Classifying materials from images: to cluster or not to cluster?

Abstract

The objective of this paper is texture classification from a single image acquired under unknown viewpoint and illumination. The method is based on the statistical distribution of filter responses in a low dimensional space. We compare two methods of representing the filter response distribution. The first is to store the entire response as a PDF, the second is to learn representative texture elements (\emph{textons}) by clustering, and represent the texture by their distribution. This effectively compares the PDF approach of Konishi and Yuille (CVPR 2000), to the texton approach of Cula and Dana (CVPR 2001) and Varma and Zisserman (ECCV 2002). The methods are assessed by classifying the material images present in the Columbia-Utrecht database, and very accurate results are obtained. We also describe preliminary results on estimating orientation co-occurrence statistics. These statistics may be used to augment the texton representation and improve classification performance.