GrabCut::GrabCut( Image<Color>* image )
{
m_image = image;
m_w = m_image->width();
m_h = m_image->height();
m_trimap = new Image<TrimapValue>( m_w, m_h );
m_trimap->fill(TrimapUnknown);
m_GMMcomponent = new Image<unsigned int>( m_w, m_h );
m_hardSegmentation = new Image<SegmentationValue>( m_w, m_h );
m_hardSegmentation->fill(SegmentationBackground);
m_softSegmentation = 0; // Not yet implemented
m_TLinksImage = new Image<Color>(m_w, m_h);
m_TLinksImage->fill(Color(0,0,0));
m_NLinksImage = new Image<Real>(m_w, m_h);
m_NLinksImage->fill(0);
m_GMMImage = new Image<Color>(m_w, m_h);
m_GMMImage->fill(Color(0,0,0));
m_AlphaImage = new Image<Real>(m_w, m_h);
m_AlphaImage->fill(0);
m_foregroundGMM = new GMM(5);
m_backgroundGMM = new GMM(5);
//set some constants
m_lambda = 50;
computeL();
computeBeta();
m_NLinks = new Image<NLinks>( m_w, m_h );
computeNLinks();
m_graph = 0;
m_nodes = new Image<Graph::node_id>( m_w, m_h );
}
void computeFSPAI(MatrixType const & A,
MatrixType const & PatternA,
MatrixType & L,
MatrixType & L_trans,
fspai_tag)
{
typedef typename MatrixType::value_type ScalarType;
typedef boost::numeric::ublas::matrix<ScalarType> DenseMatrixType;
typedef std::vector<std::map<unsigned int, ScalarType> > SparseMatrixType;
//
// preprocessing: Store A in a STL container:
//
//std::cout << "Transferring to STL container:" << std::endl;
std::vector<std::vector<ScalarType> > y_k(A.size1());
SparseMatrixType STL_A(A.size1());
sym_sparse_matrix_to_stl(A, STL_A);
//
// Step 1: Generate pattern indices
//
//std::cout << "computeFSPAI(): Generating pattern..." << std::endl;
std::vector<std::vector<vcl_size_t> > J(A.size1());
generateJ(PatternA, J);
//
// Step 2: Set up matrix blocks
//
//std::cout << "computeFSPAI(): Setting up matrix blocks..." << std::endl;
std::vector<DenseMatrixType> subblocks_A(A.size1());
fill_blocks(STL_A, subblocks_A, J, y_k);
STL_A.clear(); //not needed anymore
//
// Step 3: Cholesky-factor blocks
//
//std::cout << "computeFSPAI(): Cholesky-factorization..." << std::endl;
for (vcl_size_t i=0; i<subblocks_A.size(); ++i)
{
//std::cout << "Block before: " << subblocks_A[i] << std::endl;
cholesky_decompose(subblocks_A[i]);
//std::cout << "Block after: " << subblocks_A[i] << std::endl;
}
/*vcl_size_t num_bytes = 0;
for (vcl_size_t i=0; i<subblocks_A.size(); ++i)
num_bytes += 8*subblocks_A[i].size1()*subblocks_A[i].size2();*/
//std::cout << "Memory for FSPAI matrix: " << num_bytes / (1024.0 * 1024.0) << " MB" << std::endl;
//
// Step 4: Solve for y_k
//
//std::cout << "computeFSPAI(): Cholesky-solve..." << std::endl;
for (vcl_size_t i=0; i<y_k.size(); ++i)
{
if (subblocks_A[i].size1() > 0) //block might be empty...
{
//y_k[i].resize(subblocks_A[i].size1());
//std::cout << "y_k[" << i << "]: ";
//for (vcl_size_t j=0; j<y_k[i].size(); ++j)
// std::cout << y_k[i][j] << " ";
//std::cout << std::endl;
cholesky_solve(subblocks_A[i], y_k[i]);
}
}
//
// Step 5: Set up Cholesky factors L and L_trans
//
//std::cout << "computeFSPAI(): Computing L..." << std::endl;
L.resize(A.size1(), A.size2(), false);
L.reserve(A.nnz(), false);
L_trans.resize(A.size1(), A.size2(), false);
L_trans.reserve(A.nnz(), false);
computeL(A, L, L_trans, y_k, J);
//std::cout << "L: " << L << std::endl;
}
int main(int argc, char**) {
{
long long t1=0;
float s1=0;
long long t2=0;
float s2=0;
long long t3=0;
float s3=0;
fillR();
computeV();
for (int i=0; i!=10000; ++i) {
fillO();
t1 -= rdtsc();
computeV();
t1 += rdtsc();
s1+=sum();
t2 -= rdtsc();
computeA();
t2 += rdtsc();
s2+=sum();
t3 -= rdtsc();
computeB();
t3 += rdtsc();
s3+=sum();
}
std::cout << "native vector " << s1 << " " << double(t1)/10000 << std::endl;
std::cout << "vector by elements " << s2 << " " << double(t2)/10000 << std::endl;
std::cout << "scalar " << s3 << " " << double(t3)/10000 << std::endl << std::endl;
}
{
long long t1=0;
float s1=0;
long long t2=0;
float s2=0;
long long t3=0;
float s3=0;
fillO();
computeV();
for (int i=0; i!=10000; ++i) {
fillR();
t1 -= rdtsc();
computeV();
t1 += rdtsc();
s1+=sum();
t2 -= rdtsc();
computeA();
t2 += rdtsc();
s2+=sum();
memcpy(a,va,sizeof(float)*1024*vfl);
t3 -= rdtsc();
computeL();
t3 += rdtsc();
memcpy(vb,b,sizeof(float)*1024*vfl);
s3+=sum();
}
std::cout << "native vector " << s1 << " " << double(t1)/10000 << std::endl;
std::cout << "vector by elements " << s2 << " " << double(t2)/10000 << std::endl;
std::cout << "scalar " << s3 << " " << double(t3)/10000 << std::endl << std::endl;
}
{
long long t1=0;
float s1=0;
long long t2=0;
float s2=0;
long long t3=0;
float s3=0;
fillO();
computeV();
for (int i=0; i!=10000; ++i) {
fillW();
t1 -= rdtsc();
computeV();
t1 += rdtsc();
s1+=sum();
t2 -= rdtsc();
computeA();
t2 += rdtsc();
s2+=sum();
t3 -= rdtsc();
computeB();
t3 += rdtsc();
s3+=sum();
}
std::cout << "native vector " << s1 << " " << double(t1)/10000 << std::endl;
std::cout << "vector by elements " << s2 << " " << double(t2)/10000 << std::endl;
std::cout << "scalar " << s3 << " " << double(t3)/10000 << std::endl << std::endl;
}
return 0;
}
