Adaptive Gaussian kernel learning for sparse Bayesian classification: An approach for silhouette based vehicle classification

Kernel based approaches are one of the most well-known methods in regression and classification tasks. Type of kernel function and also its parameters have a considerable effect on the classifier performance. Usually kernel parameters are obtained by cross-validation or validation dataset. In this paper we propose a classification learning approach which learn the parameter (kernel width) of Gaussian kernel function during learning stage. The proposed method is an extension of RVM which is a Bayesian counter-part of well-known SVM classifier. The evaluation results on both synthetic and real datasets show better performance and also model sparsity compared to competing algorithms. Particularly the proposed algorithm outperforms other existing methods on vehicle classification based on their silhouettes.