|| || || || || ||

---

DoConfusionMatrix

void Impala::Core::Table::ConfusionMatrix::DoConfusionMatrix (  )  [inline] Definition at line 120 of file ConfusionMatrix.h. References conceptSimTable, Impala::Core::Table::Bayes::DumpMatrix(), Impala::Core::Table::TableTem< Col1T, Col2T, Col3T, Col4T, Col5T, Col6T, Col7T, Col8T, Col9T >::Get1(), Impala::Core::Table::SimilarityTableSet::GetName(), Impala::Core::Table::SimilarityTableSet::GetSimTable(), Impala::Core::Table::AnnotationTableSet::GetTable(), mAnnoTabSet, Impala::Core::Table::Bayes::mConceptNum, Impala::Core::Table::Bayes::mConceptSet, mConceptSimSet, mConfusionMatrix, mFeatureName, mQuidTable, Impala::Core::Table::Bayes::mSetName, Impala::Core::Table::Bayes::mTableSize, SaveConfigurationMatrix(), Impala::Core::Table::TableTem< Col1T, Col2T, Col3T, Col4T, Col5T, Col6T, Col7T, Col8T, Col9T >::Set1(), and Impala::Core::Table::Bayes::tabConcept. { std::cout << std::endl; std::cout << "DumpConfusionMatrix() - my own tool" << std::endl; std::cout << std::endl; // load annotation and similarity tables for (int c=0 ; c< mConceptNum ; c++) { tabConcept[c] = mAnnoTabSet->GetTable(c); std::cout << "table " << c << ", name = " << mConceptSimSet->GetName(c) << std::endl; conceptSimTable[c] = mConceptSimSet->GetSimTable(c); } // get the mean of each concept list Real64* pMean = new Real64[mConceptNum]; memset(pMean,0,sizeof(Real64)*mConceptNum); for (int c=0 ; c<mConceptNum ; c++) { Real64 fSum = 0; for (int i=0 ; i<mTableSize ; i++) { Quid quid = mQuidTable->Get1(i); Real64 sim = conceptSimTable[c]->Get1(i); fSum += sim; } pMean[c] = fSum/mTableSize; } // get the variance of each concept list Real64* pStd = new Real64[mConceptNum]; memset(pStd,0,sizeof(Real64)*mConceptNum); for (int c=0 ; c<mConceptNum ; c++) { Real64 fSum = 0; for (int i=0 ; i<mTableSize ; i++) { Quid quid = mQuidTable->Get1(i); Real64 sim = conceptSimTable[c]->Get1(i); fSum += pow(sim - pMean[c],2); } pStd[c] = sqrt(fSum/(mTableSize-1)); } // dump the old data distribution std::cout << std::endl; std::cout << "Before gaussian normalization:" << std::endl; for (int c=0; c<mConceptNum; c++ ) { printf(" (mean,std) = (%.10f, %.10f)\n", pMean[c], pStd[c]); } // do the gaussian normalization of the similarity table for (int c=0 ; c<mConceptNum ; c++) { Real64 fSum = 0; for (int i=0 ; i<mTableSize ; i++) { Quid quid = mQuidTable->Get1(i); Real64 old_score = conceptSimTable[c]->Get1(i); Real64 new_score = (old_score - pMean[c])/pStd[c]; // set the normalized value as the new score conceptSimTable[c]->Set1(i,new_score); } } // check the updated data distribution memset(pMean,0,sizeof(Real64)*mConceptNum); memset(pStd,0,sizeof(Real64)*mConceptNum); for (int c=0 ; c<mConceptNum ; c++) { Real64 fSum = 0; for (int i=0 ; i<mTableSize ; i++) { Quid quid = mQuidTable->Get1(i); Real64 sim = conceptSimTable[c]->Get1(i); fSum += sim; } pMean[c] = fSum/mTableSize; } // get the variance of each concept list for (int c=0 ; c<mConceptNum ; c++) { Real64 fSum = 0; for (int i=0 ; i<mTableSize ; i++) { Quid quid = mQuidTable->Get1(i); Real64 sim = conceptSimTable[c]->Get1(i); fSum += pow(sim - pMean[c],2); } pStd[c] = sqrt(fSum/(mTableSize-1)); } // dump the normalized data distribution std::cout << "After gaussian normalization:" << std::endl; for (int c=0; c<mConceptNum; c++ ) { printf(" (mean,std) = (%.10f, %.10f)\n", pMean[c], pStd[c]); } delete []pMean; delete []pStd; // loop for all the key-frames in the table int nDupLabels = 0; for (int i=0 ; i<mTableSize ; i++) { Quid quid = mQuidTable->Get1(i); // for keyframe i: find the highest score among all concepts Real64 fMaxValue = -1; int nDetect = -1; for (int c=0 ; c<mConceptNum ; c++) { //String concept = mConceptSimSet->GetName(c); //std::cout << concept << ", "; Real64 sim = conceptSimTable[c]->Get1(i); if (sim>fMaxValue) { fMaxValue = sim; nDetect = c; } } // for keyframe i: find the ground truth that concept keyframe i belongs to int nTruth = -1; int count = 0; for (int c=0 ; c<mConceptNum ; c++) { //String concept = mConceptSimSet->GetName(c); //std::cout << concept << ", "; if ( tabConcept[c]->IsPositive(i) ) { nTruth = c; count ++; } } // be sure that only one unique label if (count == 0) // nTruth == -1 { // no ground truth for current key-frame continue; } else if (count > 1) { //std::cout << "Warning: there are " << count << " labels for shot " << i << std::endl; nDupLabels++; } int nCorrect = 0; int nError = 0; // mConfusionMatrix[i=nTruth][j=nDetect]: // 1) A row (fixed i) stores all the predicted results for concept i // 2) A column (fixed j) stores all the ground truth of concept j // 3) the sum of a column j should be the total number of positive samples for concept j (before normalization) if (nDetect == nTruth) { mConfusionMatrix[nTruth][nTruth] += 1.0; nCorrect ++; } else { // re-corrected order for row/column mConfusionMatrix[nTruth][nDetect] += 1.0; nError ++; } } std::cout << "Duplicated Labels: " << nDupLabels << std::endl; DumpMatrix("Confusion Matrix", mConfusionMatrix, mConceptNum, mConceptNum, false); String outfileInt = "./ConfusionMatrix/" + mSetName + "_" + mConceptSet + "_" + mFeatureName + "_int.txt"; SaveConfigurationMatrix(outfileInt.c_str(), false); // normalizations: sum of a column y (fixed x) is 1 // mConfusionMatrix[x=nTruth][y=nDetect]: for (int x=0 ; x<mConceptNum; x++) { Real64 sum = 0; for (int y=0 ; y<mConceptNum ; y++) sum += mConfusionMatrix[x][y]; if (sum < 0) continue; for (int y=0 ; y<mConceptNum ; y++) mConfusionMatrix[x][y] /= sum; } printf("The sum of each column (actual class)\n"); for (int x=0 ; x<mConceptNum; x++) { Real64 sum = 0; for (int y=0 ; y<mConceptNum ; y++) sum += mConfusionMatrix[x][y]; printf("%.5f ", sum); } printf("\n"); DumpMatrix("Confusion Matrix", mConfusionMatrix, mConceptNum, mConceptNum); String outfile = "./ConfusionMatrix/" + mSetName + "_" + mConceptSet + "_" + mFeatureName + "_3f.txt"; SaveConfigurationMatrix(outfile.c_str()); int x=0; } Here is the call graph for this function:

---