Abstract

We propose new dense descriptors for texture segmenta-tion. Given a region of arbitrary shape in an image, thesedescriptors are formed from shape-dependent scale spacesof oriented gradients. These scale spaces are defined byPoisson-like partial differential equations. A key propertyof our new descriptors is that they do not aggregate imagedata across the boundary of the region, in contrast to exist-ing descriptors based on aggregation of oriented gradients. As an example, we show how the descriptor can be incor-porated in a Mumford-Shah energy for texture segmentation. We test our method on several challenging datasets for texture segmentation and textured object tracking. Experiments indicate that our descriptors lead to more accuratesegmentation than non-shape dependent descriptors and thestate-of-the-art in texture segmentation.