Gautam Muralidhar

Useful links for running SVM-based experiments using kernels generated by multi-resolution pyramid match –

1. Download libpmk from http://people.csail.mit.edu/jjl/libpmk/#users. This module includes a version of libsvm as well.

2. Under the extensions, download libpmk_features.tar.gz. This provides access to whole bunch of interest point detectors and feature extractors that can be used along with libpmk. To compile this successfully, you will need to first install libpmk (see unpacking and compiling code under - http://people.csail.mit.edu/jjl/libpmk/documentation/) and also install ImageMagick++ (http://www.imagemagick.org/www/Magick++/)

3. Note that libpmk has been tested only on Linux and there is no guarantee it will work on other operating systems including MAC OSX.

4. ETH-80 example code and data on the libpmk website gives a very good feel of how to create SVM-based experiments using libpmk and how to run them.

5. In case you have trouble compiling the feature extraction package (point 3), you can download the interest point detectors and descriptors from - http://www.robots.ox.ac.uk/~vgg/research/affine/index.html. You can then use Matlab to create the point set files that libpmk requires to build the multi-resolution histogram and generate the pyramid match kernel. For creating the point-set file refer to the point-set file class description on - http://people.csail.mit.edu/jjl/libpmk/documentation/reference-2.0/annotated.html