int mainPCA(const vector<string> &args, const map<string, string> &opts)
{
if(args.size() < 4) {
throw std::runtime_error("ERROR: Incorrect number of arguments. Run command with flag -h for help.");
}
double t = (double)getTickCount();
LOG(INFO) << "Performing a PCA analysis on the images database";
string dbFName = args[2];
string pcaFName = args[3];
string category;
LOG(INFO) << "Obtaining feature extractor parameters";
ParametersMap featParams;
if(opts.count("-c") == 1) {
category = opts.at("-c");
featParams = FeatureExtractor::getDefaultParameters("hog");
} else {
throw std::runtime_error("ERROR: Incorrect number of arguments. Run command with flag -h for help.");
}
if(boost::filesystem::exists(dbFName))
{
LOG(INFO) << "Creating the image database";
PascalImageDatabase db(dbFName.c_str(), category);
cout << db << endl;
FeatureExtractor *featExtractor = FeatureExtractor::create(featParams);
LOG(INFO) << "Category: " << category;
LOG(INFO) << "Extracting HOG features";
FeatureCollection features;
(*featExtractor)(db, features);
LOG(INFO) << "Performing PCA on the obtained HOG features";
int num_samples = features.size();
int num_features = features[0].size();
Mat data(num_features,num_samples,CV_32FC1,Scalar(0));
PrincipalComponentAnalysis pca;
pca.pre_process(features,data);
pca.compute(data,db);
pca.savePCAFile(pcaFName);
t = (double)getTickCount() - t;
LOG(INFO) << "PCA completed in " << t/getTickFrequency() << " seconds.";
return EXIT_SUCCESS;
}
else
{
throw std::runtime_error("ERROR: Pascal database file doesn't exist in: " + dbFName);
}
}
std::vector<float>
SupportVectorMachine::predict(const FeatureCollection &fset) const
{
std::vector<float> preds(fset.size());
for(int i = 0; i < fset.size(); i++) {
preds[i] = predict(fset[i]);
}
return preds;
}
int SvmTrain( const ICDAR2011DataSet& dataset, FeatureCollection& features ) {
for (int j = 0;j < 3;j++) {
vector<int>false_idx;
false_idx.clear();
svm_train(features,dataset, false_idx);
random_shuffle(false_idx.begin(),false_idx.end());
for (int i = 0;i<100;i++) {
int idx = false_idx[i];
features.push_back(features[idx]);
}
random_shuffle(features.begin(),features.end());
}
}
void svm_train(const FeatureCollection& features,const ICDAR2011DataSet& dataset ,vector<int>& false_idx) {
size_t feature_count = features.size();
int validation_feature_count = 5000;
size_t dim = features[0].featureArray.size();
Mat train_data(feature_count - validation_feature_count, dim, CV_32F);
Mat train_label(feature_count - validation_feature_count, 1, CV_32F);
for(int feature_idx = 0; feature_idx < feature_count - validation_feature_count; feature_idx++) {
float sum_norm = 0;
if (NORM)
sum_norm = accumulate(features[feature_idx].featureArray.begin(),features[feature_idx].featureArray.end(),0.0);
else {
for (int i = 0;i< dim;i++)
sum_norm += features[feature_idx].featureArray[i] * features[feature_idx].featureArray[i];
sum_norm = sqrt(sum_norm);
}
for(int dim_idx = 0; dim_idx < dim; dim_idx++) {
FeatureAtPoint fea = features[feature_idx];
train_data.at<float>(feature_idx, dim_idx) = features[feature_idx].featureArray[dim_idx]/sum_norm;
}
train_label.at<float>(feature_idx) = features[feature_idx].label;
}
std::cout<<"svm train-------------------------------------"<<std::endl;
CvSVM SVM;
CvBoost boost;
CvBoostParams params = CvBoostParams(CvBoost::REAL, 50, 0.95, 5, false, 0 );
boost.train(train_data, CV_ROW_SAMPLE,train_label, Mat(), Mat(), Mat(),Mat(),params);
std::cout<<"svm validad-----------------------------------"<<std::endl;
// Calculate trainning error
cout<<feature_count<<endl;
Mat test_data(1, dim, CV_32F);
float predict_correct = 0, positive_cnt = 0, negative_cnt = 0;
for(int feature_idx = feature_count - validation_feature_count; feature_idx < feature_count; feature_idx++) {
float sum_norm = 0 ;
if (NORM)
sum_norm = accumulate(features[feature_idx].featureArray.begin(),features[feature_idx].featureArray.end(),0.0);
else {
for (int i = 0;i< dim;i++)
sum_norm += features[feature_idx].featureArray[i] * features[feature_idx].featureArray[i];
sum_norm = sqrt(sum_norm);
}
for(int dim_idx = 0; dim_idx < dim; dim_idx++) {
test_data.at<float>(0, dim_idx) = features[feature_idx].featureArray[dim_idx]/sum_norm;
}
if(features[feature_idx].label == boost.predict(test_data)) {
predict_correct++;
} else {
false_idx.push_back(feature_idx);
}
positive_cnt += features[feature_idx].label == 1 ? 1 : 0;
negative_cnt += features[feature_idx].label == 0 ? 1 : 0;
}
cout << "Training accuracy:" << predict_correct / 5000 << " pos_cnt:" << positive_cnt << " neg_cnt:" << negative_cnt << endl;
boost.save(((dataset.model_dir + "boost.model").c_str()));
}
int SvmTest( const ICDAR2011DataSet& dataset, FeatureCollection& features , const vector<int>&ProposalLen, const vector< vector<Rect> >&Proposal) {
long sumProposal = accumulate( ProposalLen.begin(), ProposalLen.end(), 0 );
