Local Stereo Matching by Joining Shiftable Window and Non-parametric Transform

Abstract

In this paper, we propose a blocking-matching approach to the problem of correspondence in stereo matching. In blocking-matching methods, the correspondence or difference between pixels of a stereo pair is measured by a local window. Despite some area-based stereo matching algorithms have been developed and work well in a number of kinds of regions such as textureless or object boundary regions, their performance can debase once working in some sorts of radiometric conditions. Our proposed algorithm, in which non-parametric transform is used in the pre-processing step and an edge-preserving filter is used in the post-processing step, is an improved method of a shiftable window method and can work robustly in various radiometric conditions. Input images first are pre-processed by the census transform that makes our method robustly when the image pair is captured in various light sources or camera uncovering conditions. The window cost in our approach is calculated from the transformed images using in the Hamming distance, and the similarity is finally selected by a Winner-Takes-All strategy. The experimental results for the Middleburry images show that our proposed algorithm outperforms test local stereo algorithms in radiometrically dissimilarity images.