << "x" << p_windows_size.height << std::endl;
- std::cout << "init: FFT size " << p_roi.width << "x" << p_roi.height << std::endl;
+ std::cout << "init: img size " << img.size() << std::endl;
+ std::cout << "init: win size " << p_windows_size;
+ if (p_windows_size != fit_size)
+ std::cout << " resized to " << fit_size;
+ std::cout << std::endl;
+ std::cout << "init: FFT size " << feature_size << std::endl;
std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
- p_output_sigma = std::sqrt(p_init_pose.w * p_init_pose.h) * p_output_sigma_factor / p_cell_size;
+ p_output_sigma = std::sqrt(p_init_pose.w * p_init_pose.h * double(fit_size.area()) / p_windows_size.area())
+ * p_output_sigma_factor / p_cell_size;
- fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales);
- fft.set_window(MatDynMem(cosine_window_function(p_roi.width, p_roi.height)));
+ fft.init(feature_size.width, feature_size.height, p_num_of_feats, p_num_scales);
+ fft.set_window(MatDynMem(cosine_window_function(feature_size.width, feature_size.height)));
// window weights, i.e. labels
- MatScales gsl(1, p_roi);
- gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height).copyTo(gsl.plane(0));
- fft.forward(gsl, p_yf);
- DEBUG_PRINTM(p_yf);
+ MatScales gsl(1, feature_size);
+ gaussian_shaped_labels(p_output_sigma, feature_size.width, feature_size.height).copyTo(gsl.plane(0));
+ fft.forward(gsl, model->yf);
+ DEBUG_PRINTM(model->yf);
// train initial model
train(input_rgb, input_gray, 1.0);
BBox_c tmp = bbox;
if (p_resize_image) {
tmp.scale(p_downscale_factor);
- } else if (p_fit_to_pw2) {
- tmp.scale_x(p_fit_factor_x);
- tmp.scale_y(p_fit_factor_y);
}
p_current_center = tmp.center();
}
if (p_resize_image)
tmp.scale(1 / p_downscale_factor);
- if (p_fit_to_pw2) {
- tmp.scale_x(1 / p_fit_factor_x);
- tmp.scale_y(1 / p_fit_factor_y);
- }
return tmp;
}
if (p_resize_image) {
cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
- } else if (p_fit_to_pw2 && fabs(p_fit_factor_x - 1) > p_floating_error &&
- fabs(p_fit_factor_y - 1) > p_floating_error) {
- if (p_fit_factor_x < 1 && p_fit_factor_y < 1) {
- cv::resize(input_gray, input_gray, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_AREA);
- cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_AREA);
- } else {
- cv::resize(input_gray, input_gray, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_LINEAR);
- cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_fit_factor_x, p_fit_factor_y, cv::INTER_LINEAR);
- }
}
}
double KCF_Tracker::findMaxReponse(uint &max_idx, cv::Point2d &new_location) const
{
double max = -1.;
+ max_idx = std::numeric_limits<uint>::max();
+
#ifndef BIG_BATCH
for (uint j = 0; j < d.threadctxs.size(); ++j) {
if (d.threadctxs[j].max.response > max) {
}
}
#endif
+ assert(max_idx < IF_BIG_BATCH(p_scales.size(), d.threadctxs.size()));
+
+ if (m_visual_debug) {
+ int w = 100; //feature_size.width;
+ int h = 100; //feature_size.height;
+ cv::Mat all_responses(h * p_num_scales, w * p_num_angles,
+ d.threadctxs[0].response.type(), cv::Scalar::all(0));
+ for (size_t i = 0; i < p_num_scales; ++i) {
+ for (size_t j = 0; j < p_num_angles; ++j) {
+ cv::Mat tmp = d.threadctxs[IF_BIG_BATCH(0, p_num_angles * i + j)].response.plane(IF_BIG_BATCH(p_num_angles * i + j, 0));
+ tmp = circshift(tmp, -tmp.cols/2, -tmp.rows/2);
+ cv::resize(tmp, tmp, cv::Size(w, h));
+ cv::Mat resp_roi(all_responses, cv::Rect(j * w, i * h, w, h));
+ tmp.copyTo(resp_roi);
+ }
+ }
+ cv::namedWindow("All responses", CV_WINDOW_AUTOSIZE);
+ cv::imshow("All responses", all_responses);
+ }
+
cv::Point2i &max_response_pt = IF_BIG_BATCH(d.threadctxs[0].max[max_idx].loc, d.threadctxs[max_idx].max.loc);
cv::Mat max_response_map = IF_BIG_BATCH(d.threadctxs[0].response.plane(max_idx), d.threadctxs[max_idx].response.plane(0));
uint max_idx;
max_response = findMaxReponse(max_idx, new_location);
+ new_location.x *= double(p_windows_size.width) / fit_size.width;
+ new_location.y *= double(p_windows_size.height) / fit_size.height;