assert( sumProposal == features.size() );
CvBoost boost;
boost.load((dataset.model_dir + "boost.model").c_str());
for( size_t image_idx = 0; image_idx < ProposalLen.size() ;image_idx ++ ) {
string filename = dataset.test_set[ image_idx ];
string resultPath = dataset.result_dir + CmFile::GetFileNameWithoutExtension( filename ) + ".txt";
cout<<"result save path "<<resultPath<<" "<<image_idx<<" of "<<dataset.test_num<<endl;
ofstream score_out( resultPath.c_str() );
int num = ProposalLen[ image_idx ];
vector<Rect>proposal = Proposal[ image_idx ];
assert( num == proposal.size() );
long StartIndex = 0;
for (int i = 0;i < num ;i++) {
FeatureAtPoint featurePoint = features[ i + StartIndex ];
int dim = featurePoint.featureArray.size();
Mat test_data(1, dim, CV_32F);
double sum_norm = 0;
if (NORM) {
sum_norm = accumulate(featurePoint.featureArray.begin(),featurePoint.featureArray.end(),0.0);
} else {
for (int i = 0;i< dim;i++)
sum_norm += featurePoint.featureArray[i] * featurePoint.featureArray[i];
sum_norm = sqrt(sum_norm);
}
vector<float>sample;
for (int dim_idx = 0;dim_idx < dim;dim_idx++) {
test_data.at<float>(0, dim_idx) = featurePoint.featureArray[dim_idx]/ sum_norm;
sample.push_back(featurePoint.featureArray[dim_idx]/sum_norm);
}
double score = boost.predict(test_data,Mat(),Range::all(),false,true);
// score = score * (-1);
score_out<< proposal[i].x << " " << proposal[i].y << " "
<< proposal[i].width << " " << proposal[i].height << " "
<< score << endl;
}
StartIndex += num;
score_out.close();
}
}
void
SupportVectorMachine::train(const std::vector<float> &labels, const FeatureCollection &fset, const Size &roiSize, double C)
{
if(labels.size() != fset.size()) throw CError("Database size is different from feature set size!");
_fVecShape = fset[0].Shape();
if(roiSize.width != 0 && roiSize.height != 0) {
_roiSize = roiSize;
} else {
_roiSize.width = _fVecShape.width;
_roiSize.height = _fVecShape.height;
}
// Figure out size and number of feature vectors
int nVecs = labels.size();
CShape shape = fset[0].Shape();
int dim = shape.width * shape.height * shape.nBands;
// Parameters for SVM
svm_parameter parameter;
parameter.svm_type = C_SVC;
parameter.kernel_type = LINEAR;
parameter.degree = 0;
parameter.gamma = 0;
parameter.coef0 = 0;
parameter.nu = 0.5;
parameter.cache_size = 100; // In MB
parameter.C = C;
parameter.eps = 1e-3;
parameter.p = 0.1;
parameter.shrinking = 1;
parameter.probability = 0;
parameter.nr_weight = 0; // ?
parameter.weight_label = NULL;
parameter.weight = NULL;
//cross_validation = 0;
// Allocate memory
svm_problem problem;
problem.l = nVecs;
problem.y = new double[nVecs];
problem.x = new svm_node*[nVecs];
if(_data) delete [] _data;
/******** BEGIN TODO ********/
// Copy the data used for training the SVM into the libsvm data structures "problem".
// Put the feature vectors in _data and labels in problem.y. Also, problem.x[k]
// should point to the address in _data where the k-th feature vector starts (i.e.,
// problem.x[k] = &_data[starting index of k-th feature])
//
// Hint:
// * Don't forget to set _data[].index to the corresponding dimension in
// the original feature vector. You also need to set _data[].index to -1
// right after the last element of each feature vector
// Vector containing all feature vectors. svm_node is a struct with
// two fields, index and value. Index entry indicates position
// in feature vector while value is the value in the original feature vector,
// each feature vector of size k takes up k+1 svm_node's in _data
// the last one being simply to indicate that the feature has ended by setting the index
// entry to -1
_data = new svm_node[nVecs * (dim + 1)];
// Position in the feature vector
int posFeatVec = 0;
// Counter for elements in _data
int dataCnt = 0;
// Loop through each feature vector
for (int i = 0; i < nVecs; i++)
{
// Set labels
problem.y[i] = labels[i];
// For each vector, loop through feature elements --> (x, y, band)
for (int y = 0; y < shape.height; y++)
{
for (int x = 0; x < shape.width; x++)
{
for (int band = 0; band < shape.nBands; band++)
{
// 0 ... (dim-1)
if (posFeatVec < dim)
{
_data[dataCnt].index = posFeatVec;
_data[dataCnt].value = fset[i].Pixel(x, y, band);
posFeatVec++;
}
// dim : end of current feature vector
// If just set last feature value, set end flag in the same iteration
if (posFeatVec == dim)
{
dataCnt++;
// Set the index entry to -1
_data[dataCnt].index = -1;
_data[dataCnt].value = -1;
// Position in a new feature vector
posFeatVec = 0;
}
// Ready for next element
dataCnt++;
}
// End of band loop
}
// End of x loop
}
// End of y loop
// Position where the i-th feature vector starts
problem.x[i] = &_data[i * (dim+1)];
}
// End of i loop
/******** END TODO ********/
// Train the model
if(_model != NULL) svm_free_and_destroy_model(&_model);
_model = svm_train(&problem, ¶meter);
// Cleanup
delete [] problem.y;
delete [] problem.x;
}
