Build a custom svm kernel matrix with opencv

Question

I have to train a Support Vector Machine model and I\'d like to use a custom kernel matrix, instead of the preset ones (like RBF, Poly, ecc.). How can I do that (if is it po

Answer 1

If you're not married to OpenCV for the SVM stuff, have a look at the shogun toolbox ... lots of SVM voodoo in there.

Answer 2

AFAICT, custom kernels for SVM aren't supported directly in OpenCV. It looks like LIBSVM, which is the underlying library that OpenCV uses for this, doesn't provide a particularly easy means of defining custom kernels. So, many of the wrappers that use LIBSVM don't provide this either. There seem to be a few, e.g. scikit for python: scikit example of SVM with custom kernel

You could also take a look at a completely different library, like SVMlight. It supports custom kernels directly. Also take a look at this SO question. The answers there include a handful of SVM libraries, along with brief reviews.

If you have compelling reasons to stay within OpenCV, you might be able to accomplish it by using kernel type CvSVM::LINEAR and applying your custom kernel to the data before training the SVM. I'm a little fuzzy on whether this direction would be fruitful, so I hope someone with more experience with SVM can chime in and comment. If it is possible to use a "precomputed kernel" by choosing "linear" as your kernel, then take a look at this answer for more ideas on how to proceed.

You might also consider including LIBSVM and calling it directly, without using OpenCV. See FAQ #418 for LIBSVM, which briefly touches on how to do custom kernels:

Q: I would like to use my own kernel. Any example? In svm.cpp, there are two subroutines for kernel evaluations: k_function() and kernel_function(). Which one should I modify ?

An example is "LIBSVM for string data" in LIBSVM Tools.

The reason why we have two functions is as follows. For the RBF kernel exp(-g |xi - xj|^2), if we calculate xi - xj first and then the norm square, there are 3n operations. Thus we consider exp(-g (|xi|^2 - 2dot(xi,xj) +|xj|^2)) and by calculating all |xi|^2 in the beginning, the number of operations is reduced to 2n. This is for the training. For prediction we cannot do this so a regular subroutine using that 3n operations is needed. The easiest way to have your own kernel is to put the same code in these two subroutines by replacing any kernel.

That last option sounds like a bit of a pain, though. I'd recommend scikit or SVMlight. Best of luck to you!