Stacking of texture based filters for visual place categorization

Recent research in computer vision has shown that combining multiple features is an effective way to improve classification performance. Furthermore, the use of filters that convolve images at multiple filter responses can increase descriptions to images. The more distinctive the filter responses, the better it able to distinguish characteristics from other groups. Thus, this paper describes a combination method that combines multiple classifier outputs at several filter responses to enhance the automatic visual place categorization system. Besides, one of the goals of this study is to explore performance differences between single and dedicated combination of filter response classifier methods. One possible problem of combining multiple filter responses for describing images is that the input vector becomes very large in dimensionality, which can increase the problem of overfitting and hinder generalization performance. Therefore, the stacking of support vector machine is used to compute the right output class from each single descriptor of filter responses that has been trained at the first layer of support vector machine. Next, the second layer support vector machine is used to combine those class probability output values of all trained first layer support vector models to learn the right output class. We have performed experiments on five different categories of visual places from the KTH-IDOL2 dataset with a single descriptor using 25 different filter responses of Laws filters. Results showed that the 2-layer stacking algorithm outperform the single and naive approaches that uses single filter response input vector and combines all filter response outputs directly in a very large single input vector respectively.