//fix phase of eigensystem and store phase of first entry of each eigenvector
void fix_phase(Eigen::MatrixXcd& V, Eigen::MatrixXcd& V_fix, std::vector<double>& phase) {
const int V3 = pars -> get_int("V3");
//helper variables:
//Number of eigenvectors
int n_ev;
//negative imaginary
std::complex<double> i_neg (0,-1);
//tmp factor and phase
std::complex<double> fac (1.,1.);
double tmp_phase = 0;
//get sizes right, resize if necessary
n_ev = V.cols();
if (phase.size() != n_ev) phase.resize(n_ev);
if (V_fix.cols() != n_ev) V_fix.resize(3*V3,n_ev);
//loop over all eigenvectors of system
for (int n = 0; n < n_ev; ++n) {
tmp_phase = std::arg(V(0,n));
phase.at(n) = tmp_phase;
fac = std::exp(i_neg*tmp_phase);
//Fix phase of eigenvector with negative polar angle of first entry
V_fix.col(n) = fac * V.col(n);
}
}
void CCorrelationFilters::build_mmcf(struct CDataStruct *img, struct CParamStruct *params, struct CFilterStruct *filt)
{
/*
* This function calls the correlation filter design proposed in the following publications.
*
* A. Rodriguez, Vishnu Naresh Boddeti, B.V.K. Vijaya Kumar and A. Mahalanobis, "Maximum Margin Correlation Filter: A New Approach for Localization and Classification", IEEE Transactions on Image Processing, 2012.
*
* Vishnu Naresh Boddeti, "Advances in Correlation Filters: Vector Features, Structured Prediction and Shape Alignment" PhD thesis, Carnegie Mellon University, Pittsburgh, PA, USA, 2012.
*
* Vishnu Naresh Boddeti and B.V.K. Vijaya Kumar, "Maximum Margin Vector Correlation Filters," Arxiv 1404.6031 (April 2014).
*
* Notes: This currently the best performing Correlation Filter design, especially when the training sample size is larger than the dimensionality of the data.
*/
filt->params = *params;
filt->filter.size_data = params->size_filt_freq;
filt->filter.size_data_freq = params->size_filt_freq;
filt->filter.num_elements_freq = img->num_elements_freq;
params->num_elements_freq = img->num_elements_freq;
filt->filter.data_freq = new complex<double>[img->num_elements_freq*img->num_channels];
Eigen::ArrayXcd filt_freq = Eigen::ArrayXcd::Zero(params->num_elements_freq*img->num_channels);
// If not set default to 1
if (params->wpos < 1) params->wpos = 1;
filt->params.wpos = params->wpos;
compute_psd_matrix(img, params);
Eigen::MatrixXcd Y = Eigen::MatrixXcd::Zero(img->num_elements_freq*img->num_channels,img->num_data);
Eigen::MatrixXcd u = Eigen::MatrixXcd::Zero(img->num_data,1);
Eigen::MatrixXd temp = Eigen::MatrixXd::Zero(img->num_data,img->num_data);
Eigen::Map<Eigen::MatrixXcd> X(img->data_freq,img->num_elements_freq*img->num_channels,img->num_data);
Eigen::ArrayXXcd temp1 = Eigen::ArrayXXcd::Zero(img->num_elements_freq,img->num_channels);
Eigen::ArrayXXcd temp2 = Eigen::ArrayXXcd::Zero(img->num_elements_freq,img->num_channels);
Eigen::Vector2i num_blocks_1, num_blocks_2;
num_blocks_1 << img->num_channels,img->num_channels;
num_blocks_2 << img->num_channels,1;
for (int k=0;k<img->num_data;k++){
temp2 = X.block(0,k,img->num_elements_freq*img->num_channels,1).array();
temp2.resize(img->num_elements_freq,img->num_channels);
fusion_matrix_multiply(temp1, img->Sinv, temp2, num_blocks_1, num_blocks_2);
temp1.resize(img->num_elements_freq*img->num_channels,1);
Y.block(0,k,img->num_elements_freq*img->num_channels,1) = temp1.matrix();
temp1.resize(img->num_elements_freq,img->num_channels);
if (img->labels[k] == 1)
{
u(k) = std::complex<double>(params->wpos,0);
}
else
{
u(k) = std::complex<double>(-1,0);
}
}
esvm::SVMClassifier libsvm;
libsvm.setC(params->C);
libsvm.setKernel(params->kernel_type);
libsvm.setWpos(params->wpos);
temp = (X.conjugate().transpose()*Y).real();
Eigen::Map<Eigen::MatrixXd> y(img->labels,img->num_data,1);
libsvm.train(temp, y);
temp.resize(0,0);
int nSV;
libsvm.getNSV(&nSV);
Eigen::VectorXi sv_indices = Eigen::VectorXi::Zero(nSV);
Eigen::VectorXd sv_coef = Eigen::VectorXd::Zero(nSV);
libsvm.getSI(sv_indices);
libsvm.getCoeff(sv_coef);
for (int k=0; k<nSV; k++) {
filt_freq += (Y.block(0,sv_indices[k]-1,img->num_elements_freq*img->num_channels,1)*sv_coef[k]).array();
}
Y.resize(0,0);
Eigen::Map<Eigen::ArrayXcd>(filt->filter.data_freq,img->num_elements_freq*img->num_channels) = filt_freq;
filt->filter.num_data = 1;
filt->filter.num_channels = img->num_channels;
filt->filter.ptr_data.reserve(filt->filter.num_data);
filt->filter.ptr_data_freq.reserve(filt->filter.num_data);
ifft_data(&filt->filter);
fft_data(&filt->filter);
}
void CCorrelationFilters::build_otsdf(struct CDataStruct *img, struct CParamStruct *params, struct CFilterStruct *filt)
{
/*
* This function implements the correlation filter design proposed in the following publications.
