#include "kcf.h"
#include <numeric>
#include <thread>
-#include <future>
#include <algorithm>
+#include "threadctx.hpp"
+#include "debug.h"
+#include <limits>
#ifdef FFTW
#include "fft_fftw.h"
#include <omp.h>
#endif // OPENMP
-#define DEBUG_PRINT(obj) \
- if (m_debug) { \
- std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl; \
- }
-#define DEBUG_PRINTM(obj) \
- if (m_debug) { \
- std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl \
- << (obj) << std::endl; \
- }
+DbgTracer __dbgTracer;
+
+template <typename T>
+T clamp(const T& n, const T& lower, const T& upper)
+{
+ return std::max(lower, std::min(n, upper));
+}
+
+template <typename T>
+void clamp2(T& n, const T& lower, const T& upper)
+{
+ 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);
+}
+
+template<typename _Tp> static inline
+cv::Point_<_Tp> operator / (const cv::Point_<_Tp>& a, double b)
+{
+ return cv::Point_<_Tp>(a.x / b, a.y / b);
+}
+
+#endif
+
+class Kcf_Tracker_Private {
+ friend KCF_Tracker;
+
+ Kcf_Tracker_Private(const KCF_Tracker &kcf) : kcf(kcf) {}
+
+ const KCF_Tracker &kcf;
+#ifdef BIG_BATCH
+ std::vector<ThreadCtx> threadctxs;
+#else
+ ScaleRotVector<ThreadCtx> threadctxs{kcf.p_scales, kcf.p_angles};
+#endif
+};
KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
double output_sigma_factor, int cell_size)
- : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
- p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size)
+ : p_cell_size(cell_size), fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
+ p_lambda(lambda), p_interp_factor(interp_factor)
{
}
delete &fft;
}
+void KCF_Tracker::train(cv::Mat input_rgb, cv::Mat input_gray, double interp_factor)
+{
+ TRACE("");
+
+ // obtain a sub-window for training
+ get_features(input_rgb, input_gray, nullptr, p_current_center.x, p_current_center.y,
+ p_windows_size.width, p_windows_size.height,
+ p_current_scale, p_current_angle).copyTo(model->patch_feats.scale(0));
+ DEBUG_PRINT(model->patch_feats);
+ fft.forward_window(model->patch_feats, model->xf, model->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 = 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(feature_size));
+ ComplexMat kf(sz.height, sz.width, 1);
+ (*gaussian_correlation)(kf, model->model_xf, model->model_xf, p_kernel_sigma, true, *this);
+ DEBUG_PRINTM(kf);
+ model->model_alphaf_num = model->yf * kf;
+ model->model_alphaf_den = kf * (kf + p_lambda);
+ }
+ 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;
+ TRACE("");
+
// check boundary, enforce min size
double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
if (x1 < 0) x1 = 0.;
}
}
- 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_init_pose.w = x2 - x1;
+ p_init_pose.h = y2 - y1;
+ p_init_pose.cx = x1 + p_init_pose.w / 2.;
+ p_init_pose.cy = y1 + p_init_pose.h / 2.;
cv::Mat input_gray, input_rgb = img.clone();
if (img.channels() == 3) {
img.convertTo(input_gray, CV_32FC1);
// don't need too large image
- if (p_pose.w * p_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_pose.scale(p_downscale_factor);
+ 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 << "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 = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
- if (fabs(tmp - fit_size_x) > p_floating_error) p_scale_factor_x = fit_size_x / tmp;
- tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
- if (fabs(tmp - fit_size_y) > p_floating_error) 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 (fabs(p_scale_factor_x - 1) > p_floating_error && fabs(p_scale_factor_y - 1) > p_floating_error) {
- 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
- p_windows_size.width = int(round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size);
- p_windows_size.height = int(round(p_pose.h * (1. + p_padding) / p_cell_size) * p_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;
+
+ 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);
+ }
- 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.width / p_cell_size;
- p_roi_height = p_windows_size.height / p_cell_size;
+ feature_size = fit_size / p_cell_size;
p_scales.clear();
- if (m_use_scale)
- for (int i = -p_num_scales / 2; i <= p_num_scales / 2; ++i)
- p_scales.push_back(std::pow(p_scale_step, i));
- else
- p_scales.push_back(1.);
+ for (int i = -int(p_num_scales - 1) / 2; i <= int(p_num_scales) / 2; ++i)
+ p_scales.push_back(std::pow(p_scale_step, i));
-#ifdef CUFFT
