Index of resulting single detector vector probably being incorrect

[jpjodoin]

As found here: https://github.com/DaHoC/trainHOG/issues/2

"The first value (Index 0) of the vector being generated in function getSingleDetectingVector() being always 0 due to the way SVMlight handles indices, as well as the last value (Index 3780) being overwritten by -b. An indexshift does not fix this issue because the positions/indices of the detecting vector components do matter for HOG."

I think I've found the solution to solve this issue if you want to try. You have to shift the indice by -1, remove the last line with the -model->b, and rescale the output to 3780.

We then use -model->b as the detection threshold.

Example: void getSingleDetectingVector(std::vector& singleDetectorVector, std::vector& singleDetectorVectorIndices) { // Now we use the trained svm to retrieve the single detector vector DOC* supveclist = model->supvec; printf("Calculating single descriptor vector out of support vectors (may take some time)\n"); // Retrieve single detecting vector (v1 | b) from returned ones by calculating vec1 = sum_1_n (alpha_yx_i). (vec1 is a n x1 column vector. n = feature vector length ) singleDetectorVector.clear(); singleDetectorVector.resize(model->totwords, 0.); printf("Resulting vector size %lu\n", singleDetectorVector.size());

        // Walk over every support vector
        for (long ssv = 1; ssv < model->sv_num; ++ssv) { // Don't know what's inside model->supvec[0] ?!
            // Get a single support vector
            DOC* singleSupportVector = supveclist[ssv]; // Get next support vector
            SVECTOR* singleSupportVectorValues = singleSupportVector->fvec;
            WORD singleSupportVectorComponent;
            // Walk through components of the support vector and populate our detector vector
            for (long singleFeature = 0; singleFeature < model->totwords; ++singleFeature) {
                singleSupportVectorComponent = singleSupportVectorValues->words[singleFeature];
                singleDetectorVector.at(singleSupportVectorComponent.wnum-1) += (singleSupportVectorComponent.weight * model->alpha[ssv]);
            }
        }
    }

float getThreshold() const { return model->b; }

And when you want to detect, you can use something like this: hogdetector.detect(im, ptList, svm.getThreshold(), cv::Size(8, 8),cv::Size(0, 0));

[tengwang]

How you test the correctness of the classifier with your specific modification?

[jpjodoin]

To test it, I have used my training data as my test data in order to make sure I was getting the correct result for the known case of the classifier. With the previous implementation, I was getting poor results even when using my training data as test data.