A joint stereo matching in the pixel and image level

Image noise, textureless regions, and occlusions are still problems of stereo matching. We propose a novel method to address these problems. Firstly, initial disparity map and reliability map are obtained by stereo matching in the pixel level. In this stage, the proposed approach not only imposes the photo-consistency constraint, but also explicitly associates the geometric coherence to solve the problem of occlusion. However, after this stage in the textureless regions there are still some bad pixels whose disparities are wrongly estimated. In the second stage, we use twice surface interpolation in the image level to correct these bad pixels via image segmentation and initial disparity map segmentation. Quantitative evaluation results show that it outperforms all the other local methods using edge-aware filtering in terms of accuracy on Middlebury benchmark. And subjective comfort of experimental results outperform other stereo matching algorithms. The experiment results show that the proposed method can obtain accurate disparity map and handle problems of stereo matching very well.