#ifdef FFTW
#include "fft_fftw.h"
#define FFT Fftw
+#elif defined(CUFFT)
+ #include "fft_cufft.h"
+ #define FFT cuFFT
#else
#include "fft_opencv.h"
#define FFT FftOpencv
KCF_Tracker::~KCF_Tracker()
{
delete &fft;
+#ifdef CUFFT
+ for (int i = 0;i < p_num_scales;++i) {
+ CudaSafeCall(cudaFreeHost(scale_vars[i].xf_sqr_norm));
+ CudaSafeCall(cudaFreeHost(scale_vars[i].yf_sqr_norm));
+ CudaSafeCall(cudaFreeHost(scale_vars[i].data_i_1ch));
+ CudaSafeCall(cudaFreeHost(scale_vars[i].data_i_features));
+ CudaSafeCall(cudaFree(scale_vars[i].gauss_corr_res));
+ CudaSafeCall(cudaFreeHost(scale_vars[i].rot_labels_data));
+ CudaSafeCall(cudaFreeHost(scale_vars[i].data_features));
+ }
+#else
+ for (int i = 0;i < p_num_scales;++i) {
+ free(scale_vars[i].xf_sqr_norm);
+ free(scale_vars[i].yf_sqr_norm);
+ }
+#endif
}
-void KCF_Tracker::init(cv::Mat &img, const cv::Rect & bbox)
+void KCF_Tracker::init(cv::Mat &img, const cv::Rect & bbox, int fit_size_x, int fit_size_y)
{
//check boundary, enforce min size
double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
p_pose.w = x2-x1;
p_pose.h = y2-y1;
p_pose.cx = x1 + p_pose.w/2.;
- p_pose.cy = y1 + p_pose.h/2.;
+ p_pose.cy = y1 + p_pose.h /2.;
cv::Mat input_gray, input_rgb = img.clone();
img.convertTo(input_gray, CV_32FC1);
// don't need too large image
- if (p_pose.w * p_pose.h > 100.*100.) {
+ if (p_pose.w * p_pose.h > 100.*100. && (fit_size_x == -1 || fit_size_y == -1)) {
std::cout << "resizing image by factor of " << 1/p_downscale_factor << std::endl;
p_resize_image = true;
p_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 << "Fit size does not fit to hog cell size. The dimensions have to be divisible by HOG cell size, which is: " << p_cell_size << std::endl;;
+ std::exit(EXIT_FAILURE);
+ }
+ double tmp;
+ if (( tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_x)
+ p_scale_factor_x = fit_size_x/tmp;
+ if (( tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_y)
+ p_scale_factor_y = fit_size_y/tmp;
+ std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x
+ << " and verticaly by factor of " << p_scale_factor_y << std::endl;
+ p_fit_to_pw2 = true;
+ p_pose.scale_x(p_scale_factor_x);
+ p_pose.scale_y(p_scale_factor_y);
+ if (p_scale_factor_x != 1 && p_scale_factor_y != 1) {
+ if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
+ cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
+ cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
+ } else {
+ cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
+ cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
+ }
+ }
}
//compute win size + fit to fhog cell size
else
p_scales.push_back(1.);
+ for (int i = 0;i<p_num_scales;++i) {
+ scale_vars.push_back(Scale_vars());
+ }
+
+ p_num_of_feats = 31;
+ if(m_use_color) p_num_of_feats += 3;
+ if(m_use_cnfeat) p_num_of_feats += 10;
+ p_roi_width = p_windows_size[0]/p_cell_size;
+ p_roi_height = p_windows_size[1]/p_cell_size;
+
+ init_scale_vars();
+
p_current_scale = 1.;
double min_size_ratio = std::max(5.*p_cell_size/p_windows_size[0], 5.*p_cell_size/p_windows_size[1]);
p_output_sigma = std::sqrt(p_pose.w*p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
- //window weights, i.e. labels
- num_of_feats = 31;
- if(m_use_color) num_of_feats += 3;
- if(m_use_cnfeat) num_of_feats += 10;
- fft.init(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size, num_of_feats, p_scales.size());