- if (p_windows_size.height / p_cell_size * (p_windows_size.width / 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.width << "x" << p_windows_size.height
- << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
- std::exit(EXIT_FAILURE);
- }
+ p_angles.clear();
+ for (int i = -int(p_num_angles - 1) / 2; i <= int(p_num_angles) / 2; ++i)
+ p_angles.push_back(i * p_angle_step);
+#ifdef CUFFT
if (m_use_linearkernel) {
std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
std::exit(EXIT_FAILURE);
}
- CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
- p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size) / 2 + 1, p_num_of_feats, this->stream);
- p_rot_labels_data = DynMem(
- ((uint(p_windows_size.width) / p_cell_size) * (uint(p_windows_size.height) / p_cell_size)) * sizeof(float));
- p_rot_labels = cv::Mat(p_windows_size.height / int(p_cell_size), windows_size.width / int(p_cell_size), CV_32FC1,
- p_rot_labels_data.hostMem());
-#else
- p_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.height / p_cell_size)) / 2 + 1,
- p_num_of_feats);
#endif
-#if defined(CUFFT) || defined(FFTW)
- p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1,
- uint(p_num_of_feats));
- p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)) / 2 + 1, 1);
- p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size) / 2 + 1,
- p_num_of_feats);
+ model.reset(new Model(feature_size, p_num_of_feats));
+ d.reset(new Kcf_Tracker_Private(*this));
+
+#ifndef BIG_BATCH
+ for (auto scale: p_scales)
+ for (auto angle : p_angles)
+ d->threadctxs.emplace_back(feature_size, p_num_of_feats, scale, angle);
#else
- p_model_xf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)),
- uint(p_num_of_feats));
- p_yf.create(uint(p_windows_size.height / p_cell_size), (uint(p_windows_size.width / p_cell_size)), 1);
- p_xf.create(uint(p_windows_size.height) / p_cell_size, (uint(p_windows_size.width) / p_cell_size), p_num_of_feats);
+ d->threadctxs.emplace_back(feature_size, p_num_of_feats, p_scales, p_angles);
#endif
- int max = m_use_big_batch ? 2 : p_num_scales;
- for (int i = 0; i < max; ++i) {
- if (m_use_big_batch && i == 1) {
- p_scale_vars.emplace_back(
- new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats * p_num_scales, p_num_scales));
- } else {
- p_scale_vars.emplace_back(new ThreadCtx(p_windows_size, p_cell_size, p_num_of_feats, 1));
- }
- }
+ gaussian_correlation.reset(new GaussianCorrelation(1, p_num_of_feats, feature_size));
+ p_current_center = p_init_pose.center();
p_current_scale = 1.;
double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
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 << " " << img.rows << std::endl;
- std::cout << "init: win size. " << p_windows_size.width << " " << p_windows_size.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_pose.w * p_pose.h) * p_output_sigma_factor / static_cast<double>(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(uint(p_windows_size.width / p_cell_size), uint(p_windows_size.height / p_cell_size), uint(p_num_of_feats),
- uint(p_num_scales), m_use_big_batch);
- fft.set_window(cosine_window_function(p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size));
+ fft.init(feature_size.width, feature_size.height, p_num_of_feats, p_num_scales * p_num_angles);
+ fft.set_window(MatDynMem(cosine_window_function(feature_size.width, feature_size.height)));
// window weights, i.e. labels
- fft.forward(
- gaussian_shaped_labels(p_output_sigma, p_windows_size.width / p_cell_size, p_windows_size.height / p_cell_size), p_yf,
- m_use_cuda ? p_rot_labels_data.deviceMem() : nullptr, p_scale_vars.front()->stream);
- DEBUG_PRINTM(p_yf);
-
- // obtain a sub-window for training initial model
- p_scale_vars.front()->patch_feats.clear();
- get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
- *p_scale_vars.front());
- fft.forward_window(p_scale_vars.front()->patch_feats, p_model_xf, p_scale_vars.front()->fw_all,
- m_use_cuda ? p_scale_vars.front()->data_features.deviceMem() : nullptr, p_scale_vars.front()->stream);
- DEBUG_PRINTM(p_model_xf);
-#if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
- p_scale_vars.front()->model_xf = p_model_xf;
- p_scale_vars.front()->model_xf.set_stream(p_scale_vars.front()->stream);
- p_yf.set_stream(p_scale_vars.front()->stream);
- p_model_xf.set_stream(p_scale_vars.front()->stream);
- p_xf.set_stream(p_scale_vars.front()->stream);
-#endif
-