+
p_current_center += p_current_scale * p_cell_size * new_location;
- if (p_fit_to_pw2) {
- clamp2(p_current_center.x, 0.0, (img.cols * p_fit_factor_x) - 1);
- clamp2(p_current_center.y, 0.0, (img.rows * p_fit_factor_y) - 1);
- } else {
- clamp2(p_current_center.x, 0.0, img.cols - 1.0);
- clamp2(p_current_center.y, 0.0, img.rows - 1.0);
- }
+ clamp2(p_current_center.x, 0.0, img.cols - 1.0);
+ clamp2(p_current_center.y, 0.0, img.rows - 1.0);
// sub grid scale interpolation
if (m_use_subgrid_scale) {
DEBUG_PRINTM(zf);
if (kcf.m_use_linearkernel) {
- kzf = zf.mul(kcf.p_model_alphaf).sum_over_channels();
+ kzf = zf.mul(kcf.model->model_alphaf).sum_over_channels();
} else {
- gaussian_correlation(kzf, zf, kcf.p_model_xf, kcf.p_kernel_sigma, false, kcf);
+ gaussian_correlation(kzf, zf, kcf.model->model_xf, kcf.p_kernel_sigma, false, kcf);
DEBUG_PRINTM(kzf);
- kzf = kzf.mul(kcf.p_model_alphaf);
+ kzf = kzf.mul(kcf.model->model_alphaf);
}
kcf.fft.inverse(kzf, response);
cv::Mat KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy,
int size_x, int size_y, double scale) const
{
- int size_x_scaled = floor(size_x * scale);
- int size_y_scaled = floor(size_y * scale);
+ cv::Size scaled = cv::Size(floor(size_x * scale), floor(size_y * scale));
- cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
- cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
+ cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, scaled.width, scaled.height);
+ cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, scaled.width, scaled.height);
// resize to default size
- if (scale > 1.) {
+ if (scaled.area() > fit_size.area()) {
// if we downsample use INTER_AREA interpolation
- cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
+ // note: this is just a guess - we may downsample in X and upsample in Y (or vice versa)
+ cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_AREA);
} else {
- cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
+ cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_LINEAR);
}
// get hog(Histogram of Oriented Gradients) features
std::vector<cv::Mat> color_feat;
if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
// resize to default size
- if (scale > 1.) {
+ if (scaled.area() > (fit_size / p_cell_size).area()) {
// if we downsample use INTER_AREA interpolation
- cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_AREA);
+ cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_AREA);
} else {
- cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_LINEAR);
+ cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_LINEAR);
}
}
hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
- int size[] = {p_num_of_feats, p_roi.height, p_roi.width};
+ int size[] = {p_num_of_feats, feature_size.height, feature_size.width};
cv::Mat result(3, size, CV_32F);
for (uint i = 0; i < hog_feat.size(); ++i)
hog_feat[i].copyTo(cv::Mat(size[1], size[2], CV_32FC1, result.ptr(i)));
return rot_labels;
}
-cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
+cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot) const
{
cv::Mat rot_patch(patch.size(), CV_32FC1);
cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
{
TRACE("");
xf.sqr_norm(xf_sqr_norm);
+ DEBUG_PRINTM(xf_sqr_norm[0]);
if (auto_correlation) {
yf_sqr_norm = xf_sqr_norm;
} else {
yf.sqr_norm(yf_sqr_norm);
}
+ DEBUG_PRINTM(yf_sqr_norm[0]);
xyf = auto_correlation ? xf.sqr_mag() : xf * yf.conj(); // xf.muln(yf.conj());
DEBUG_PRINTM(xyf);
DEBUG_PRINTM(xyf_sum);
kcf.fft.inverse(xyf_sum, ifft_res);
DEBUG_PRINTM(ifft_res);
-#ifdef CUFFT
+#if 0 && defined(CUFFT)
// FIXME
cuda_gaussian_correlation(ifft_res.deviceMem(), k.deviceMem(), xf_sqr_norm.deviceMem(),
auto_correlation ? xf_sqr_norm.deviceMem() : yf_sqr_norm.deviceMem(), sigma,
- xf.n_channels, xf.n_scales, kcf.p_roi.height, kcf.p_roi.width);
+ xf.n_channels, xf.n_scales, kcf.feature_size.height, kcf.feature_size.width);
#else
float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));