- p_yf = fft.forward(gaussian_shaped_labels(p_output_sigma, p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
+ fft.init(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size, p_num_of_feats, p_num_scales, m_use_big_batch);
fft.set_window(cosine_window_function(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
+ scale_vars[0].flag = Tracker_flags::TRACKER_INIT;
+ //window weights, i.e. labels
+ gaussian_shaped_labels(scale_vars[0], p_output_sigma, p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size);
+ DEBUG_PRINTM(scale_vars[0].rot_labels);
+
+ fft.forward(scale_vars[0]);
+ DEBUG_PRINTM(p_yf);
+
//obtain a sub-window for training initial model
- std::vector<cv::Mat> path_feat = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1]);
- p_model_xf = fft.forward_window(path_feat);
+ get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], scale_vars[0]);
+ fft.forward_window(scale_vars[0]);
DEBUG_PRINTM(p_model_xf);
+ scale_vars[0].flag = Tracker_flags::AUTO_CORRELATION;
+
if (m_use_linearkernel) {
ComplexMat xfconj = p_model_xf.conj();
p_model_alphaf_den = (p_model_xf * xfconj);
} else {
//Kernel Ridge Regression, calculate alphas (in Fourier domain)
- ComplexMat kf = gaussian_correlation(p_model_xf, p_model_xf, p_kernel_sigma, true);
- DEBUG_PRINTM(kf);
- p_model_alphaf_num = p_yf * kf;
+ gaussian_correlation(scale_vars[0], p_model_xf, p_model_xf, p_kernel_sigma, true);
+ DEBUG_PRINTM(scale_vars[0].kf);
+ p_model_alphaf_num = p_yf * scale_vars[0].kf;
DEBUG_PRINTM(p_model_alphaf_num);
- p_model_alphaf_den = kf * (kf + p_lambda);
+ p_model_alphaf_den = scale_vars[0].kf * (scale_vars[0].kf + p_lambda);
DEBUG_PRINTM(p_model_alphaf_den);
}
p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
// p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
}
-void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat & img)
+void KCF_Tracker::init_scale_vars()
+{
+ double alloc_size;
+
+#ifdef CUFFT
+ if (p_windows_size[1]/p_cell_size*(p_windows_size[0]/p_cell_size/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[0] << "x" << p_windows_size[1] <<
+ " which is " << p_windows_size[0]*p_windows_size[1] << " pixels. " << std::endl;
+ std::exit(EXIT_FAILURE);
+ }
+
+ if (m_use_linearkernel){
+ std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
+ std::exit(EXIT_FAILURE);
+ }
+ cudaSetDeviceFlags(cudaDeviceMapHost);
+
+ for (int i = 0;i<p_num_scales;++i) {
+ alloc_size = p_windows_size[0]/p_cell_size*p_windows_size[1]/p_cell_size*sizeof(cufftReal);
+ CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].data_i_1ch, alloc_size, cudaHostAllocMapped));
+ CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].data_i_1ch_d, (void*)scale_vars[i].data_i_1ch, 0));
+
+ alloc_size = p_windows_size[0]/p_cell_size*p_windows_size[1]/p_cell_size*p_num_of_feats*sizeof(cufftReal);
+ CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].data_i_features, alloc_size, cudaHostAllocMapped));
+ CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].data_i_features_d, (void*)scale_vars[i].data_i_features, 0));
+
+
+ scale_vars[i].ifft2_res = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC(p_num_of_feats), scale_vars[i].data_i_features);
+ scale_vars[i].response = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC1, scale_vars[i].data_i_1ch);
+
+ scale_vars[i].zf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
+ scale_vars[i].kzf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
+ scale_vars[i].kf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
