Disparity Refinement Based on Feature Classification and Local Propagation for Stereo Matching

Stereo matching usually makes up of four steps: cost computation, cost aggregation, disparity optimization, and disparity refinement. The disparity refinement is used to further eliminate mismatches caused by occlusion, low texture, and other factors. The popular refinement methods are based on the consistency check of left and right two disparity maps. For efficiency, we propose a novel multistep disparity refinement framework using only one-sided image, which is organized into four main steps: leftmost occlusion detection, four-directional scanline outlier detection, black hole detection and eight-directional disparity propagation. Experimental results on Middlebury datasets show that our method is comparable with other postprocessing strategies, especially in occlusion handling, retaining object shapes and preserving discontinuities.