- if (m_use_linearkernel) {
- ComplexMat xfconj = p_model_xf.conj();
- p_model_alphaf_num = xfconj.mul(p_yf);
- p_model_alphaf_den = (p_model_xf * xfconj);
- } else {
- // Kernel Ridge Regression, calculate alphas (in Fourier domain)
-#if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
- gaussian_correlation(*p_scale_vars.front(), p_scale_vars.front()->model_xf, p_scale_vars.front()->model_xf,
- p_kernel_sigma, true);
-#else
- gaussian_correlation(*p_scale_vars.front(), p_model_xf, p_model_xf, p_kernel_sigma, true);
-#endif
- DEBUG_PRINTM(p_scale_vars.front()->kf);
- p_model_alphaf_num = p_yf * p_scale_vars.front()->kf;
- DEBUG_PRINTM(p_model_alphaf_num);
- p_model_alphaf_den = p_scale_vars.front()->kf * (p_scale_vars.front()->kf + float(p_lambda));
- DEBUG_PRINTM(p_model_alphaf_den);
- }
- p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
- DEBUG_PRINTM(p_model_alphaf);
- // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
+ 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);
-#if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
- for (auto it = p_scale_vars.begin(); it != p_scale_vars.end(); ++it) {
- (*it)->model_xf = p_model_xf;
- (*it)->model_xf.set_stream((*it)->stream);
- (*it)->model_alphaf = p_model_alphaf;
- (*it)->model_alphaf.set_stream((*it)->stream);
- }
-#endif
+ // train initial model
+ train(input_rgb, input_gray, 1.0);
}
void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
{
+ BBox_c tmp = bbox;
if (p_resize_image) {
- BBox_c tmp = 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;
}
+ p_current_center = tmp.center();
}
BBox_c KCF_Tracker::getBBox()
{
- BBox_c tmp = p_pose;
- tmp.w *= p_current_scale;
- tmp.h *= p_current_scale;
-
- 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);
- }
+ BBox_c tmp;
+ tmp.cx = p_current_center.x;
+ tmp.cy = p_current_center.y;
+ tmp.w = p_init_pose.w * p_current_scale;
+ tmp.h = p_init_pose.h * p_current_scale;
+ tmp.a = p_current_angle;
+
+ if (p_resize_image)
+ tmp.scale(1 / p_downscale_factor);
return tmp;
}
-void KCF_Tracker::track(cv::Mat &img)
+double KCF_Tracker::getFilterResponse() const
{
- if (m_debug) std::cout << "NEW FRAME" << '\n';
- cv::Mat input_gray, input_rgb = img.clone();
- if (img.channels() == 3) {
- cv::cvtColor(img, input_gray, CV_BGR2GRAY);
- input_gray.convertTo(input_gray, CV_32FC1);
- } else
- img.convertTo(input_gray, CV_32FC1);
+ return this->max_response;
+}
- // don't need too large image
+void KCF_Tracker::resizeImgs(cv::Mat &input_rgb, cv::Mat &input_gray)
+{
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_scale_factor_x - 1) > p_floating_error &&
- fabs(p_scale_factor_y - 1) > p_floating_error) {
- 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);
- }
}
+}
- double max_response = -1.;
- int scale_index = 0;
- cv::Point2i *max_response_pt = nullptr;
- cv::Mat *max_response_map = nullptr;
-
- if (m_use_multithreading) {
- std::vector<std::future<void>> async_res(p_scales.size());
- for (auto it = p_scale_vars.begin(); it != p_scale_vars.end(); ++it) {
- uint index = uint(std::distance(p_scale_vars.begin(), it));
- async_res[index] = std::async(std::launch::async, [this, &input_gray, &input_rgb, index, it]() -> void {
- return scale_track(*(*it), input_rgb, input_gray, this->p_scales[index]);
- });
- }
- for (auto it = p_scale_vars.begin(); it != p_scale_vars.end(); ++it) {
- uint index = uint(std::distance(p_scale_vars.begin(), it));
- async_res[index].wait();
- if ((*it)->max_response > max_response) {
- max_response = (*it)->max_response;
- max_response_pt = &(*it)->max_loc;
- max_response_map = &(*it)->response;
- scale_index = int(index);
- }
- }
+static void drawCross(cv::Mat &img, cv::Point center, bool green)
+{
+ cv::Scalar col = green ? cv::Scalar(0, 1, 0) : cv::Scalar(0, 0, 1);
+ cv::line(img, cv::Point(center.x, 0), cv::Point(center.x, img.size().height), col);
+ cv::line(img, cv::Point(0, center.y), cv::Point(img.size().height, center.y), col);
+}
+
+static cv::Point2d wrapAroundFreq(cv::Point2d pt, cv::Mat &resp_map)
+{
+ if (pt.y > resp_map.rows / 2) // wrap around to negative half-space of vertical axis
+ pt.y = pt.y - resp_map.rows;
+ if (pt.x > resp_map.cols / 2) // same for horizontal axis
+ pt.x = pt.x - resp_map.cols;
+ return pt;
+}
+
+double KCF_Tracker::findMaxReponse(uint &max_idx, cv::Point2d &new_location) const
+{
+ double max;
+ const auto &vec = IF_BIG_BATCH(d->threadctxs[0].max, d->threadctxs);