+
+#ifdef BIG_BATCH
+ alloc_size = p_num_of_feats;
+#else
+ alloc_size = 1;
+#endif
+
+ CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].xf_sqr_norm, alloc_size*sizeof(float), cudaHostAllocMapped));
+ CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].xf_sqr_norm_d, (void*)scale_vars[i].xf_sqr_norm, 0));
+
+ CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].yf_sqr_norm, sizeof(float), cudaHostAllocMapped));
+ CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].yf_sqr_norm_d, (void*)scale_vars[i].yf_sqr_norm, 0));
+
+ alloc_size =(p_windows_size[0]/p_cell_size)*(p_windows_size[1]/p_cell_size)*alloc_size*sizeof(float);
+ CudaSafeCall(cudaMalloc((void**)&scale_vars[i].gauss_corr_res, alloc_size));
+ scale_vars[i].in_all = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC1, scale_vars[i].gauss_corr_res);
+
+ alloc_size = (p_windows_size[0]/p_cell_size)*(p_windows_size[1]/p_cell_size)*alloc_size*sizeof(float);
+ CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].rot_labels_data, alloc_size, cudaHostAllocMapped));
+ CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].rot_labels_data_d, (void*)scale_vars[i].rot_labels_data, 0));
+ scale_vars[i].rot_labels = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC1, scale_vars[i].rot_labels_data);
+
+ alloc_size = (p_windows_size[0]/p_cell_size)*((p_windows_size[1]/p_cell_size)*p_num_of_feats)*sizeof(cufftReal);
+ CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].data_features, alloc_size, cudaHostAllocMapped));
+ CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].data_features_d, (void*)scale_vars[i].data_features, 0));
+ scale_vars[i].fw_all = cv::Mat((p_windows_size[1]/p_cell_size)*p_num_of_feats, p_windows_size[0]/p_cell_size, CV_32F, scale_vars[i].data_features);
+ }
+#else
+if(m_use_big_batch)
+ alloc_size = p_num_of_feats;
+else
+ alloc_size = 1;
+
+ for (int i = 0;i<p_num_scales;++i) {
+ scale_vars[i].xf_sqr_norm = (float*) malloc(alloc_size*sizeof(float));
+ scale_vars[i].yf_sqr_norm = (float*) malloc(sizeof(float));
+
+ scale_vars[i].patch_feats.reserve(p_num_of_feats);
+#ifdef FFTW
+ scale_vars[i].ifft2_res = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC(p_num_of_feats));
+ scale_vars[i].response = cv::Mat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, CV_32FC1);
+
+ scale_vars[i].zf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
+ scale_vars[i].kzf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
+ scale_vars[i].kf = ComplexMat(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
+
+ scale_vars[i].in_all = cv::Mat((p_windows_size[1]/p_cell_size)*p_num_of_feats, p_windows_size[0]/p_cell_size, CV_32F);
+ scale_vars[i].fw_all = cv::Mat((p_windows_size[1]/p_cell_size)*p_num_of_feats, p_windows_size[0]/p_cell_size, CV_32F);
+#else
+ scale_vars[i].zf = ComplexMat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, p_num_of_feats);
+ //We use scale_vars[0] for updating the tracker, so we only allocate memory for its xf only.
+ if (i==0)
+ scale_vars[i].xf = ComplexMat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, p_num_of_feats);
+#endif
+ }
+#endif
+#if defined(FFTW) || defined(CUFFT)
+ p_model_xf.create(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
+ p_yf.create(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, 1);
+ //We use scale_vars[0] for updating the tracker, so we only allocate memory for its xf only.
