Abstract
In this paper we introduce a principled approach to modeling the im- age brightness constraint for optical flow algorithms. Using a simple noise model, we derive a probabilistic representation for optical flow. This representation sub- sumes existing approaches to flow modeling, provides insights into the behaviour and limitations of existing methods and leads to modified algorithms that out- perform other approaches that use the brightness constraint. Based on this repre- sentation we develop algorithms for flow estimation using different smoothness assumptions, namely constant and affine flow. Experiments on standard data sets demonstrate the superiority of our approach.