Recovering Intrinsic Images by Minimizing Image Complexity

This paper tackles the problem of decomposing a single image into two intrinsic images – a shading image (the illumination at each point) and a reflectance image (the colour at each point). Assuming a linear response of the camera, the acquired image I(x,y) is modelled as the product of the shading S(x,y) and the reflectance R(x,y) (collectively called intrinsic images): the goal is to recover S and R from I(x,y). The proposed method stems from the observation that R is "simpler" than I, in some sense related to its information content. This allows to formulate the problem as the minimization over all the possible S of a cost function describing the complexity of a tentative reflectance image given a shading image S. Given a 3D model of the scene, the orientation of the camera, and an illumination model, S can be parameterized with the position of light sources on a hemisphere. Preliminary experiments in a simulated environment validate the substance of the method, although many details will be subject of further improvement.