+
+#ifndef BIG_BATCH
+ auto max_it = std::max_element(vec.begin(), vec.end(),
+ [](const ThreadCtx &a, const ThreadCtx &b)
+ { return a.max.response < b.max.response; });
+#else
+ auto max_it = std::max_element(vec.begin(), vec.end(),
+ [](const ThreadCtx::Max &a, const ThreadCtx::Max &b)
+ { return a.response < b.response; });
+#endif
+ assert(max_it != vec.end());
+ max = max_it->IF_BIG_BATCH(response, max.response);
+
+ max_idx = std::distance(vec.begin(), max_it);
+
+ cv::Point2i max_response_pt = IF_BIG_BATCH(max_it->loc, max_it->max.loc);
+ cv::Mat max_response_map = IF_BIG_BATCH(d->threadctxs[0].response.plane(max_idx),
+ max_it->response.plane(0));
+
+ DEBUG_PRINTM(max_response_map);
+ DEBUG_PRINT(max_response_pt);
+
+ max_response_pt = wrapAroundFreq(max_response_pt, max_response_map);
+
+ // sub pixel quadratic interpolation from neighbours
+ if (m_use_subpixel_localization) {
+ new_location = sub_pixel_peak(max_response_pt, max_response_map);
} else {
- uint start = m_use_big_batch ? 1 : 0;
- uint end = m_use_big_batch ? 2 : uint(p_num_scales);
- NORMAL_OMP_PARALLEL_FOR
- for (uint i = start; i < end; ++i) {
- auto it = p_scale_vars.begin();
- std::advance(it, i);
- scale_track(*(*it), input_rgb, input_gray, this->p_scales[i]);
-
- if (m_use_big_batch) {
- for (size_t j = 0; j < p_scales.size(); ++j) {
- if ((*it)->max_responses[j] > max_response) {
- max_response = (*it)->max_responses[j];
- max_response_pt = &(*it)->max_locs[j];
- max_response_map = &(*it)->response_maps[j];
- scale_index = int(j);
- }
+ new_location = max_response_pt;
+ }
+ DEBUG_PRINT(new_location);
+
+ if (m_visual_debug != vd::NONE) {
+ const bool fit = 1;
+ int w = fit ? 100 : (m_visual_debug == vd::PATCH ? fit_size.width : feature_size.width);
+ int h = fit ? 100 : (m_visual_debug == vd::PATCH ? fit_size.height : feature_size.height);
+ cv::Mat all_responses((h + 1) * p_num_scales - 1,
+ (w + 1) * p_num_angles - 1, CV_32FC3, cv::Scalar::all(0));
+ for (size_t i = 0; i < p_num_scales; ++i) {
+ for (size_t j = 0; j < p_num_angles; ++j) {
+ auto &threadctx = d->IF_BIG_BATCH(threadctxs[0], threadctxs(i, j));
+ cv::Mat tmp;
+ cv::Point2d cross = threadctx.IF_BIG_BATCH(max(i, j), max).loc;
+ cross = wrapAroundFreq(cross, max_response_map);
+ if (m_visual_debug == vd::PATCH ) {
+ threadctx.dbg_patch IF_BIG_BATCH((i, j),)
+ .convertTo(tmp, all_responses.type(), 1.0 / 255);
+ cross.x = cross.x / fit_size.width * tmp.cols + tmp.cols / 2;
+ cross.y = cross.y / fit_size.height * tmp.rows + tmp.rows / 2;
+ } else {
+ cv::cvtColor(threadctx.response.plane(IF_BIG_BATCH(threadctx.max.getIdx(i, j), 0)),
+ tmp, cv::COLOR_GRAY2BGR);
+ tmp /= max; // Normalize to 1
+ cross += cv::Point2d(tmp.size())/2;
+ tmp = circshift(tmp, -tmp.cols/2, -tmp.rows/2);
+ //drawCross(tmp, cross, false);
}
- } else {
- NORMAL_OMP_CRITICAL
- {
- if ((*it)->max_response > max_response) {
- max_response = (*it)->max_response;
- max_response_pt = &(*it)->max_loc;
- max_response_map = &(*it)->response;
- scale_index = int(i);
- }
+ bool green = false;
+ if (&*max_it == &IF_BIG_BATCH(threadctx.max(i, j), threadctx)) {
+ // Show the green cross at position of sub-pixel interpolation (if enabled)
+ cross = new_location + cv::Point2d(tmp.size())/2;
+ green = true;
}
+ // Move to the center of pixes (if scaling up) and scale
+ cross.x = (cross.x + 0.5) * double(w)/tmp.cols;
+ cross.y = (cross.y + 0.5) * double(h)/tmp.rows;
+ cv::resize(tmp, tmp, cv::Size(w, h)); //, 0, 0, cv::INTER_NEAREST);
+ drawCross(tmp, cross, green);
+ cv::Mat resp_roi(all_responses, cv::Rect(j * (w+1), i * (h+1), w, h));
+ tmp.copyTo(resp_roi);
}
}
+ cv::namedWindow("KCF visual debug", CV_WINDOW_AUTOSIZE);
+ cv::imshow("KCF visual debug", all_responses);
}
- 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;
+ return max;
+}
- cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
- DEBUG_PRINT(new_location);
+void KCF_Tracker::track(cv::Mat &img)
+{
+ __dbgTracer.debug = m_debug;
+ TRACE("");
- if (m_use_subpixel_localization) new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
- DEBUG_PRINT(new_location);
+ cv::Mat input_gray, input_rgb = img.clone();
+ if (img.channels() == 3) {
+ cv::cvtColor(img, input_gray, CV_BGR2GRAY);
+ input_gray.convertTo(input_gray, CV_32FC1);
+ } else
+ img.convertTo(input_gray, CV_32FC1);
- p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
