Local Disparity Refinement with Disparity Inheritance

It is well known that it is difficult for local stereo algorithms to obtain correct disparities at occluded regions and depth discontinuities. In order to increase matching accuracy, current algorithms always use some disparity refinement measures. An adaptive algorithm to refine disparities based on color and distances has got good results but still not satisfactory. This paper proposes a simple and novel method which employs disparity inheritance within above algorithm, further improving the accuracy and robustness of stereo correspondence. According to our experimental results on Middlebury images, disparity inheritance can connect regions at different depths, and implement the diffusion of initial disparities from accurate areas to less accurate areas. Disparities are obviously improved especially at occlusion areas and depth discontinuities, and we get disparity maps with sharper edges.