Extend to support not only multiple scales but also multiple angles

[hercules2020/kcf.git] / src / kcf.h

diff --git a/src/kcf.h b/src/kcf.h
index 4ca1066e033efeb732bb65b0de476f2df43d6338..76abc5ea4ea48989f9dcfe7693268b857bc43d7b 100644 (file)
--- a/src/kcf.h
+++ b/src/kcf.h
@@ -6,13 +6,10 @@
 #include <memory>
 #include "fhog.hpp"
 
+#include "complexmat.hpp"
 #ifdef CUFFT
-#include "complexmat.cuh"
-#include "cuda_functions.cuh"
 #include "cuda_error_check.hpp"
 #include <cuda_runtime.h>
-#else
-#include "complexmat.hpp"
 #endif
 
 #include "cnfeat.hpp"
@@ -24,7 +21,7 @@ struct ThreadCtx;
 
 struct BBox_c
 {
-    double cx, cy, w, h;
+    double cx, cy, w, h, a;
 
     inline cv::Point2d center() const { return cv::Point2d(cx, cy); }
 
@@ -36,18 +33,6 @@ struct BBox_c
         h  *= factor;
     }
 
-    inline void scale_x(double factor)
-    {
-        cx *= factor;
-        w  *= factor;
-    }
-
-    inline void scale_y(double factor)
-    {
-        cy *= factor;
-        h  *= factor;
-    }
-
     inline cv::Rect get_rect()
     {
         return cv::Rect(int(cx-w/2.), int(cy-h/2.), int(w), int(h));
@@ -58,6 +43,7 @@ struct BBox_c
 class KCF_Tracker
 {
     friend ThreadCtx;
+    friend Kcf_Tracker_Private;
 public:
     bool m_debug {false};
     bool m_visual_debug {false};
@@ -101,15 +87,13 @@ private:
     // Information to calculate current pose of the tracked object
     cv::Point2d p_current_center;
     double p_current_scale = 1.;
+    double p_current_angle = 0.;
 
     double max_response = -1.;
 
     bool p_resize_image = false;
-    bool p_fit_to_pw2 = false;
 
     const double p_downscale_factor = 0.5;
-    double p_fit_factor_x = 1;
-    double p_fit_factor_y = 1;
     const double p_floating_error = 0.0001;
 
     const double p_padding = 1.5;
@@ -118,7 +102,8 @@ private:
     const double p_kernel_sigma = 0.5;    //def = 0.5
     const double p_lambda = 1e-4;         //regularization in learning step
     const double p_interp_factor = 0.02;  //def = 0.02, linear interpolation factor for adaptation
-    cv::Size p_windows_size;
+    cv::Size p_windows_size;              // size of the patch to find the tracked object in
+    cv::Size fit_size;                    // size to which rescale the patch for better FFT performance
 
     const uint p_num_scales = m_use_scale ? 7 : 1;
     const double p_scale_step = 1.02;
@@ -127,26 +112,44 @@ private:
 
     const uint p_num_angles = 1;
     const int p_angle_step = 10;
-    std::vector<double> p_angles = {0};
+    std::vector<double> p_angles;
 
     const int p_num_of_feats = 31 + (m_use_color ? 3 : 0) + (m_use_cnfeat ? 10 : 0);
     cv::Size feature_size;
 
-    Kcf_Tracker_Private &d;
+    std::unique_ptr<Kcf_Tracker_Private> d;
+
+    class Model {
+        cv::Size feature_size;
+        uint height, width, n_feats;
+    public:
+        ComplexMat yf {height, width, 1};
+        ComplexMat model_alphaf {height, width, 1};
+        ComplexMat model_alphaf_num {height, width, 1};
+        ComplexMat model_alphaf_den {height, width, 1};
+        ComplexMat model_xf {height, width, n_feats};
+        ComplexMat xf {height, width, n_feats};
+
+        // Temporary variables for trainig
+        MatScaleFeats patch_feats{1, n_feats, feature_size};
+        MatScaleFeats temp{1, n_feats, feature_size};
+
+
+
+        Model(cv::Size feature_size, uint _n_feats)
+            : feature_size(feature_size)
+            , height(Fft::freq_size(feature_size).height)
+            , width(Fft::freq_size(feature_size).width)
+            , n_feats(_n_feats) {}
+    };
 
-    //model
-    ComplexMat p_yf;
-    ComplexMat p_model_alphaf;
-    ComplexMat p_model_alphaf_num;
-    ComplexMat p_model_alphaf_den;
-    ComplexMat p_model_xf;
-    ComplexMat p_xf;
+    std::unique_ptr<Model> model;
 
     class GaussianCorrelation {
       public:
-        GaussianCorrelation(uint num_scales, cv::Size size)
+        GaussianCorrelation(uint num_scales, uint num_feats, cv::Size size)
             : xf_sqr_norm(num_scales)
-            , xyf(Fft::freq_size(size), 1, num_scales)
+            , xyf(Fft::freq_size(size), num_feats, num_scales)
             , ifft_res(num_scales, size)
             , k(num_scales, size)
         {}
@@ -162,12 +165,12 @@ private:
 
     //helping functions
     void scale_track(ThreadCtx &vars, cv::Mat &input_rgb, cv::Mat &input_gray);
-    cv::Mat get_subwindow(const cv::Mat &input, int cx, int cy, int size_x, int size_y) const;
+    cv::Mat get_subwindow(const cv::Mat &input, int cx, int cy, int size_x, int size_y, double angle) const;
     cv::Mat gaussian_shaped_labels(double sigma, int dim1, int dim2);
     std::unique_ptr<GaussianCorrelation> gaussian_correlation;
-    cv::Mat circshift(const cv::Mat &patch, int x_rot, int y_rot);
+    cv::Mat circshift(const cv::Mat &patch, int x_rot, int y_rot) const;
     cv::Mat cosine_window_function(int dim1, int dim2);
-    cv::Mat get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy, int size_x, int size_y, double scale) const;
+    cv::Mat 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;
     cv::Point2f sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const;
     double sub_grid_scale(uint index);
     void resizeImgs(cv::Mat &input_rgb, cv::Mat &input_gray);