Shape modeling and categorization using fisher kernels

This paper describes an efficient approach for the problem of shape modeling and classification. It is shown that this problem can be approached within a hybrid generative discriminative framework that integrates both finite mixture models and support vectors machines (SVM). The proposed framework is based on the generation of Fisher SVM kernel from the multinomial Beta-Liouville finite mixture model (MBLM). The MBLM is introduced as an efficient and flexible approach to model shape contexts represented by count vectors. Through extensive experiments concerning the categorization of well-known challenging shape databases, we show the merits of the proposed learning technique.