Abstract
We present a method for estimating intrinsic images from a fixed-viewpoint image sequence captured under changing illumination di- rections. Previous work on this problem reduces the infiuence of shadows on refiectance images, but does not address shading efiects which can sig- nificantly degrade refiectance image estimation under the typically biased sampling of illumination directions. In this paper, we describe how biased illumination sampling leads to biased estimates of re?ectance image de- rivatives. To avoid the efiects of illumination bias, we propose a solution that explicitly models spatial and temporal constraints over the image sequence. With this constraint network, our technique minimizes a regu- larization function that takes advantage of the biased image derivatives to yield refiectance images less infiuenced by shading.