A Non-local Sparse Model for Intrinsic Images

Abstract

This paper deals with the intrinsic image decomposition problem, a long-standing ill-posed problem that decomposes an input image into shading and reflectance ones. Based on the observation that colors in the scene are usually dominated by a set of representative material colors, we sample material colors in the scene and recover a set of dominant material colors through a voting scheme. With this set of material colors, based on the assumption that pixels with similar chroma likely have similar reflectance values, we adopt global sparsity and non-local constraints on the reflectance and formulate the problem as a least-square minimization problem. We show the effectiveness of our method on a benchmark and demonstrate its use on a few applications.