+ scale_vars[0].xf.create(p_windows_size[1]/p_cell_size, (p_windows_size[0]/p_cell_size)/2+1, p_num_of_feats);
+#else
+ p_model_xf.create(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, p_num_of_feats);
+ p_yf.create(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, 1);
+#endif
+ scale_vars[0].p_model_xf_ptr = & p_model_xf;
+ scale_vars[0].p_yf_ptr = & p_yf;
+}
+
+void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat & img, int fit_size_x, int fit_size_y)
{
- init(img, bbox.get_rect());
+ init(img, bbox.get_rect(), fit_size_x, fit_size_y);
}
void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
tmp.scale(p_downscale_factor);
p_pose.cx = tmp.cx;
p_pose.cy = tmp.cy;
+ } else if (p_fit_to_pw2) {
+ BBox_c tmp = bbox;
+ tmp.scale_x(p_scale_factor_x);
+ tmp.scale_y(p_scale_factor_y);
+ p_pose.cx = tmp.cx;
+ p_pose.cy = tmp.cy;
} else {
p_pose.cx = bbox.cx;
p_pose.cy = bbox.cy;
if (p_resize_image)
tmp.scale(1/p_downscale_factor);
+ if (p_fit_to_pw2) {
+ tmp.scale_x(1/p_scale_factor_x);
+ tmp.scale_y(1/p_scale_factor_y);
+ }
return tmp;
}
void KCF_Tracker::track(cv::Mat &img)
{
if (m_debug)
- std::cout << "NEW FRAME" << '\n';
+ std::cout << "NEW FRAME" << '\n';
cv::Mat input_gray, input_rgb = img.clone();
if (img.channels() == 3){
cv::cvtColor(img, input_gray, CV_BGR2GRAY);
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 && p_scale_factor_x != 1 && p_scale_factor_y != 1) {
+ if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
+ cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
+ cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
+ } else {
+ cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
+ cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
+ }
}
-
- std::vector<cv::Mat> patch_feat;
double max_response = -1.;
- cv::Mat max_response_map;
- cv::Point2i max_response_pt;
int scale_index = 0;
- std::vector<double> scale_responses;
+ cv::Point2i *max_response_pt = nullptr;
+ cv::Mat *max_response_map = nullptr;
- if (m_use_multithreading){
- std::vector<std::future<cv::Mat>> async_res(p_scales.size());
- for (size_t i = 0; i < p_scales.size(); ++i) {
+ if(m_use_multithreading) {
+ std::vector<std::future<void>> async_res(p_scales.size());
+ for (size_t i = 0; i < scale_vars.size(); ++i) {
async_res[i] = std::async(std::launch::async,
- [this, &input_gray, &input_rgb, i]() -> cv::Mat
- {
- std::vector<cv::Mat> patch_feat_async = get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size[0],
- this->p_windows_size[1], this->p_current_scale * this->p_scales[i]);
- ComplexMat zf = fft.forward_window(patch_feat_async);
- if (m_use_linearkernel)
- return fft.inverse((p_model_alphaf * zf).sum_over_channels());
- else {
- ComplexMat kzf = gaussian_correlation(zf, this->p_model_xf, this->p_kernel_sigma);
- return fft.inverse(this->p_model_alphaf * kzf);
- }
- });
+ [this, &input_gray, &input_rgb, i]() -> void
+ {return scale_track(this->scale_vars[i], input_rgb, input_gray, this->p_scales[i]);});
}
-
for (size_t i = 0; i < p_scales.size(); ++i) {
- // wait for result
async_res[i].wait();
- cv::Mat response = async_res[i].get();
-
- double min_val, max_val;
- cv::Point2i min_loc, max_loc;
- cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
-
- double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
- if (max_val*weight > max_response) {
- max_response = max_val*weight;
- max_response_map = response;
- max_response_pt = max_loc;
+ if (this->scale_vars[i].max_response > max_response) {
+ max_response = this->scale_vars[i].max_response;
+ max_response_pt = & this->scale_vars[i].max_loc;
+ max_response_map = & this->scale_vars[i].response;
scale_index = i;
}
- scale_responses.push_back(max_val*weight);
- }
- } else if(m_use_big_batch){
- for (size_t i = 0; i < p_scales.size(); ++i) {
