Update sample and code with external computation of HOG detector.

modules/ml/include/opencv2/ml.hpp

@@ -518,8 +518,7 @@ public:
     virtual CvSVMParams get_params() const { return params; };
     CV_WRAP virtual void clear();
 
-    // return a single vector for HOG detector.
-    virtual void get_svm_detector( std::vector< float > & detector ) const;
+    virtual const CvSVMDecisionFunc* get_decision_function() const { return decision_func; }
 
     static CvParamGrid get_default_grid( int param_id );
 

modules/ml/src/svm.cpp

@@ -1245,38 +1245,6 @@ const float* CvSVM::get_support_vector(int i) const
     return sv && (unsigned)i < (unsigned)sv_total ? sv[i] : 0;
 }
 
-void CvSVM::get_svm_detector( std::vector< float > & detector ) const
-{
-    CV_Assert( var_all > 0 &&
-               sv_total > 0 &&
-               sv != 0 &&
-               decision_func != 0 &&
-               decision_func->alpha != 0 &&
-               decision_func->sv_count == sv_total );
-    float svi = 0.f;
-
-    detector.clear(); //clear stuff in vector.
-    detector.reserve( var_all + 1 ); //reserve place for memory efficiency.
-
-    /**
-     * detector^i = \sum_j support_vector_j^i * \alpha_j
-     * detector^dim = -\rho
-     */
-    for( int i = 0 ; i < var_all ; ++i )
-    {
-        svi = 0.f;
-        for( int j = 0 ; j < sv_total ; ++j )
-        {
-            if( decision_func->sv_index != NULL ) // sometime the sv_index isn't store on YML/XML.
-                svi += (float)( sv[decision_func->sv_index[j]][i] * decision_func->alpha[ j ] );
-            else
-                svi += (float)( sv[j][i] * decision_func->alpha[ j ] );
-        }
-        detector.push_back( svi );
-    }
-    detector.push_back( (float)-decision_func->rho );
-}
-
 bool CvSVM::set_params( const CvSVMParams& _params )
 {
     bool ok = false;

samples/cpp/train_HOG.cpp

@@ -11,11 +11,53 @@ using namespace cv;
 using namespace std;
 
 
-/*
- * Convert training/testing set to be used by OpenCV Machine Learning algorithms.
- * TrainData is a matrix of size (#samples x max(#cols,#rows) per samples), in 32FC1.
- * Transposition of samples are made if needed.
+void get_svm_detector(const SVM& svm, vector< float > & hog_detector )
+{
+    // get the number of variables
+    const int var_all = svm.get_var_count();
+    // get the number of support vectors
+    const int sv_total = svm.get_support_vector_count();
+    // get the decision function
+    const CvSVMDecisionFunc* decision_func = svm.get_decision_function();
+    // get the support vectors
+    const float** sv = &(svm.get_support_vector(0)); 
+
+    CV_Assert( var_all > 0 &&
+        sv_total > 0 &&
+        decision_func != 0 &&
+        decision_func->alpha != 0 &&
+        decision_func->sv_count == sv_total );
+    
+    float svi = 0.f;
+
+    hog_detector.clear(); //clear stuff in vector.
+    hog_detector.reserve( var_all + 1 ); //reserve place for memory efficiency.
+
+     /**
+    * hog_detector^i = \sum_j support_vector_j^i * \alpha_j
+    * hog_detector^dim = -\rho
     */
+   for( int i = 0 ; i < var_all ; ++i )
+    {
+        svi = 0.f;
+        for( int j = 0 ; j < sv_total ; ++j )
+        {
+            if( decision_func->sv_index != NULL ) // sometime the sv_index isn't store on YML/XML.
+                svi += (float)( sv[decision_func->sv_index[j]][i] * decision_func->alpha[ j ] );
+            else
+                svi += (float)( sv[j][i] * decision_func->alpha[ j ] );
+        }
+        hog_detector.push_back( svi );
+    }
+    hog_detector.push_back( (float)-decision_func->rho );
+}
+
+
+/*
+* Convert training/testing set to be used by OpenCV Machine Learning algorithms.
+* TrainData is a matrix of size (#samples x max(#cols,#rows) per samples), in 32FC1.
+* Transposition of samples are made if needed.
+*/
 void convert_to_ml(const std::vector< cv::Mat > & train_samples, cv::Mat& trainData )
 {
     //--Convert data
@@ -322,7 +364,7 @@ void test_it( const Size & size )
     Scalar reference( 0, 255, 0 );
     Scalar trained( 0, 0, 255 );
     Mat img, draw;
-    SVM svm;
+    MySVM svm;
     HOGDescriptor hog;
     HOGDescriptor my_hog;
     my_hog.winSize = size;
@@ -333,7 +375,7 @@ void test_it( const Size & size )
     svm.load( "my_people_detector.yml" );
     // Set the trained svm to my_hog
     vector< float > hog_detector;
-    svm.get_svm_detector( hog_detector );
+    get_svm_detector( svm, hog_detector );
     my_hog.setSVMDetector( hog_detector );
     // Set the people detector.
     hog.setSVMDetector( hog.getDefaultPeopleDetector() );