Abstract

While intrinsic image decomposition has been studied exten- sively during the past a few decades, it is still a challenging problem. This is partly because commonly used constraints on shading and reflectance are often too restrictive to capture an important property of natural im- ages, i.e., rich textures. In this paper, we propose a novel image model for handling textures in intrinsic image decomposition, which enables us to produce high quality results even with simple constraints. We also propose a novel constraint based on surface normals obtained from an RGB-D image. Assuming Lambertian surfaces, we formulate the con- straint based on a locally linear embedding framework to promote local and global consistency on the shading layer. We demonstrate that com- bining the novel texture-aware image model and the novel surface normal based constraint can produce superior results to existing approaches.