Efficient Handling of Over/Under-exposure in Stereo Vision

Stereo vision involves deriving scene depth from the differences in two images of a scene. It is classically formulated as a matching problem, with the results of matching easily translated to scene depth. Recent research has shown that the stereo vision problem can be represented as a succession of smaller problems, one of which is the problem of calculating the likelihood of two pixels matching each other, otherwise known as calculating the matching costs of the two pixels. This paper examines the assumption of intensity constancy inherent in all pixel matching costs currently in use, as well as the specific case where this assumption is violated, i.e. when one image of the stereo pair is over or under exposed. A generalization of the commonly used absolute difference measure for pixel matching is proposed and shown to perform well regardless of over/under-exposure of the input images, taken from the Middlebury Stereo Vision dataset.