- p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
- 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;
- }
+ // don't need too large image
+ resizeImgs(input_rgb, input_gray);
+
+#ifdef ASYNC
+ for (auto &it : d->threadctxs)
+ it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
+ it.track(*this, input_rgb, input_gray);
+ });
+ for (auto const &it : d->threadctxs)
+ it.async_res.wait();
+
+#else // !ASYNC
+ NORMAL_OMP_PARALLEL_FOR
+ for (uint i = 0; i < d->threadctxs.size(); ++i)
+ d->threadctxs[i].track(*this, input_rgb, input_gray);
+#endif
- // sub grid scale interpolation
- double new_scale = p_scales[uint(scale_index)];
- if (m_use_subgrid_scale) new_scale = sub_grid_scale(scale_index);
+ cv::Point2d new_location;
+ uint max_idx;
+ max_response = findMaxReponse(max_idx, new_location);
- p_current_scale *= new_scale;
+ double angle_change = d->IF_BIG_BATCH(threadctxs[0].max, threadctxs).angle(max_idx);
+ p_current_angle += angle_change;
- if (p_current_scale < p_min_max_scale[0]) p_current_scale = p_min_max_scale[0];
- if (p_current_scale > p_min_max_scale[1]) p_current_scale = p_min_max_scale[1];
+ new_location.x = new_location.x * cos(-p_current_angle/180*M_PI) + new_location.y * sin(-p_current_angle/180*M_PI);
+ new_location.y = new_location.y * cos(-p_current_angle/180*M_PI) - new_location.x * sin(-p_current_angle/180*M_PI);
- // obtain a subwindow for training at newly estimated target position
- p_scale_vars.front()->patch_feats.clear();
- get_features(input_rgb, input_gray, int(p_pose.cx), int(p_pose.cy), p_windows_size.width, p_windows_size.height,
- *p_scale_vars.front(), p_current_scale);
- fft.forward_window(p_scale_vars.front()->patch_feats, p_xf, p_scale_vars.front()->fw_all,
- m_use_cuda ? p_scale_vars.front()->data_features.deviceMem() : nullptr, p_scale_vars.front()->stream);
+ new_location.x *= double(p_windows_size.width) / fit_size.width;
+ new_location.y *= double(p_windows_size.height) / fit_size.height;
- // subsequent frames, interpolate model
- p_model_xf = p_model_xf * float((1. - p_interp_factor)) + p_xf * float(p_interp_factor);
+ p_current_center += p_current_scale * p_cell_size * new_location;
- ComplexMat alphaf_num, alphaf_den;
+ clamp2(p_current_center.x, 0.0, img.cols - 1.0);
+ clamp2(p_current_center.y, 0.0, img.rows - 1.0);
- if (m_use_linearkernel) {
- ComplexMat xfconj = p_xf.conj();
- alphaf_num = xfconj.mul(p_yf);
- alphaf_den = (p_xf * xfconj);
+ // sub grid scale interpolation
+ if (m_use_subgrid_scale) {
+ p_current_scale *= sub_grid_scale(max_idx);
} else {
- // Kernel Ridge Regression, calculate alphas (in Fourier domain)
- gaussian_correlation(*p_scale_vars.front(), p_xf, p_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 * p_scale_vars.front()->kf;
- alphaf_den = p_scale_vars.front()->kf * (p_scale_vars.front()->kf + float(p_lambda));
+ p_current_scale *= d->IF_BIG_BATCH(threadctxs[0].max, threadctxs).scale(max_idx);
}
- p_model_alphaf_num = p_model_alphaf_num * float((1. - p_interp_factor)) + alphaf_num * float(p_interp_factor);
- p_model_alphaf_den = p_model_alphaf_den * float((1. - p_interp_factor)) + alphaf_den * float(p_interp_factor);
- p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
+ clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
-#if defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
- for (auto it = p_scale_vars.begin(); it != p_scale_vars.end(); ++it) {
- (*it)->model_xf = p_model_xf;
- (*it)->model_xf.set_stream((*it)->stream);
- (*it)->model_alphaf = p_model_alphaf;
- (*it)->model_alphaf.set_stream((*it)->stream);
- }
-#endif
+
+ // train at newly estimated target position
+ train(input_rgb, input_gray, p_interp_factor);
}
-void KCF_Tracker::scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray, double scale)
+void ThreadCtx::track(const KCF_Tracker &kcf, cv::Mat &input_rgb, cv::Mat &input_gray)
{
- if (m_use_big_batch) {
- vars.patch_feats.clear();
- BIG_BATCH_OMP_PARALLEL_FOR
- for (uint i = 0; i < uint(p_num_scales); ++i) {
- get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
- this->p_windows_size.height, vars, this->p_current_scale * this->p_scales[i]);
- }
- } else {
- vars.patch_feats.clear();
- get_features(input_rgb, input_gray, int(this->p_pose.cx), int(this->p_pose.cy), this->p_windows_size.width,
- this->p_windows_size.height, vars, this->p_current_scale *scale);
+ TRACE("");
+
+ BIG_BATCH_OMP_PARALLEL_FOR
