Abstract
We present a generative, probabilistic model that decom- poses an image into reflectance and shading components. The proposed approach uses a Dirichlet process Gaussian mixture model where the mean parameters evolve jointly according to a Gaussian process. In con- trast to prior methods, we eliminate the Retinex term and adopt more general smoothness assumptions for the shading image. Markov chain Monte Carlo sampling techniques are used for inference, yielding state- of-the-art results on the MIT Intrinsic Image Dataset.