Weighted omnidirectional semi-global stereo matching

Abstract

Traditional semi-global matching (SGM) lacks interaction between scanlines and struggles to deal with the ambiguity of pixels in homogenous areas. In this paper, we propose a novel path-centering graph to perform weighted omnidirectional SGM (WOdSGM), in which the input image is divided into eight sub-trees, corresponding to eight optimization directions. In each pass, the outputs of pixels are recursively computed from leaf nodes to the root node along the tree structure. As the results of SGM from multiple scanlines often show different biases owing to respective message propagation directions, one solution is accurate in certain areas while others are not. By delving into the distribution of aggregated costs for pixels from various structures, we find that the minimum cost of reliable output is heavily biased from others. Therefore, we perform line fitting to approximate the distribution of aggregated costs and use the distance between the fitted line and the point corresponding to the minimum cost to evaluate the reliability of aggregated costs. Furthermore, we present a weighted fusion strategy to incorporate outputs from multiple directions, enabling our method to provide directionally biased constraints. Extensive experiments and analyses on widely used stereo datasets show that our approach outperforms typical traditional SGM-based algorithms.