+ for (uint i = 0; i < IF_BIG_BATCH(max.size(), 1); ++i)
+ {
+ kcf.get_features(input_rgb, input_gray, &dbg_patch IF_BIG_BATCH([i],),
+ kcf.p_current_center.x, kcf.p_current_center.y,
+ kcf.p_windows_size.width, kcf.p_windows_size.height,
+ kcf.p_current_scale * IF_BIG_BATCH(max.scale(i), scale),
+ kcf.p_current_angle + IF_BIG_BATCH(max.angle(i), angle))
+ .copyTo(patch_feats.scale(i));
+ DEBUG_PRINT(patch_feats.scale(i));
}
- fft.forward_window(vars.patch_feats, vars.zf, vars.fw_all, m_use_cuda ? vars.data_features.deviceMem() : nullptr,
- vars.stream);
- DEBUG_PRINTM(vars.zf);
+ kcf.fft.forward_window(patch_feats, zf, temp);
+ DEBUG_PRINTM(zf);
- if (m_use_linearkernel) {
- vars.kzf = m_use_big_batch ? (vars.zf.mul2(this->p_model_alphaf)).sum_over_channels()
- : (p_model_alphaf * vars.zf).sum_over_channels();
- fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
+ if (kcf.m_use_linearkernel) {
+ kzf = zf.mul(kcf.model->model_alphaf).sum_over_channels();
} else {
-#if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
- gaussian_correlation(vars, vars.zf, vars.model_xf, this->p_kernel_sigma);
- vars.kzf = vars.model_alphaf * vars.kzf;
-#else
- gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
- DEBUG_PRINTM(this->p_model_alphaf);
- DEBUG_PRINTM(vars.kzf);
- vars.kzf = m_use_big_batch ? vars.kzf.mul(this->p_model_alphaf) : this->p_model_alphaf * vars.kzf;
-#endif
- fft.inverse(vars.kzf, vars.response, m_use_cuda ? vars.data_i_1ch.deviceMem() : nullptr, vars.stream);
+ gaussian_correlation(kzf, zf, kcf.model->model_xf, kcf.p_kernel_sigma, false, kcf);
+ DEBUG_PRINTM(kzf);
+ kzf = kzf.mul(kcf.model->model_alphaf);
}
+ kcf.fft.inverse(kzf, response);
- DEBUG_PRINTM(vars.response);
+ 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. */
- if (m_use_big_batch) {
- cv::split(vars.response, vars.response_maps);
-
- for (size_t i = 0; i < p_scales.size(); ++i) {
- double min_val, max_val;
- cv::Point2i min_loc, max_loc;
- cv::minMaxLoc(vars.response_maps[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];
- vars.max_responses[i] = max_val * weight;
- vars.max_locs[i] = max_loc;
- }
- } else {
- double min_val;
- cv::Point2i min_loc;
- cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
+ double min_val, max_val;
+ cv::Point2i min_loc, max_loc;
+#ifdef BIG_BATCH
+ for (size_t i = 0; i < max.size(); ++i) {
+ cv::minMaxLoc(response.plane(i), &min_val, &max_val, &min_loc, &max_loc);
+ DEBUG_PRINT(max_loc);
+ double weight = kcf.p_scales[i] < 1. ? kcf.p_scales[i] : 1. / kcf.p_scales[i];
+ max[i].response = max_val * weight;
+ max[i].loc = max_loc;
+ }
+#else
+ cv::minMaxLoc(response.plane(0), &min_val, &max_val, &min_loc, &max_loc);
- DEBUG_PRINT(vars.max_loc);
+ DEBUG_PRINT(max_loc);
+ DEBUG_PRINT(max_val);
- double weight = scale < 1. ? scale : 1. / scale;
- vars.max_response = vars.max_val * weight;
- }
- return;
+ double weight = scale < 1. ? scale : 1. / scale;
+ max.response = max_val * weight;
+ max.loc = max_loc;
+#endif
}
// ****************************************************************************
-void KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y,
- ThreadCtx &vars, double scale)
+cv::Mat KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, cv::Mat *dbg_patch,
+ int cx, int cy, int size_x, int size_y, double scale, double angle) const
{
- int size_x_scaled = int(floor(size_x * scale));
- int size_y_scaled = int(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, angle);
+ cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, scaled.width, scaled.height, angle);
+
+ if (dbg_patch)
+ patch_rgb.copyTo(*dbg_patch);
// 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
- FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
+ std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
// get color rgb features (simple r,g,b channels)
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);
}
}
std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
}
- BIG_BATCH_OMP_ORDERED
- vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
- return;
+
+ hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
+
+ 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 result;
}
cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
float *row_ptr = labels.ptr<float>(j);
double y_s = y * y;
for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
- row_ptr[i] = float(std::exp(-0.5 * (y_s + x * x) / sigma_s)); //-1/2*e^((y^2+x^2)/sigma^2)
