Novel multistage feedback approach for stereo correspondence

Stereoscopic depth estimation is widely used in ranging, machine vision and image rendering. Local methods and global optimization are the two main branches of depth estimation. Works on the former has been concentrated on window-correlation approaches. In this paper, we propose a multistage local method that operates on image segments instead of rectangular windows. This new approach not only addresses the unique characteristics of image segments, but also reduces occlusion through a feedback system. The proposed method has a low computational complexity. Experimental results show that it is very effective for natural images.