Segmentation-Based Stereo Matching Using Improved Self-Adapting Dissimilarity Measure Based on Second Order Smoothness Priors

Generally, image segmentation and self-adapting dissimilarity only based on first-order priors is not accurate enough for stereo matching. In this paper, to remedy this problem, segmentationbased stereo matching using improved self-adapting dissimilarity measure based on second order smoothness priors is proposed. We know differential geometric information can provide inherent interrelation in space geometry, which can be greatly helpful to acquire accurate correspondences. Inspired by differential geometry incorporating into global methods, this good property is once more incorporated into local methods. However, complex differential geometric property is easy to reduce speed of our methods. Therefore, to improve the accuracy and stability while keeping fast speed, the second order smoothness prior with geometric contextual information is incorporated into both image segmentation and self-adapting dissimilarity measure. Experimental results are evaluated on Middlebury data sets, and demonstrate usefulness of the second-order prior can be robust and accurate, especially in the complex conditions.