- std::vector<cv::Mat> tmp = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], p_current_scale * p_scales[i]);
- patch_feat.insert(std::end(patch_feat), std::begin(tmp), std::end(tmp));
}
- ComplexMat zf = fft.forward_window(patch_feat);
- DEBUG_PRINTM(zf);
- cv::Mat response;
-
- if (m_use_linearkernel)
- response = fft.inverse((zf.mul2(p_model_alphaf)).sum_over_channels());
- else {
- ComplexMat kzf = gaussian_correlation(zf, p_model_xf, p_kernel_sigma);
- DEBUG_PRINTM(p_model_alphaf);
- DEBUG_PRINTM(kzf);
- response = fft.inverse(kzf.mul(p_model_alphaf));
- }
- DEBUG_PRINTM(response);
- std::vector<cv::Mat> scales;
- cv::split(response,scales);
-
- /* target location is at the maximum response. we must take into
- account the fact that, if the target doesn't move, the peak
- will appear at the top-left corner, not at the center (this is
- discussed in the paper). the responses wrap around cyclically. */
- for (size_t i = 0; i < p_scales.size(); ++i) {
- double min_val, max_val;
- cv::Point2i min_loc, max_loc;
- cv::minMaxLoc(scales[i], &min_val, &max_val, &min_loc, &max_loc);
- DEBUG_PRINT(max_loc);
-
- double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
- if (max_val*weight > max_response) {
- max_response = max_val*weight;
- max_response_map = scales[i];
- max_response_pt = max_loc;
- scale_index = i;
- }
- scale_responses.push_back(max_val*weight);
- }
} else {
-#pragma omp parallel for ordered private(patch_feat) schedule(dynamic)
- for (size_t i = 0; i < p_scales.size(); ++i) {
- patch_feat = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], p_current_scale * p_scales[i]);
- ComplexMat zf = fft.forward_window(patch_feat);
- DEBUG_PRINTM(zf);
- cv::Mat response;
- if (m_use_linearkernel)
- response = fft.inverse((p_model_alphaf * zf).sum_over_channels());
- else {
- ComplexMat kzf = gaussian_correlation(zf, p_model_xf, p_kernel_sigma);
- DEBUG_PRINTM(p_model_alphaf);
- DEBUG_PRINTM(kzf);
- DEBUG_PRINTM(p_model_alphaf * kzf);
- response = fft.inverse(p_model_alphaf * kzf);
- }
- DEBUG_PRINTM(response);
-
- /* target location is at the maximum response. we must take into
- account the fact that, if the target doesn't move, the peak
- will appear at the top-left corner, not at the center (this is
- discussed in the paper). the responses wrap around cyclically. */
- double min_val, max_val;
- cv::Point2i min_loc, max_loc;
- cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
- DEBUG_PRINT(max_loc);
-
- double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
+#pragma omp parallel for schedule(dynamic)
+ for (size_t i = 0; i < scale_vars.size(); ++i) {
+ scale_track(this->scale_vars[i], input_rgb, input_gray, this->p_scales[i]);
#pragma omp critical
{
- if (max_val*weight > max_response) {
- max_response = max_val*weight;
- max_response_map = response;
- max_response_pt = max_loc;
+ if (this->scale_vars[i].max_response > max_response) {
+ max_response = this->scale_vars[i].max_response;
+ max_response_pt = & this->scale_vars[i].max_loc;
+ max_response_map = & this->scale_vars[i].response;
scale_index = i;
}
}
-#pragma omp ordered
- scale_responses.push_back(max_val*weight);
}
}
- DEBUG_PRINTM(max_response_map);
- DEBUG_PRINT(max_response_pt);
+
+ DEBUG_PRINTM(*max_response_map);
+ DEBUG_PRINT(*max_response_pt);
+
//sub pixel quadratic interpolation from neighbours
- if (max_response_pt.y > max_response_map.rows / 2) //wrap around to negative half-space of vertical axis
- max_response_pt.y = max_response_pt.y - max_response_map.rows;
- if (max_response_pt.x > max_response_map.cols / 2) //same for horizontal axis
- max_response_pt.x = max_response_pt.x - max_response_map.cols;
+ if (max_response_pt->y > max_response_map->rows / 2) //wrap around to negative half-space of vertical axis
+ max_response_pt->y = max_response_pt->y - max_response_map->rows;
