void HandPairCollection::makePairs( const cv::PCA &pca,
const ContourSet &singles ) {
int size = singles.size();
hands.reserve( size * size + 2 );
centers.reserve( size * size + 2 );
for( int dom = 0; dom < size; ++dom ) {
for( int weak = 0; weak < size; ++weak ) {
Projection proj;
CenterPair centerp;
if( weak != dom ) {
proj = pca.project( flattenHistogram(
generateHandHistogram( singles[dom], &singles[weak]) ) );
centerp = std::make_pair( ::center( singles[dom] ),
new CenterPoint( ::center( singles[weak] ) ) );
}
else {
proj = pca.project(
flattenHistogram( generateHandHistogram( singles[dom], (Contour *)0 ) ));
centerp = std::make_pair( ::center( singles[dom] ), (CenterPoint *)0 );
}
hands.push_back( proj );
centers.push_back( centerp );
}
}
}
void testSVM(vector<string>& pPositiveFiles, vector<string>& pNegativeFiles, Descriptor_Hog& pDescriptor, CvSVM& pSvm, cv::PCA pPca, float pPortion = 1.0)
{
float portion;
if (pPortion < 1.f)
portion = pPortion;
else
portion = 0.f;
// We compute the total detection time (descriptor creation + prediction)
unsigned long long totalTime = 0;
unsigned long long chronoTime;
auto chronoStart = std::chrono::high_resolution_clock::now();
/***/
if (!gTable)
cout << "Testing positive files... " << flush;
int positive = 0;
int negative = 0;
int portionOfPositiveFiles = (int)((float)pPositiveFiles.size() * portion);
for (int i = portionOfPositiveFiles; i < pPositiveFiles.size(); ++i)
{
cv::Mat image = cv::imread(pPositiveFiles[i]);
chronoStart = std::chrono::high_resolution_clock::now();
chronoTime = chrono::duration_cast<chrono::microseconds>(chronoStart.time_since_epoch()).count();
pDescriptor.setImage(image);
vector<float> description = pDescriptor.getDescriptor(_roiPosition);
cv::Mat descriptionMat(1, (int)description.size(), CV_32FC1, &description[0]);
// If we used PCA for training
if (gPca < 1.0)
descriptionMat = pPca.project(descriptionMat);
float value = pSvm.predict(descriptionMat, true);
if (gVerbose)
cout << pPositiveFiles[i] << " -> " << value << endl;
if (value < -gMargin)
positive++;
totalTime += timeSince(chronoTime);
}
/***/
if (!gTable)
cout << "Testing negative files" << endl;
int totalNegatives = 0;
int portionOfNegativeFiles = (int)((float)pNegativeFiles.size() * portion);
for (int i = portionOfNegativeFiles; i < pNegativeFiles.size(); ++i)
{
cv::Mat image = cv::imread(pNegativeFiles[i]);
chronoStart = std::chrono::high_resolution_clock::now();
chronoTime = chrono::duration_cast<chrono::microseconds>(chronoStart.time_since_epoch()).count();
pDescriptor.setImage(image);
for (int x = 0; x < image.cols - _roiSize.x; x += image.cols/5)
{
for (int y = 0; y < image.rows - _roiSize.y; y += image.rows/5)
{
vector<float> description = pDescriptor.getDescriptor(_roiPosition + cv::Point_<int>(x, y));
cv::Mat descriptionMat(1, (int)description.size(), CV_32FC1, &description[0]);
if (description.size() == 0)
continue;
// If we used PCA for training
if (gPca < 1.0)
descriptionMat = pPca.project(descriptionMat);
float value = pSvm.predict(descriptionMat, true);
if (gVerbose)
cout << pNegativeFiles[i] << " -> " << value << endl;
totalNegatives++;
if (value > -gMargin)
negative++;
}
}
totalTime += timeSince(chronoTime);
}
if (!gTable)
{
cout << "Positive: " << positive << " / " << pPositiveFiles.size() - portionOfPositiveFiles << endl;
cout << "Negative: " << negative << " / " << totalNegatives << endl;
cout << "Time per prediction (us): " << totalTime / (pPositiveFiles.size() - portionOfPositiveFiles + totalNegatives) << endl;
}
else
{
cout << gIterations << " ";
cout << gEpsilon << " ";
cout << gCPenalty << " ";
cout << gOutput << " ";
cout << gPosition << " ";
cout << gCellSize << " ";
cout << gBlockSize << " ";
cout << gRoiSize << " ";
cout << gBins << " ";
cout << gSigma << " ";
cout << gPca << " ";
cout << totalTime / (pPositiveFiles.size() - portionOfPositiveFiles + totalNegatives) << " ";
