Classification methods used in OpenCV

edit

The documentation notes, that the implementation is based on the paper:

• Ahonen, T., Hadid, A., and Pietikainen, M. "Face Recognition with Local Binary Patterns". Computer Vision - ECCV 2004 (2004), 469–481.

In this paper a 1-Nearest Neighbor classifier with a Chi-Square Distance is used and so does the implementation. You can easily implement a different classifier, by for example wrapping the LBPHFaceRecognizer to get the histograms out. If you want to implement it and not rely on the given classes, you'll find the implementation in:

• facerec.cpp#L559

Do I think using a SVM is going to give you much better recognition rates? I don't think so and none of my experiments showed significantly improved recognition rates by simply employing a SVM. I have used all the kernel functions libsvm has, did the normalization for the RBF Kernel of course and optimized the parameters with a Grid Search. I didn't employ something like a Chi-Square Kernel, which might work better for histograms.

I wrote a Python framework once, where you could easily combine feature extraction methods with different classifiers (yeah, yeah basically I rewrote everything scikit-learn already has). It also has a Python implementation of Local Binary Pattern Histograms:

• https://github.com/bytefish/facerec

This has always been a fast way for me to validate the performance of feature extraction and classifier combinations. If I would do it now, I would use scikit-learn for almost all of the stuff. And this is what I recommend, simply because with Python it is much faster to prototype. Once you are confident the algorithm works as expected, then I would go ahead and implement it with OpenCV.