Empirical Study of Multi-scale Filter Banks for Object Categorization

The aim of this work is the evaluation of different multi-scale filter banks, mainly based on oriented Gaussian derivatives and Gabor functions, to be used in the generation of robust features for visual object categorization. In order to combine the responses obtained from several spatial scales, we use the biologically inspired HMAX model (Riesenhuber and Poggio, 1999). We have tested the different sets of features on the challenging Caltech-101 database, and we have performed the categorizarion procedure with AdaBoost, support vector machines and JointBoosting classifiers, achieving remarkable results