MatrixXf featureGradient( const MatrixXf & a, const MatrixXf & b ) const {
if (ntype_ == NO_NORMALIZATION )
return kernelGradient( a, b );
else if (ntype_ == NORMALIZE_SYMMETRIC ) {
MatrixXf fa = lattice_.compute( a*norm_.asDiagonal(), true );
MatrixXf fb = lattice_.compute( b*norm_.asDiagonal() );
MatrixXf ones = MatrixXf::Ones( a.rows(), a.cols() );
VectorXf norm3 = norm_.array()*norm_.array()*norm_.array();
MatrixXf r = kernelGradient( 0.5*( a.array()*fb.array() + fa.array()*b.array() ).matrix()*norm3.asDiagonal(), ones );
return - r + kernelGradient( a*norm_.asDiagonal(), b*norm_.asDiagonal() );
}
else if (ntype_ == NORMALIZE_AFTER ) {
MatrixXf fb = lattice_.compute( b );
MatrixXf ones = MatrixXf::Ones( a.rows(), a.cols() );
VectorXf norm2 = norm_.array()*norm_.array();
MatrixXf r = kernelGradient( ( a.array()*fb.array() ).matrix()*norm2.asDiagonal(), ones );
return - r + kernelGradient( a*norm_.asDiagonal(), b );
}
else /*if (ntype_ == NORMALIZE_BEFORE )*/ {
MatrixXf fa = lattice_.compute( a, true );
MatrixXf ones = MatrixXf::Ones( a.rows(), a.cols() );
VectorXf norm2 = norm_.array()*norm_.array();
MatrixXf r = kernelGradient( ( fa.array()*b.array() ).matrix()*norm2.asDiagonal(), ones );
return -r+kernelGradient( a, b*norm_.asDiagonal() );
}
}
BPPottsPotential(const float* features1, const float* features2, int D, int N1, int N2, float w, bool per_pixel_normalization=true) :N1_(N1), N2_(N2), w_(w) {
float * features = new float[ (N1_+N2_)*D ];
memset( features, 0, (N1_+N2_)*D*sizeof(float) );
memcpy( features , features1, N1_*D*sizeof(float) );
memcpy( features+N1_*D, features2, N2_*D*sizeof(float) );
lattice_.init( features, D, N1_+N2_ );
delete [] features;
norm_ = allocate( N2_ );
float * tmp = allocate( N1_ );
for( int i=0; i<N1_; i++ )
tmp[i] = 1;
// Compute the normalization factor
lattice_.compute( norm_, tmp, 1, 0, N1_, N1_, N2_ );
if( per_pixel_normalization ){
// use a per pixel normalization
for( int i=0; i<N2_; i++ )
norm_[i] = 1.f / (norm_[i]+1e-20f);
}
else{
float mean_norm = 0;
for( int i=0; i<N2_; i++ )
mean_norm += norm_[i];
mean_norm = N2_ / mean_norm;
// use a per pixel normalization
for( int i=0; i<N2_; i++ )
norm_[i] = mean_norm;
}
deallocate( tmp );
}
void filter( MatrixXf & out, const MatrixXf & in, bool transpose ) const {
// Read in the values
if( ntype_ == NORMALIZE_SYMMETRIC || (ntype_ == NORMALIZE_BEFORE && !transpose) || (ntype_ == NORMALIZE_AFTER && transpose))
out = in*norm_.asDiagonal();
else
out = in;
// Filter
if( transpose )
lattice_.compute( out, out, true );
else
lattice_.compute( out, out );
// lattice_.compute( out.data(), out.data(), out.rows() );
// Normalize again
if( ntype_ == NORMALIZE_SYMMETRIC || (ntype_ == NORMALIZE_BEFORE && transpose) || (ntype_ == NORMALIZE_AFTER && !transpose))
out = out*norm_.asDiagonal();
}
PottsPotential(const float* features, int D, int N, float w, bool per_pixel_normalization=true) :N_(N), w_(w) {
lattice_.init( features, D, N );
norm_ = allocate( N );
for ( int i=0; i<N; i++ )
norm_[i] = 1;
// Compute the normalization factor
lattice_.compute( norm_, norm_, 1 );
if ( per_pixel_normalization ) {
// use a per pixel normalization
for ( int i=0; i<N; i++ )
norm_[i] = 1.f / (norm_[i]+1e-20f);
}
else {
float mean_norm = 0;
for ( int i=0; i<N; i++ )
mean_norm += norm_[i];
mean_norm = N / mean_norm;
// use a per pixel normalization
for ( int i=0; i<N; i++ )
norm_[i] = mean_norm;
}
}
void initLattice( const MatrixXf & f ) {
const int N = f.cols();
lattice_.init( f );
norm_ = lattice_.compute( VectorXf::Ones( N ).transpose() ).transpose();
if ( ntype_ == NO_NORMALIZATION ) {
float mean_norm = 0;
for ( int i=0; i<N; i++ )
mean_norm += norm_[i];
mean_norm = N / mean_norm;
for ( int i=0; i<N; i++ )
norm_[i] = mean_norm;
}
else if ( ntype_ == NORMALIZE_SYMMETRIC ) {
for ( int i=0; i<N; i++ )
norm_[i] = 1.0 / sqrt(norm_[i]+1e-20);
}
else {
for ( int i=0; i<N; i++ )
norm_[i] = 1.0 / (norm_[i]+1e-20);
}
}
void apply(float* out_values, const float* in_values, float* tmp, int value_size) const {
lattice_.compute( tmp, in_values, value_size );
for ( int i=0,k=0; i<N_; i++ )
for ( int j=0; j<value_size; j++, k++ )
out_values[k] += w_*norm_[i]*tmp[k];
}