Make Stump real-valued

apps/icf/acffeature.hpp

@@ -65,7 +65,7 @@ public:
     ACFFeatureEvaluator(const std::vector<Point>& features);
 
     /* Set channels for feature evaluation */
-    void setChannels(const std::vector<Mat> >& channels);
+    void setChannels(const std::vector<Mat>& channels);
 
     /* Set window position */
     void setPosition(Size position);

apps/icf/waldboost.cpp

@@ -1,3 +1,8 @@
+#include <cmath>
+
+#include <algorithm>
+using std::swap;
+
 #include "waldboost.hpp"
 
 using cv::Mat;
@@ -6,6 +11,7 @@ using cv::sort;
 using cv::sortIdx;
 using cv::adas::Stump;
 using cv::adas::WaldBoost;
+using cv::Ptr;
 
 /* Cumulative sum by rows */
 static void cumsum(const Mat_<float>& src, Mat_<float> dst)
@@ -81,28 +87,37 @@ int Stump::train(const Mat& data, const Mat& labels, const Mat& weights)
     cumsum(neg_weights, neg_cum_weights);
 
     /* Compute total weights of positive and negative samples */
-    int pos_total_weight = 0, neg_total_weight = 0;
-    for( int col = 0; col < labels.cols; ++col )
-    {
-        if( labels.at<int>(0, col) == +1)
-            pos_total_weight += weights.at<float>(0, col);
-        else
-            neg_total_weight += weights.at<float>(0, col);
-    }
+    float pos_total_weight = pos_cum_weights.at<float>(0, weights.cols - 1);
+    float neg_total_weight = neg_cum_weights.at<float>(0, weights.cols - 1);
+
+
+    float eps = 1. / 4 * labels.cols;
 
     /* Compute minimal error */
     float min_err = FLT_MAX;
     int min_row = -1;
     int min_col = -1;
     int min_polarity = 0;
+    float min_pos_value = 1, min_neg_value = -1;
+
     for( int row = 0; row < sorted_weights.rows; ++row )
     {
         for( int col = 0; col < sorted_weights.cols - 1; ++col )
         {
-            float err;
+            float err, h_pos, h_neg;
+
+            // Direct polarity
 
-            err = pos_cum_weights.at<float>(row, col) +
-                  (neg_total_weight - neg_cum_weights.at<float>(row, col));
+            float pos_wrong = pos_cum_weights.at<float>(row, col);
+            float pos_right = pos_total_weight - pos_wrong;
+
+            float neg_right = neg_cum_weights.at<float>(row, col);
+            float neg_wrong = neg_total_weight - neg_right;
+
+            h_pos = .5 * log((pos_right + eps) / (pos_wrong + eps));
+            h_neg = .5 * log((neg_wrong + eps) / (neg_right + eps));
+
+            err = sqrt(pos_right * neg_wrong) + sqrt(pos_wrong * neg_right);
 
             if( err < min_err )
             {
@@ -110,11 +125,19 @@ int Stump::train(const Mat& data, const Mat& labels, const Mat& weights)
                 min_row = row;
                 min_col = col;
                 min_polarity = +1;
+                min_pos_value = h_pos;
+                min_neg_value = h_neg;
             }
 
+            // Opposite polarity
+            swap(pos_right, pos_wrong);
+            swap(neg_right, neg_wrong);
+
+            h_pos = -h_pos;
+            h_neg = -h_neg;
+
+            err = sqrt(pos_right * neg_wrong) + sqrt(pos_wrong * neg_right);
 
-            err = (pos_total_weight - pos_cum_weights.at<float>(row, col)) +
-                  neg_cum_weights.at<float>(row, col);
 
             if( err < min_err )
             {
@@ -122,6 +145,8 @@ int Stump::train(const Mat& data, const Mat& labels, const Mat& weights)
                 min_row = row;
                 min_col = col;
                 min_polarity = -1;
+                min_pos_value = h_pos;
+                min_neg_value = h_neg;
             }
         }
     }
@@ -130,18 +155,13 @@ int Stump::train(const Mat& data, const Mat& labels, const Mat& weights)
     threshold_ = ( sorted_data.at<int>(min_row, min_col) +
                    sorted_data.at<int>(min_row, min_col + 1) ) / 2;
     polarity_ = min_polarity;
+    pos_value_ = min_pos_value;
+    neg_value_ = min_neg_value;
 
     return min_row;
 }
 
-static inline int sign(int value)
-{
-    if (value > 0)
-        return +1;
-    return -1;
-}
-
-int Stump::predict(int value)
+float Stump::predict(int value)
 {
-    return polarity_ * sign(value - threshold_);
+    return polarity_ * (value - threshold_) > 0 ? pos_value_ : neg_value_;
 }

apps/icf/waldboost.hpp

@@ -56,11 +56,13 @@ class Stump
 public:
 
     /* Initialize zero stump */
-    Stump(): threshold_(0), polarity_(1) {};
+    Stump(): threshold_(0), polarity_(1), pos_value_(1), neg_value_(-1) {}
 
-    /* Initialize stump with given threshold and polarity */
-    Stump(int threshold, int polarity): threshold_(threshold),
-        polarity_(polarity) {};
+    /* Initialize stump with given threshold, polarity
+        and classification values */
+    Stump(int threshold, int polarity, float pos_value, float neg_value):
+        threshold_(threshold), polarity_(polarity),
+        pos_value_(pos_value), neg_value_(neg_value) {}
 
     /* Train stump for given data
 
@@ -77,21 +79,20 @@ public:
 
     /* Predict object class given
 
-        value — feature value. Feature must be the same as chose during training
-        stump
+        value — feature value. Feature must be the same as was chosen
+        during training stump
 
-    Returns object class from {-1, +1}
+    Returns real value, sign(value) means class
     */
-    int predict(int value);
+    float predict(int value);
 
 private:
     /* Stump decision threshold */
     int threshold_;
     /* Stump polarity, can be from {-1, +1} */
     int polarity_;
-    /* Stump decision rule:
-        h(value) = polarity * sign(value - threshold)
-    */
+    /* Classification values for positive and negative classes  */
+    float pos_value_, neg_value_;
 };
 
 struct WaldBoostParams