void RPCA::fit(const cv::Mat1f & descr)
{
// make data zero mean
cv::reduce(descr, mean, 0, CV_REDUCE_AVG);
cv::Mat1f X(descr.rows, descr.cols);
#ifdef USE_TBB
tbb::parallel_for(int32_t(0), X.rows, [&](int32_t y) {
X.row(y) = descr.row(y) - mean;
}
);
#else
for( int32_t y = 0; y < X.rows; y++){
X.row(y) = descr.row(y) - mean;
}
#endif
cv::Mat1f U, S, V;
cv::SVD::compute(X, S, U, V);
cv::pow(S, 2, variance);
variance /= X.rows;
if (n_components <= 0) {
n_components = X.cols;
}
components = V.rowRange(0, n_components);
if(variance.rows < n_components) {
whiten = false;
} else {
variance = variance.rowRange(0, n_components);
}
this->dataLoaded = true;
}
void RPCA::transform(const cv::Mat1f & descr,
cv::Mat1f & out) const
{
if(descr.cols != mean.cols) {
std::stringstream s;
s << "Input and mean data missmatch." << std::endl;
s << "Input: " << descr.size() << " Mean: " << mean.size() << std::endl;
throw pl::DimensionalityReductionError(s.str(), currentMethod, currentLine);
} else if(descr.cols != components.cols) {
std::stringstream s;
s << "Input and transformation data missmatch." << std::endl;
s << "Input: " << descr.size() << " Transform: " << components.size() << std::endl;
throw pl::DimensionalityReductionError(s.str(), currentMethod, currentLine);
}
// make data zero-mean
cv::Mat1f X(descr.rows,descr.cols);
#ifdef USE_TBB
tbb::parallel_for(int32_t(0), X.rows, [&](int32_t i) {
X.row(i) = descr.row(i) - mean;
}
);
#else
for( int32_t y = 0; y < X.rows; y++){
X.row(y) = descr.row(y) - mean;
}
#endif
// the actual transformation
X = X * components.t();
// whiten the data
if ( whiten ) {
cv::Mat1f var_reg = variance.clone();
var_reg = var_reg.reshape(1, 1);
if ( reg > 0.0 )
var_reg += reg;
cv::sqrt(var_reg, var_reg);
#ifdef USE_TBB
tbb::parallel_for(int32_t(0), X.rows, [&](int32_t i) {
cv::Mat1f row = X.row(i);
row /= var_reg;
}
);
#else
for( int32_t i = 0; i < X.rows; i++ ) {
cv::Mat1f row = X.row(i);
row /= var_reg;
}
#endif
}
out = X;
}