+ if (max_response_pt->x > max_response_map->cols / 2) //same for horizontal axis
+ max_response_pt->x = max_response_pt->x - max_response_map->cols;
- cv::Point2f new_location(max_response_pt.x, max_response_pt.y);
+ cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
DEBUG_PRINT(new_location);
if (m_use_subpixel_localization)
- new_location = sub_pixel_peak(max_response_pt, max_response_map);
+ new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
DEBUG_PRINT(new_location);
p_pose.cx += p_current_scale*p_cell_size*new_location.x;
p_pose.cy += p_current_scale*p_cell_size*new_location.y;
- if (p_pose.cx < 0) p_pose.cx = 0;
- if (p_pose.cx > img.cols-1) p_pose.cx = img.cols-1;
- if (p_pose.cy < 0) p_pose.cy = 0;
- if (p_pose.cy > img.rows-1) p_pose.cy = img.rows-1;
+ if (p_fit_to_pw2) {
+ if (p_pose.cx < 0) p_pose.cx = 0;
+ if (p_pose.cx > (img.cols*p_scale_factor_x)-1) p_pose.cx = (img.cols*p_scale_factor_x)-1;
+ if (p_pose.cy < 0) p_pose.cy = 0;
+ if (p_pose.cy > (img.rows*p_scale_factor_y)-1) p_pose.cy = (img.rows*p_scale_factor_y)-1;
+ } else {
+ if (p_pose.cx < 0) p_pose.cx = 0;
+ if (p_pose.cx > img.cols-1) p_pose.cx = img.cols-1;
+ if (p_pose.cy < 0) p_pose.cy = 0;
+ if (p_pose.cy > img.rows-1) p_pose.cy = img.rows-1;
+ }
//sub grid scale interpolation
double new_scale = p_scales[scale_index];
if (m_use_subgrid_scale)
- new_scale = sub_grid_scale(scale_responses, scale_index);
+ new_scale = sub_grid_scale(scale_index);
p_current_scale *= new_scale;
if (p_current_scale > p_min_max_scale[1])
p_current_scale = p_min_max_scale[1];
//obtain a subwindow for training at newly estimated target position
- patch_feat = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], p_current_scale);
- ComplexMat xf = fft.forward_window(patch_feat);
+ get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], scale_vars[0], p_current_scale);
+ scale_vars[0].flag = Tracker_flags::TRACKER_UPDATE;
+ fft.forward_window(scale_vars[0]);
//subsequent frames, interpolate model
- p_model_xf = p_model_xf * (1. - p_interp_factor) + xf * p_interp_factor;
+ p_model_xf = p_model_xf * (1. - p_interp_factor) + scale_vars[0].xf * p_interp_factor;
ComplexMat alphaf_num, alphaf_den;
if (m_use_linearkernel) {
- ComplexMat xfconj = xf.conj();
+ ComplexMat xfconj = scale_vars[0].xf.conj();
alphaf_num = xfconj.mul(p_yf);
- alphaf_den = (xf * xfconj);
+ alphaf_den = (scale_vars[0].xf * xfconj);
} else {
+ scale_vars[0].flag = Tracker_flags::AUTO_CORRELATION;
//Kernel Ridge Regression, calculate alphas (in Fourier domain)
- ComplexMat kf = gaussian_correlation(xf, xf, p_kernel_sigma, true);
+ gaussian_correlation(scale_vars[0], scale_vars[0].xf, scale_vars[0].xf, p_kernel_sigma, true);
// ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
// p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
- alphaf_num = p_yf * kf;
- alphaf_den = kf * (kf + p_lambda);
+ alphaf_num = p_yf * scale_vars[0].kf;
+ alphaf_den = scale_vars[0].kf * (scale_vars[0].kf + p_lambda);
}
p_model_alphaf_num = p_model_alphaf_num * (1. - p_interp_factor) + alphaf_num * p_interp_factor;
p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
}
+void KCF_Tracker::scale_track(Scale_vars & vars, cv::Mat & input_rgb, cv::Mat & input_gray, double scale)
+{
+ get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size[0], this->p_windows_size[1],
+ vars, this->p_current_scale * scale);
+
+ vars.flag = Tracker_flags::SCALE_RESPONSE;
+ fft.forward_window(vars);
+ DEBUG_PRINTM(vars.zf);
+
+ if (m_use_linearkernel) {
+ vars.kzf = (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels();
+ vars.flag = Tracker_flags::RESPONSE;
+ fft.inverse(vars);
+ } else {
+ vars.flag = Tracker_flags::CROSS_CORRELATION;
+ gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
+ DEBUG_PRINTM(this->p_model_alphaf);
+ DEBUG_PRINTM(vars.kzf);
+ DEBUG_PRINTM(this->p_model_alphaf * vars.kzf);
