Efficient Classification for Additive Kernel SVMs

(Updated June 17, 2012)

Below is an implementation of the fast additive kernel SVM classifier described in the papers:

Efficient Classification for Additive Kernel SVMs
Subhransu Maji and Alexander C. Berg and Jitendra Malik.
IEEE PAMI (to appear) pdf preprint

Classification Using Intersection Kernel SVMs is efficient
Subhransu Maji and Alexander C. Berg and Jitendra Malik.
In Proceedings, CVPR 2008, Anchorage, Alaska. pdf

The source code is available as a tar.gz

  • (June 17, 2012) fast-additive-svms.tar.gz | README
    • (Updated June 17, 2012) Added a bug fix for sparse features.
    • Better support for all platforms. make.m compiles all the binaries from MATLAB itself.
    • Note: Only dense data is supported for now.
    • Includes libsvm-mat-3.0-1 : latest version of LIBSVM at the time of release.
    • Support for learning models using the intersection, chi-squared, and JS kernel.
    • Fast approximate classification for these kernels using piecewise linear approximations. ONLY piecewise linear approximations are supported. The binary search based exact classification for intersection kernel is no longer supported.
    • C/C++ based mex code for picewise linear interpolation, weighted kernel sampling and new fast binary search based weighted intersection kernel sampling.
  • (July 10, 2009) libsvm-mat-2.8.1-fast.v3.tar.gz
    • Support for float/double features.
    • Precomputation and prediction are now separate modules. Lightweight approximate models can be precomputed and stored without the need to store all the support vectors.
    • Approximate(or exact) predictions can be done directly using the approximate(or exact) precomputed models.
    • Check out the README/READMEFAST for documentation.
  • (May 19, 2008) libsvm-mat-2.8.1-fast.tar.gz