Hello there,
I am developing a coin recognizer for iOS and I'm using opencv for feature extraction and SVM training. So far, I managed to train my classifier on my computer (OSX) using opencv 3.0, SIFT extraction and SVM. Resubstitution error rates are perfect (i.e. 100% accuracy).
Moving to the iOS end, I am loading both my BoVW descriptors and SVM model from the bundle (xml files) and using them for recognition. Nevertheless, I am getting 0% accuracy.
Things already verified: - SVM, when loaded on iOS, possesses the same parameters obtained by the OSX training step (grid search went well I believe). - BoVW descriptors look good too.
I'm thinking about a floating point error or something like that?
Regards,
Jean