| Vector::VectorTem<Real64> Impala::Core::Feature::MarkovStationaryFeature | ( | Matrix::Mat * | cooccuranceMatrix | ) | [inline] |
Definition at line 39 of file MarkovStationaryFeature.h.
References Impala::Core::Array::Add(), Impala::Core::Matrix::GetDiagonal(), Impala::Core::Matrix::MatCopy(), Impala::Core::Matrix::MatE(), Impala::Core::Matrix::MatMul(), Impala::Core::Matrix::MatNrCol(), Impala::Core::Matrix::MatNrRow(), Impala::Core::Matrix::MatSumAxis0(), Impala::Core::Matrix::MatSumAxis1(), Impala::Core::Array::MulVal(), Impala::Core::Array::SetVal(), and Impala::Core::Vector::Sum().
Referenced by Impala::Core::VideoSet::LbpEval::HandleNewFrame(), MarkovStationaryFeatureTimed(), and testMSF().
00040 { 00041 using namespace Matrix; 00042 const int ITERATION_COUNT = 100; 00043 00044 int n = MatNrRow(cooccuranceMatrix); 00045 if(MatNrCol(cooccuranceMatrix) != n) 00046 { 00047 // it cannot be! 00048 std::cerr << "cooccuranceMatrix must be square" << std::endl; 00049 throw "bye"; 00050 } 00051 00052 /* descriptor storage */ 00053 Vector::VectorTem<Real64> descriptor(2 * n); 00054 00055 /* normalize the co-ocurrance matrix */ 00056 Mat* p = MatCreate<Mat>(n, n); 00057 Mat* normalizationFactors = MatSumAxis1(cooccuranceMatrix); 00058 for(int vj = 0; vj < n; vj++) 00059 { 00060 if(*MatE(normalizationFactors, vj, 0) == 0.0) 00061 { 00062 *MatE(normalizationFactors, vj, 0) = 1.0; 00063 } 00064 for(int i = 0; i < n; i++) 00065 { 00066 *MatE(p, vj, i) = *MatE(cooccuranceMatrix, vj, i) / *MatE(normalizationFactors, vj, 0); 00067 } 00068 } 00069 //Dump(cooccuranceMatrix); 00070 //Dump(normalizationFactors); 00071 //Dump(p); 00072 delete normalizationFactors; 00073 00074 /* we now have a transition matrix in p */ 00075 Mat* identity = MatCreate<Mat>(n, n); 00076 SetVal(identity, 0.0); 00077 // set to identity matrix 00078 for(int i = 0; i < n; i++) 00079 { 00080 *MatE(identity, i, i) = 1.0; 00081 } 00082 Mat* a = MatCopy(identity); 00083 Mat* previousPower = identity; 00084 // identity will be deleted in the loop as previousPower 00085 for(int k = 0; k < ITERATION_COUNT; k++) 00086 { 00087 Mat* currentPower = MatMul(previousPower, p); 00088 // perform matrix addition 00089 Add(a, a, currentPower); 00090 delete previousPower; 00091 previousPower = currentPower; 00092 } 00093 //std::cout << "previous:" << std::endl; Dump(previousPower); 00094 delete previousPower; 00095 MulVal(a, a, 1.0 / (ITERATION_COUNT + 1)); 00096 //std::cout << "a_n:" << std::endl; Dump(a); 00097 00098 /* construct the stationary disribution by taking the averages of the 00099 rows of a (all rows of a should be equal upon convergence) */ 00100 Mat* stationarySum = MatSumAxis0(a); 00101 //Dump(stationarySum); 00102 for(int i = 0; i < n; i++) 00103 { 00104 descriptor[n + i] = *MatE(stationarySum, 0, i) / n; 00105 } 00106 delete a; 00107 delete stationarySum; 00108 00109 /* construct initial distribution by normalizing the self-transition */ 00110 Vector::VectorTem<Real64> diagonal = GetDiagonal(cooccuranceMatrix); 00111 Real64 normalizationFactor = Sum(diagonal); 00112 if(normalizationFactor == 0.0) normalizationFactor = 1.0; 00113 Vector::VectorTem<Real64> initialDistribution = diagonal * (1.0 / normalizationFactor); 00114 00115 /* combine initial distribution and stationary distribution into a single 00116 vector */ 00117 for(int i = 0; i < n; i++) 00118 { 00119 descriptor[i] = initialDistribution[i]; 00120 } 00121 delete p; 00122 00123 return descriptor; 00124 }
Here is the call graph for this function:
