Tuesday, March 28, 2006

SVM Library in Matlab

For the past few weeks I’ve been using SVM as part of my research project. Until recently I used the SVM functions available in Piotr Dollar’s Image and Video Toolbox for Matlab. I am, however, constantly looking for alternatives. From what I saw on the Internet, it seems that one quite popular SVM implementation is SVM-Light by Thorsten Joachims. Add to it Tom Brigg’s MATLAB/MEX Interface to SVM-Light and you can run it from Matlab. The only thing you need is to put the Patch.exe utility in a folder that is in your system path and follow the installation instructions.

I’m still new with SVM-Light, so I don’t know exactly what all the parameters mean and how to tune them. A-priori it seems to work slightly better in terms of generalization accuracy and also run a little faster (with mostly default parameters and the same kernel function and parameters as Piotr Dollar’s implementation).

If you know of some other interesting SVM implementation, especially one that is available in Matlab, please drop me a line.

No comments: