Alternative confidence measure for local matching stereo algorithms

Abstract

The authors present a confidence measure applied to individual disparity estimates in local matching stereo correspondence algorithms. It aims at identifying textureless areas, where most local matching algorithms fail. The confidence measure works by analyzing the correlation curve produced during the matching process. The authors also test the confidence measure by developing an easily parallelized local matching algorithm, and use our confidence measure to filter out unreliable disparity estimates. Using the Middlebury dataset and our own evaluation scheme, the results show that the confidence measure significantly decreases the disparity estimate errors at a low computational overhead.