+ vars.flag = Tracker_flags::RESPONSE;
+ vars.kzf = this->p_model_alphaf * vars.kzf;
+ //TODO Add support for fft.inverse(vars) for CUFFT
+ fft.inverse(vars);
+ }
+
+ DEBUG_PRINTM(vars.response);
+
+ /* target location is at the maximum response. we must take into
+ account the fact that, if the target doesn't move, the peak
+ will appear at the top-left corner, not at the center (this is
+ discussed in the paper). the responses wrap around cyclically. */
+ double min_val;
+ cv::Point2i min_loc;
+ cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
+
+ DEBUG_PRINT(vars.max_loc);
+
+ double weight = scale < 1. ? scale : 1./scale;
+ vars.max_response = vars.max_val*weight;
+}
+
// ****************************************************************************
-std::vector<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)
+void KCF_Tracker::get_features(cv::Mat & input_rgb, cv::Mat & input_gray, int cx, int cy, int size_x, int size_y, Scale_vars &vars, double scale)
{
int size_x_scaled = floor(size_x*scale);
int size_y_scaled = floor(size_y*scale);
}
// get hog(Histogram of Oriented Gradients) features
- std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
+ FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
//get color rgb features (simple r,g,b channels)
std::vector<cv::Mat> color_feat;
color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
}
- hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
- return hog_feat;
+ vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
+ return;
}
-cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
+void KCF_Tracker::gaussian_shaped_labels(Scale_vars & vars, double sigma, int dim1, int dim2)
{
cv::Mat labels(dim2, dim1, CV_32FC1);
int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
}
//rotate so that 1 is at top-left corner (see KCF paper for explanation)
- cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
+ cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
+ tmp.copyTo(vars.rot_labels);
//sanity check, 1 at top left corner
- assert(rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
+ assert(vars.rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
- return rot_labels;
+ return;
}
cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
// Returns sub-window of image input centered at [cx, cy] coordinates),
// with size [width, height]. If any pixels are outside of the image,
// they will replicate the values at the borders.
-cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
+cv::Mat KCF_Tracker::get_subwindow(const cv::Mat & input, int cx, int cy, int width, int height)
{
cv::Mat patch;
return patch;
}
-ComplexMat KCF_Tracker::gaussian_correlation(const ComplexMat &xf, const ComplexMat &yf, double sigma, bool auto_correlation)
+void KCF_Tracker::gaussian_correlation(struct Scale_vars & vars, const ComplexMat & xf, const ComplexMat & yf, double sigma, bool auto_correlation)
{
- std::vector<float> xf_sqr_norm = xf.sqr_norm();
- std::vector<float> yf_sqr_norm = auto_correlation ? xf_sqr_norm : yf.sqr_norm();
-
- ComplexMat xyf;
- xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
- DEBUG_PRINTM(xyf);
-
+#ifdef CUFFT
+ xf.sqr_norm(vars.xf_sqr_norm_d);
+ if (!auto_correlation)
+ yf.sqr_norm(vars.yf_sqr_norm_d);
+#else
+ xf.sqr_norm(vars.xf_sqr_norm);
+ if (auto_correlation){
+ vars.yf_sqr_norm[0] = vars.xf_sqr_norm[0];
+ } else {
+ yf.sqr_norm(vars.yf_sqr_norm);
+ }
+#endif
+ vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
+ DEBUG_PRINTM(vars.xyf);
+#ifdef CUFFT
+ fft.inverse(vars);
+ if(auto_correlation)
+ cuda_gaussian_correlation(vars.data_i_features, vars.gauss_corr_res, vars.xf_sqr_norm_d, vars.xf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
+ else
+ cuda_gaussian_correlation(vars.data_i_features, vars.gauss_corr_res, vars.xf_sqr_norm_d, vars.yf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
