Robust estimation of multiple surface shapes from occluded textures

This paper examines the problem of estimating surface shape from texture in situations in which there are multiple textures present due to texture discontinuities, occlusion, and pseudo-transparency (for example looking through a picket fence at a textured surface). Previous shape-from-texture methods that use changes in the spatial frequency representation of neighboring image patches assume that only a single texture is present in each of the patches. The authors extend these approaches to situations in which multiple textures may be present. The authors provide a theoretical analysis of the multiple texture problem and the effect of the texture discontinuities, occlusion, etc. on the spatial frequency representation. The authors also present an algorithm, using robust mixture models, for recovering multiple surface shapes from occluded textures. The method performs well on real and synthetic images with results which are comparable to that of shape from texture with only one texture.