Stereo matching with superpixels

Abstract

The purpose of a short-baseline stereo matching algorithm is calculation of the disparity map for a given rectified stereo pair. Most stereo algorithms presented in the literature achieve this goal by searching all possible disparities for all pixels. This may be extremely costly if an image size and search range are large. In this paper an efficient method of computing a disparity map for the whole image, by searching disparities for only a selected subset of pixels, is presented. It is shown how to select good initial set of pixels for matching given the over-segmentation of the input images. It is also presented that such a limited set of pixels is sufficient to obtain a rough estimate of the disparity map. Next, it is described how this disparity map can be improved. The results indicate that despite searching only a fraction of possible disparities for just a subset of input image pixels, a computed disparity maps are of quality comparable to those obtained with a full search algorithm. Moreover, the method is general enough to be used with almost any local stereo matching algorithm.