+ row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
}
}
// rotate so that 1 is at top-left corner (see KCF paper for explanation)
-#ifdef CUFFT
- cv::Mat tmp = circshift(labels, range_x[0], range_y[0]);
- tmp.copyTo(p_rot_labels);
-
- assert(p_rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
- return tmp;
-#else
- cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
+ MatDynMem rot_labels = circshift(labels, range_x[0], range_y[0]);
// sanity check, 1 at top left corner
assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
return rot_labels;
-#endif
}
-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);
+ cv::Mat rot_patch(patch.size(), patch.type());
+ cv::Mat tmp_x_rot(patch.size(), patch.type());
// circular rotate x-axis
if (x_rot < 0) {
// 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, double angle) const
{
cv::Mat patch;
- int x1 = cx - width / 2;
- int y1 = cy - height / 2;
- int x2 = cx + width / 2;
- int y2 = cy + height / 2;
+ cv::Size sz(width, height);
+ cv::RotatedRect rr(cv::Point2f(cx, cy), sz, angle);
+ cv::Rect bb = rr.boundingRect();
+
+ int x1 = bb.tl().x;
+ int y1 = bb.tl().y;
+ int x2 = bb.br().x;
+ int y2 = bb.br().y;
// out of image
if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
// cv::waitKey();
}
+ cv::Point2f src_pts[4];
+ cv::RotatedRect(cv::Point2f(patch.size()) / 2.0, sz, angle).points(src_pts);
+ cv::Point2f dst_pts[3] = { cv::Point2f(0, height), cv::Point2f(0, 0), cv::Point2f(width, 0)};
+ auto rot = cv::getAffineTransform(src_pts, dst_pts);
+ cv::warpAffine(patch, patch, rot, sz);
+
// sanity check
assert(patch.cols == width && patch.rows == height);
return patch;
}
-void KCF_Tracker::gaussian_correlation(struct ThreadCtx &vars, const ComplexMat &xf, const ComplexMat &yf,
- double sigma, bool auto_correlation)
+void KCF_Tracker::GaussianCorrelation::operator()(ComplexMat &result, const ComplexMat &xf, const ComplexMat &yf,
+ double sigma, bool auto_correlation, const KCF_Tracker &kcf)
{
-#ifdef CUFFT
- xf.sqr_norm(vars.xf_sqr_norm.deviceMem());
- if (!auto_correlation) yf.sqr_norm(vars.yf_sqr_norm.deviceMem());
-#else
- xf.sqr_norm(vars.xf_sqr_norm.hostMem());
+ TRACE("");
+ DEBUG_PRINTM(xf);
+ DEBUG_PRINT(xf_sqr_norm.num_elem);
+ xf.sqr_norm(xf_sqr_norm);
+ for (uint s = 0; s < xf.n_scales; ++s)
+ DEBUG_PRINT(xf_sqr_norm[s]);
if (auto_correlation) {
- vars.yf_sqr_norm.hostMem()[0] = vars.xf_sqr_norm.hostMem()[0];
+ yf_sqr_norm = xf_sqr_norm;
} else {
- yf.sqr_norm(vars.yf_sqr_norm.hostMem());
- }
-#endif
- vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
- DEBUG_PRINTM(vars.xyf);
- fft.inverse(vars.xyf, vars.ifft2_res, m_use_cuda ? vars.data_i_features.deviceMem() : nullptr, vars.stream);
-#ifdef CUFFT
- if (auto_correlation)
- cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.xf_sqr_norm.deviceMem(),
- sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
- else
- cuda_gaussian_correlation(vars.data_i_features.deviceMem(), vars.gauss_corr_res.deviceMem(), vars.xf_sqr_norm.deviceMem(), vars.yf_sqr_norm.deviceMem(),
- sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width, vars.stream);
-#else
- // ifft2 and sum over 3rd dimension, we dont care about individual channels
- DEBUG_PRINTM(vars.ifft2_res);
- cv::Mat xy_sum;
- if (xf.channels() != p_num_scales * p_num_of_feats)
- xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
- else
- xy_sum.create(vars.ifft2_res.size(), CV_32FC(int(p_scales.size())));
- xy_sum.setTo(0);
- 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 < 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 * 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(yf);
+ yf.sqr_norm(yf_sqr_norm);
}
- DEBUG_PRINTM(xy_sum);
+ for (uint s = 0; s < yf.n_scales; ++s)
+ DEBUG_PRINTM(yf_sqr_norm[s]);
+ xyf = auto_correlation ? xf.sqr_mag() : xf * yf.conj(); // xf.muln(yf.conj());
+ DEBUG_PRINTM(xyf);
- std::vector<cv::Mat> scales;
- cv::split(xy_sum, scales);
+ // ifft2 and sum over 3rd dimension, we dont care about individual channels
+ ComplexMat xyf_sum = xyf.sum_over_channels();
+ DEBUG_PRINTM(xyf_sum);
+ kcf.fft.inverse(xyf_sum, ifft_res);
+ DEBUG_PRINTM(ifft_res);
float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
- for (uint i = 0; i < uint(xf.n_scales); ++i) {
- cv::Mat in_roi(vars.in_all, cv::Rect(0, int(i) * scales[0].rows, scales[0].cols, scales[0].rows));
- cv::exp(
