#include "agglomerative_clustering.h" void write(FileStorage& fs, const std::string&, const HCluster& x) { x.write(fs); } void read(const FileNode& node, HCluster& x, const HCluster& default_value) { if(node.empty()) x = default_value; else x.read(node); } HierarchicalClustering::HierarchicalClustering(vector &_regions): regions(_regions) { #ifndef _TRAIN_ boost = StatModel::load( "./trained_boost_groups.xml" ); if( boost.empty() ) { cout << "Could not read the classifier ./trained_boost_groups.xml" << endl; CV_Error(Error::StsBadArg, "Could not read the default classifier!"); } #endif } //For feature space void HierarchicalClustering::operator()(t_float *data, unsigned int num, int dim, unsigned char method, unsigned char metric, vector &merge_info, float x_coord_mult, cv::Size imsize) { t_float *Z = (t_float*)malloc(((num-1)*4) * sizeof(t_float)); // we need 4 floats foreach merge linkage_vector(data, (int)num, dim, Z, method, metric); build_merge_info(Z, data, (int)num, dim, merge_info, x_coord_mult, imsize); free(Z); } void HierarchicalClustering::build_merge_info(t_float *Z, t_float *X, int N, int dim, vector &merge_info, float x_coord_mult, cv::Size imsize) { int this_node = 0; // walk the whole dendogram for (int i=0; i<(N-1)*4; i=i+4) { HCluster cluster; cluster.childs.push_back(this_node); cluster.num_elem = Z[i+3]; //number of elements int node1 = Z[i]; int node2 = Z[i+1]; float dist = Z[i+2]; if (node1 point; //for (int n=0; n point; //for (int n=0; nN) && (cluster.rect == merge_info.at(node1-N).rect)) ) cluster.inherit_cnn_probability = 1; else if ( ((node2N) && (cluster.rect == merge_info.at(node2-N).rect)) ) cluster.inherit_cnn_probability = 2; else cluster.inherit_cnn_probability = 0; cluster.node1 = node1-N; cluster.node2 = node2-N; //Minibox mb; //for (int i=0; i= 1) volume = 0.999999; //TODO this may never happen!!! if (volume == 0) volume = 0.000001; //TODO is this the minimum we can get? // better if we just quantize to a given grid of possible volumes ... cluster.nfa = (int)NFA( N, cluster.points.size(), (double) volume, 0); //this uses an approximation for the nfa calculations (faster) /* predict group class with boost */ int nfa2 = (int)NFA( N, cluster.points.size(), (double) ext_volume, 0); //this uses an approximation for the nfa calculations (faster) cluster.feature_vector.push_back(cluster.elements.size()); cluster.feature_vector.push_back(cluster.nfa); cluster.feature_vector.push_back(nfa2); if (cluster.nfa == 0) cluster.feature_vector.push_back(-1.0); else cluster.feature_vector.push_back((float)nfa2/cluster.nfa); cluster.feature_vector.push_back((float)(volume/ext_volume)); cluster.feature_vector.push_back((float)centers_rect.width/cluster.rect.width); int left_diff = centers_rect.x - cluster.rect.x; int right_diff = (cluster.rect.x+cluster.rect.width) - (centers_rect.x+centers_rect.width); cluster.feature_vector.push_back((float)left_diff/cluster.rect.width); cluster.feature_vector.push_back((float)right_diff/cluster.rect.width); if (max(left_diff,right_diff) == 0) cluster.feature_vector.push_back(1.0); else cluster.feature_vector.push_back((float)min(left_diff,right_diff)/max(left_diff,right_diff)); cluster.feature_vector.push_back((float)centers_rect.height/cluster.rect.height); int top_diff = centers_rect.y - cluster.rect.y; int bottom_diff = (cluster.rect.y+cluster.rect.height) - (centers_rect.y+centers_rect.height); cluster.feature_vector.push_back((float)top_diff/cluster.rect.height); cluster.feature_vector.push_back((float)bottom_diff/cluster.rect.height); if (max(top_diff,bottom_diff) == 0) cluster.feature_vector.push_back(1.0); else cluster.feature_vector.push_back((float)min(top_diff,bottom_diff)/max(top_diff,bottom_diff)); Scalar mean,std; meanStdDev( cluster.elements_diameters, mean, std ); cluster.feature_vector.push_back(std[0]/mean[0]); meanStdDev( cluster.elements_strokes, mean, std ); cluster.feature_vector.push_back(std[0]/mean[0]); meanStdDev( cluster.elements_gradients, mean, std ); cluster.feature_vector.push_back(std[0]); meanStdDev( cluster.elements_intensities, mean, std ); cluster.feature_vector.push_back(std[0]); meanStdDev( cluster.elements_b_intensities, mean, std ); cluster.feature_vector.push_back(std[0]); #ifndef _TRAIN_ float votes_group = boost->predict( Mat(cluster.feature_vector), noArray(), DTrees::PREDICT_SUM | StatModel::RAW_OUTPUT); cluster.probability = (double)1-(double)1/(1+exp(-2*votes_group)); #endif merge_info.push_back(cluster); this_node++; } }