Efficient disparity estimation using region based segmentation and multistage feedback

Stereoscopic analysis is widely used in machine vision applications. Local and global methods are two main branches of stereoscopic analysis. The global methods typically minimize a cost function over the entire scene. Although these methods provide high estimation accuracy, because of its high complexity, they are not suitable for real-time implementation. The local methods typically use window-correlation approaches, and the associated complexity is generally low. However, the estimation accuracy is sensitive to the selected window size. In this paper, we propose a multistage local method that operates on image segments instead of traditional rectangular windows. This new approach exploits the unique characteristics of image segments, and reduces occlusion through a feedback system. Experimental results show that it is very effective for natural images. In addition, it has a low computational complexity which may be suitable for real-time implementation.