Reducing Search Space for Stereo Correspondence with Graph Cuts

In recent years, stereo correspondence algorithms based on graph cuts have gained popularity due to the significant improvement in accu racy over the local methods. Even though there has been a noticeable progress in efficient max-flow algorithms, the computational cost for graph cut st ereo is still quite heavy, especially if the disparity search range is large. In this paper, we investigate and compare several ways of limiting the disparity search range. We show that the immediately obvious ideas based on thresholding or the hierarchical approach do not work reasonably well. We do, however, find that we can utilise the results of fast local correspondence meth ods for disparity range reduction of the more expensive graph cuts method. The idea is to understand and exploit the ways in which the local stereo correspondence methods fail. We are able to achieve 2.8 times average speed-up with only a modest degradation in performance, 1.7% average energy increase.