- -1. / (sigma * sigma) *
- cv::max((double(vars.xf_sqr_norm.hostMem()[i] + vars.yf_sqr_norm.hostMem()[0]) - 2 * scales[i]) * double(numel_xf_inv), 0),
- in_roi);
- DEBUG_PRINTM(in_roi);
+ for (uint i = 0; i < xf.n_scales; ++i) {
+ cv::Mat plane = ifft_res.plane(i);
+ DEBUG_PRINT(ifft_res.plane(i));
+ cv::exp(-1. / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * ifft_res.plane(i))
+ * numel_xf_inv, 0), plane);
+ DEBUG_PRINTM(plane);
}
-#endif
- DEBUG_PRINTM(vars.in_all);
- fft.forward(vars.in_all, auto_correlation ? vars.kf : vars.kzf, m_use_cuda ? vars.gauss_corr_res.deviceMem() : nullptr,
- vars.stream);
- return;
+
+ kcf.fft.forward(ifft_res, result);
}
float get_response_circular(cv::Point2i &pt, cv::Mat &response)
{
int x = pt.x;
int y = pt.y;
+ assert(response.dims == 2); // ensure .cols and .rows are valid
if (x < 0) x = response.cols + x;
if (y < 0) y = response.rows + y;
if (x >= response.cols) x = x - response.cols;
return response.at<float>(y, x);
}
-cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response)
+cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const
{
// find neighbourhood of max_loc (response is circular)
// 1 2 3
return sub_peak;
}
-double KCF_Tracker::sub_grid_scale(int index)
+double KCF_Tracker::sub_grid_scale(uint max_index)
{
cv::Mat A, fval;
- if (index < 0 || index > int(p_scales.size()) - 1) {
+ const auto &vec = d->IF_BIG_BATCH(threadctxs[0].max, threadctxs);
+ uint index = vec.getScaleIdx(max_index);
+ uint angle_idx = vec.getAngleIdx(max_index);
+
+ if (index >= vec.size()) {
// interpolate from all values
// fit 1d quadratic function f(x) = a*x^2 + b*x + c
- A.create(int(p_scales.size()), 3, CV_32FC1);
- fval.create(int(p_scales.size()), 1, CV_32FC1);
- for (auto it = p_scale_vars.begin(); it != p_scale_vars.end(); ++it) {
- uint i = uint(std::distance(p_scale_vars.begin(), it));
- int j = int(i);
- A.at<float>(j, 0) = float(p_scales[i] * p_scales[i]);
- A.at<float>(j, 1) = float(p_scales[i]);
- A.at<float>(j, 2) = 1;
- fval.at<float>(j) =
- m_use_big_batch ? float(p_scale_vars.back()->max_responses[i]) : float((*it)->max_response);
+ A.create(p_scales.size(), 3, CV_32FC1);
+ fval.create(p_scales.size(), 1, CV_32FC1);
+ for (size_t i = 0; i < p_scales.size(); ++i) {
+ A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
+ A.at<float>(i, 1) = float(p_scales[i]);
+ A.at<float>(i, 2) = 1;
+ fval.at<float>(i) = d->IF_BIG_BATCH(threadctxs[0].max[i].response, threadctxs(i, angle_idx).max.response);
}
} else {
// only from neighbours
- if (index == 0 || index == int(p_scales.size()) - 1) return p_scales[uint(index)];
-
- A = (cv::Mat_<float>(3, 3) << p_scales[uint(index) - 1] * p_scales[uint(index) - 1], p_scales[uint(index) - 1],
- 1, p_scales[uint(index)] * p_scales[uint(index)], p_scales[uint(index)], 1,
- p_scales[uint(index) + 1] * p_scales[uint(index) + 1], p_scales[uint(index) + 1], 1);
- auto it1 = p_scale_vars.begin();
- std::advance(it1, index - 1);
- auto it2 = p_scale_vars.begin();
- std::advance(it2, index);
- auto it3 = p_scale_vars.begin();
- std::advance(it3, index + 1);
- fval = (cv::Mat_<float>(3, 1) << (m_use_big_batch ? p_scale_vars.back()->max_responses[uint(index) - 1]
- : (*it1)->max_response),
- (m_use_big_batch ? p_scale_vars.back()->max_responses[uint(index)] : (*it2)->max_response),
- (m_use_big_batch ? p_scale_vars.back()->max_responses[uint(index) + 1] : (*it3)->max_response));
+ if (index == 0 || index == p_scales.size() - 1)
+ return p_scales[index];
+
+ A = (cv::Mat_<float>(3, 3) <<
+ p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
+ p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
+ p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
+#ifdef BIG_BATCH
+ fval = (cv::Mat_<float>(3, 1) <<
+ d->threadctxs[0].max(index - 1, angle_idx).response,
+ d->threadctxs[0].max(index + 0, angle_idx).response,
+ d->threadctxs[0].max(index + 1, angle_idx).response);
+#else
+ fval = (cv::Mat_<float>(3, 1) <<
+ d->threadctxs(index - 1, angle_idx).max.response,
+ d->threadctxs(index + 0, angle_idx).max.response,
+ d->threadctxs(index + 1, angle_idx).max.response);
+#endif
}
cv::Mat x;
cv::solve(A, fval, x, cv::DECOMP_SVD);
float a = x.at<float>(0), b = x.at<float>(1);
- double scale = p_scales[uint(index)];
- if (a > 0 || a < 0) scale = double(-b / (2 * a));
+ double scale = p_scales[index];
+ if (a > 0 || a < 0)
+ scale = -b / (2 * a);
return scale;
}