Abstract

We propose a new approach to associate supervisedlearning-based confidence prediction with the stereo match-ing problem. First of all, we analyze the characteristics ofvarious confidence measures in the regression forest frame-work to select effective confidence measures using trainingdata. We then train regression forests again to predict thecorrectness (confidence) of a match by using selected confidence measures. In addition, we present a confidence-basedmatching cost modulation scheme based on the predictedcorrectness for improving the robustness and accuracy ofvarious stereo matching algorithms. We apply the proposedscheme to the semi-global matching algorithm to make itrobust under unexpected difficulties that can occur in out-door environments. We verify the proposed confidence mea-sure selection and cost modulation methods through exten-sive experimentation with various aspects using KITTI and challenging outdoor datasets.