cout << (float)positive / ((float)pPositiveFiles.size() - (float)portionOfPositiveFiles) << " " << (float)negative/(float)totalNegatives << endl;
}
}
void trainSVM(vector<string>& pPositiveFiles, vector<string>& pNegativeFiles, Descriptor_Hog& pDescriptor, CvSVM& pSvm, cv::PCA& pPca, float pPortion = 1.0)
{
int result;
if (!gTable)
{
cout << "Training parameters: " << endl;
if (_svmCriteria == CV_TERMCRIT_ITER)
cout << " iterations = " << gIterations << endl;
if (_svmCriteria == CV_TERMCRIT_EPS)
cout << " epsilon = " << gEpsilon << endl;
cout << " C penalty = " << gCPenalty << endl;
cout << " output file = " << gOutput << endl;
cout << " ROI position = " << gPosition << endl;
cout << " cell size = " << gCellSize << endl;
cout << " block size = " << gBlockSize << endl;
cout << " roi size = " << gRoiSize << endl;
cout << " bins per cell = " << gBins << endl;
cout << " sigma = " << gSigma << endl;
cout << " pca = " << gPca << endl;
}
// Set up training data
if (!gVerbose && !gTable)
cout << "Loading training data... " << flush;
int nbrImg = 0;
vector<float> labels;
vector<float> trainingData;
int portionOfPositiveFiles = (int)((float)pPositiveFiles.size() * pPortion);
for (int i = 0; i < portionOfPositiveFiles; ++i)
{
if (gVerbose)
cout << "Analysing " << pPositiveFiles[i] << endl;
cv::Mat image = cv::imread(pPositiveFiles[i]);
pDescriptor.setImage(image);
vector<float> description = pDescriptor.getDescriptor(_roiPosition);
if (description.size() == 0)
continue;
labels.push_back(1.f);
for (int j = 0; j < description.size(); ++j)
trainingData.push_back(description[j]);
nbrImg++;
}
if (!gVerbose && !gTable)
cout << "Positive data loaded... " << flush;
int portionOfNegativeFiles = (int)((float)pNegativeFiles.size() * pPortion);
for (int i = 0; i < portionOfNegativeFiles; ++i)
{
if (gVerbose)
cout << "Analysing " << pNegativeFiles[i] << endl;
cv::Mat image = cv::imread(pNegativeFiles[i]);
pDescriptor.setImage(image);
for (int x = 0; x < image.cols - _roiSize.x; x += image.cols/5)
{
for (int y = 0; y < image.rows - _roiSize.y; y += image.rows/5)
{
vector<float> description = pDescriptor.getDescriptor(_roiPosition + cv::Point_<int>(x, y));
if (description.size() == 0)
continue;
labels.push_back(-1.f);
for (int j = 0; j < description.size(); ++j)
trainingData.push_back(description[j]);
nbrImg++;
}
}
}
if (!gVerbose && !gTable)
cout << "Negative data loaded" << endl;
// Applying the PCA and reformating input data
if (gPca < 1.0)
{
cout << "Applying PCA before SVM train... ";
cout.flush();
int vectorSize = trainingData.size() / nbrImg;
cv::Mat trainingPCA = cv::Mat::zeros(nbrImg, vectorSize, CV_32F);
for (int i = 0; i < nbrImg; ++i)
{
for (int j = 0; j < vectorSize; ++j)
{
trainingPCA.at<float>(i, j) = trainingData[i*vectorSize + j];
}
}
int keptComponents = (int)((float)vectorSize * gPca);
pPca(trainingPCA, cv::noArray(), CV_PCA_DATA_AS_ROW, keptComponents);
// Reproject all vectors
cv::Mat projected;
pPca.project(trainingPCA, projected);
// Store the result as the training data
trainingData.clear();
for (int i = 0; i < projected.rows; ++i)
for (int j = 0; j < projected.cols; ++j)
{
trainingData.push_back(projected.at<float>(i, j));
}
cout << "Kept " << keptComponents << " components" << endl;
// Save the resulting space
cv::FileStorage file("pca.xml", cv::FileStorage::WRITE);
file << "eigenVectors" << pPca.eigenvectors;
file << "mean" << pPca.mean;
}
if (!gTable)
cout << "Training the SVM classifier..." << endl;
// Set up the SVM train parameter
CvSVMParams svmParams;
svmParams.svm_type = CvSVM::C_SVC;
svmParams.C = gCPenalty;
svmParams.kernel_type = CvSVM::LINEAR;
svmParams.term_crit = cvTermCriteria(_svmCriteria, gIterations, gEpsilon);
// Train the SVM
cv::Mat labelsMat((int)labels.size(), 1, CV_32FC1, &labels[0]);
cv::Mat trainingMat(nbrImg, (int)trainingData.size() / nbrImg, CV_32FC1, &trainingData[0]);
result = pSvm.train(trainingMat, labelsMat, cv::Mat(), cv::Mat(), svmParams);
pSvm.save((const char*)gOutput);
}
