Efficient Stereo Algorithm using Multiscale Belief Propagation on Segmented Images

TL;DR: This paper introduces a novel approach, based on a combination of segmentation and BP, which inherits the idea of Multiscale BP, however at each level of the hierarchy, each graph node corresponds to an image segment, which is called superpixel, instead of a fixed rectangular block of pixels.

Abstract: A variety of approaches using BP and image segmentation have been proposed for the stereo correspondence problem. In this paper, we introduce a novel approach, based on a combination of segmentation and BP. Our method inherits the idea of Multiscale BP, however at each level of the hierarchy, each graph node corresponds to an image segment, which we call superpixel, instead of a fixed rectangular block of pixels. The resulting de pth map at the coarser level is used to initialize the depths at the finer lev el. At the lowest level, we perform loopy BP on the four-connected pixel subgrid within each superpixel. The proposed method is applied to stereo images in the standard Middlebury dataset, and to real outdoor stereo images and car sequences. Experimental results show quite acceptable accuracy of depth inference, with running time fast enough for practical use.