*
* [1] Optimal trade-off synthetic discriminant function filters for arbitrary devices, B.V.K.Kumar, D.W.Carlson, A.Mahalanobis - Optics Letters, 1994.
*
* [2] Jason Thornton, "Matching deformed and occluded iris patterns: a probabilistic model based on discriminative cues." PhD thesis, Carnegie Mellon University, Pittsburgh, PA, USA, 2007.
*
* [3] Vishnu Naresh Boddeti, Jonathon M Smereka, and B. V. K. Vijaya Kumar, "A comparative evaluation of iris and ocular recognition methods on challenging ocular images." IJCB, 2011.
*
* [4] A. Mahalanobis, B.V.K. Kumar, D. Casasent, "Minimum average correlation energy filters," Applied Optics, 1987
*
* Notes: This filter design is good when the dimensionality of the data is greater than the training sample size. Setting the filter parameter params->alpha=0 results in the popular MACE filter. However, it is usually better to set alpha to a small number rather than setting it to 0. If you use MACE cite [4].
*/
filt->params = *params;
filt->filter.size_data = params->size_filt_freq;
filt->filter.size_data_freq = params->size_filt_freq;
filt->filter.num_elements_freq = img->num_elements_freq;
params->num_elements_freq = img->num_elements_freq;
filt->filter.data_freq = new complex<double>[img->num_elements_freq*img->num_channels];
Eigen::ArrayXcd filt_freq = Eigen::ArrayXcd::Zero(params->num_elements_freq*img->num_channels);
// If not set default to 1
if (params->wpos < 1) params->wpos = 1;
filt->params.wpos = params->wpos;
compute_psd_matrix(img, params);
Eigen::MatrixXcd Y = Eigen::MatrixXcd::Zero(img->num_elements_freq*img->num_channels,img->num_data);
Eigen::MatrixXcd u = Eigen::MatrixXcd::Zero(img->num_data,1);
Eigen::MatrixXcd temp = Eigen::MatrixXcd::Zero(img->num_data,img->num_data);
Eigen::MatrixXd tmp = Eigen::MatrixXd::Zero(img->num_data,img->num_data);
Eigen::Map<Eigen::MatrixXcd> X(img->data_freq,img->num_elements_freq*img->num_channels,img->num_data);
Eigen::ArrayXXcd temp1 = Eigen::ArrayXXcd::Zero(img->num_elements_freq,img->num_channels);
Eigen::ArrayXXcd temp2 = Eigen::ArrayXXcd::Zero(img->num_elements_freq,img->num_channels);
Eigen::Vector2i num_blocks_1, num_blocks_2;
num_blocks_1 << img->num_channels,img->num_channels;
num_blocks_2 << img->num_channels,1;
for (int k=0;k<img->num_data;k++){
temp2 = X.block(0,k,img->num_elements_freq*img->num_channels,1).array();
temp2.resize(img->num_elements_freq,img->num_channels);
fusion_matrix_multiply(temp1, img->Sinv, temp2, num_blocks_1, num_blocks_2);
temp1.resize(img->num_elements_freq*img->num_channels,1);
Y.block(0,k,img->num_elements_freq*img->num_channels,1) = temp1.matrix();
temp1.resize(img->num_elements_freq,img->num_channels);
if (img->labels[k] == 1)
{
u(k) = std::complex<double>(params->wpos,0);
}
else
{
u(k) = std::complex<double>(1,0);
}
}
temp = X.conjugate().transpose()*Y;
temp = temp.ldlt().solve(u);
filt_freq = Y*temp;
Y.resize(0,0);
Eigen::Map<Eigen::ArrayXcd>(filt->filter.data_freq,img->num_elements_freq*img->num_channels) = filt_freq;
filt->filter.num_data = 1;
filt->filter.num_channels = img->num_channels;
filt->filter.ptr_data.reserve(filt->filter.num_data);
filt->filter.ptr_data_freq.reserve(filt->filter.num_data);
ifft_data(&filt->filter);
fft_data(&filt->filter);
}