+#else
//ifft2 and sum over 3rd dimension, we dont care about individual channels
- cv::Mat ifft2_res = fft.inverse(xyf);
+ fft.inverse(vars);
+ DEBUG_PRINTM(vars.ifft2_res);
cv::Mat xy_sum;
- if(xf.channels() != 308)
- xy_sum.create(ifft2_res.size(), CV_32FC1);
+ if (xf.channels() != p_num_scales*p_num_of_feats)
+ xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
else
- xy_sum.create(ifft2_res.size(), CV_32FC(p_scales.size()));
+ xy_sum.create(vars.ifft2_res.size(), CV_32FC(p_scales.size()));
xy_sum.setTo(0);
- for (int y = 0; y < ifft2_res.rows; ++y) {
- float * row_ptr = ifft2_res.ptr<float>(y);
+ for (int y = 0; y < vars.ifft2_res.rows; ++y) {
+ float * row_ptr = vars.ifft2_res.ptr<float>(y);
float * row_ptr_sum = xy_sum.ptr<float>(y);
- for (int x = 0; x < ifft2_res.cols; ++x) {
+ for (int x = 0; x < vars.ifft2_res.cols; ++x) {
for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
- row_ptr_sum[(x*xy_sum.channels())+sum_ch] += std::accumulate(row_ptr + x*ifft2_res.channels() + sum_ch*(ifft2_res.channels()/xy_sum.channels()),
- (row_ptr + x*ifft2_res.channels() + (sum_ch+1)*(ifft2_res.channels()/xy_sum.channels())), 0.f);
- }
+ row_ptr_sum[(x*xy_sum.channels())+sum_ch] += std::accumulate(row_ptr + x*vars.ifft2_res.channels() + sum_ch*(vars.ifft2_res.channels()/xy_sum.channels()),
+ (row_ptr + x*vars.ifft2_res.channels() + (sum_ch+1)*(vars.ifft2_res.channels()/xy_sum.channels())), 0.f);
+ }
}
}
- DEBUG_PRINTM(ifft2_res);
DEBUG_PRINTM(xy_sum);
std::vector<cv::Mat> scales;
cv::split(xy_sum,scales);
- cv::Mat in_all(scales[0].rows * xf.n_scales, scales[0].cols, CV_32F);
float numel_xf_inv = 1.f/(xf.cols * xf.rows * (xf.channels()/xf.n_scales));
- for(int i = 0; i < xf.n_scales; ++i){
- cv::Mat in_roi(in_all, cv::Rect(0, i*scales[0].rows, scales[0].cols, scales[0].rows));
- cv::exp(- 1.f / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0), in_roi);
- DEBUG_PRINTM(in_roi);
+ for (int i = 0; i < xf.n_scales; ++i){
+ cv::Mat in_roi(vars.in_all, cv::Rect(0, i*scales[0].rows, scales[0].cols, scales[0].rows));
+ cv::exp(- 1.f / (sigma * sigma) * cv::max((vars.xf_sqr_norm[i] + vars.yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0), in_roi);
+ DEBUG_PRINTM(in_roi);
}
-
- DEBUG_PRINTM(in_all);
- return fft.forward(in_all);
+#endif
+ DEBUG_PRINTM(vars.in_all);
+ fft.forward(vars);
+ return;
}
float get_response_circular(cv::Point2i & pt, cv::Mat & response)
cv::Point2i p4(max_loc.x-1, max_loc.y), p5(max_loc.x+1, max_loc.y);
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);
+ // clang-format off
// fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
cv::Mat A = (cv::Mat_<float>(9, 6) <<
p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
get_response_circular(p7, response),
get_response_circular(p8, response),
get_response_circular(max_loc, response));
+ // clang-format on
cv::Mat x;
cv::solve(A, fval, x, cv::DECOMP_SVD);
return sub_peak;
}
-double KCF_Tracker::sub_grid_scale(std::vector<double> & responses, int index)
+double KCF_Tracker::sub_grid_scale(int index)
{
cv::Mat A, fval;
if (index < 0 || index > (int)p_scales.size()-1) {
A.at<float>(i, 0) = p_scales[i] * p_scales[i];
A.at<float>(i, 1) = p_scales[i];
A.at<float>(i, 2) = 1;
- fval.at<float>(i) = responses[i];
+ fval.at<float>(i) = scale_vars[i].max_response;
}
} else {
//only from neighbours
p_scales[index-1] * p_scales[index-1], p_scales[index-1], 1,
p_scales[index] * p_scales[index], p_scales[index], 1,
p_scales[index+1] * p_scales[index+1], p_scales[index+1], 1);
- fval = (cv::Mat_<float>(3, 1) << responses[index-1], responses[index], responses[index+1]);
+ fval = (cv::Mat_<float>(3, 1) << scale_vars[index-1].max_response, scale_vars[index].max_response, scale_vars[index+1].max_response);
}
cv::Mat x;
projects / hercules2020 / kcf.git / blobdiff