This paper presents a statistical approach to 3D texture classification from a single image obtained under unknown viewpoint and illumination. Unlike in prior work, in which texture primitives (textons) are defined in a filter-response space, and texture classes modeled by frequency histograms of these textons, we seek to extract and model geometric and photometric properties of image regions defining the texture. To this end, texture images are first segmented by a multiscale segmentation algorithm, and a universal set of texture primitives is specified over all texture classes in the domain of region geometric and photometric properties. Then, for each class, a tree-structured belief network (TSBN) is learned, where nodes represent the corresponding image regions, and edges, their statistical dependecies. A given unknown texture is classified with respect to the maximum posterior distribution of the TSBN. Experimental results on the benchmark CUReT database demonstrate that our approach outperforms the state-of-the-artmethods.