/**
* trains an svm model using the InriaPerson training dataset
* choses a number of negative detections randomly chosen from each image
* saves the freshly trained model to the harddisk
* @param string modelFile: path to the model file
*/
void trainSVM(string modelFile){
vector<string> pos_examples;
vector<string> neg_examples;
Classifier classifier = Classifier();
RNG random_index = RNG( time (NULL) );
Rect roi(Point(16,16), Point(80, 144));
Preprocessing::loadPathsByDirectory(TRAIN_POS_NORMALIZED, pos_examples);
Preprocessing::loadPathsByDirectory(TRAIN_NEG_ORIGINAL, neg_examples);
for(vector<string>::iterator posIt = pos_examples.begin(); posIt != pos_examples.end(); ++posIt)
{
OriginalImage img( *posIt );
( img ).image = ( img ).image(roi);
FeatureExtraction::extractHOGFeatures( img );
classifier.addPositiveExample(img.hog_features);
}
for(vector<string>::iterator negIt = neg_examples.begin(); negIt != neg_examples.end(); ++negIt)
{
OriginalImage img( *negIt );
FeatureExtraction::computeHOGPyramid( img );
FeatureExtraction::computeSlidingWindows( img );
for(std::vector<Image>::iterator scaleIt = ( img ).lower_images.begin(); scaleIt != ( img ).lower_images.end(); ++scaleIt)
{
FeatureExtraction::computeSlidingWindows(*scaleIt);
}
int count = 0;
double howMuch = ( (double)( MAXRANDOMELEMENTS ) / (double)( img.lower_images.size() + 1 ) );
for( int i = 0; i < (int)( howMuch ); i++ )
{
classifier.addNegativeExample( img.slidingWindows[ random_index.uniform( 0, img.slidingWindows.size() ) ].hog_features );
count++;
}
howMuch = howMuch + ( (double)( MAXRANDOMELEMENTS ) / (double)( img.lower_images.size() + 1) ) - (int)howMuch;
if( howMuch < MAXRANDOMELEMENTS )
for( int current = 0; current < img.lower_images.size(); current++ )
{
for( int i = 0; i < (int)( howMuch ); i++ )
{
int y = random_index.uniform( 0, img.lower_images[ current ].slidingWindows.size() );
Mat hog_features = img.lower_images[ current ].slidingWindows[ y ].hog_features;
classifier.addNegativeExample( hog_features );
count++;
}
howMuch = howMuch + ( (double)( MAXRANDOMELEMENTS ) / (double)( img.lower_images.size() + 1) ) - (int)howMuch;
}
while( count < MAXRANDOMELEMENTS )
{
int x = random_index.uniform( 0, img.lower_images.size() );
int y = random_index.uniform( 0, img.lower_images[ x ].slidingWindows.size() );
Mat hog_features = img.lower_images[ x ].slidingWindows[ y ].hog_features;
classifier.addNegativeExample( hog_features );
count++;
}
}
classifier.train( MODEL_STANDARD_FILE );
classifier.loadModel( MODEL_STANDARD_FILE );
for(vector<string>::iterator negIt = neg_examples.begin(); negIt != neg_examples.end(); ++negIt)
{
OriginalImage img( *negIt );
FeatureExtraction::computeHOGPyramid( img );
FeatureExtraction::computeSlidingWindows( img );
for( vector<SlidingWindow>::iterator it= img.slidingWindows.begin(); it != img.slidingWindows.end(); ++it )
{
if( classifier.classify( ( *it ).hog_features ) > 0.0 )
{
classifier.addNegativeExample( (*it ).hog_features );
}
}
for(std::vector<Image>::iterator scaleIt = ( img ).lower_images.begin(); scaleIt != ( img ).lower_images.end(); ++scaleIt)
{
FeatureExtraction::computeSlidingWindows(*scaleIt);
for( vector<SlidingWindow>::iterator it= ( *scaleIt ).slidingWindows.begin(); it != ( *scaleIt ).slidingWindows.end(); ++it )
{
if( classifier.classify( ( *it ).hog_features ) > 0.0 )
{
classifier.addNegativeExample( (*it ).hog_features );
}
}
}
}
classifier.train( MODEL_HARD_EX_FILE );
}
