Outlier detection and disparity refinement in stereo matching

Disparity estimation in ill-posed regions, such as occlusions, repetitive patterns and textureless regions, is a challenging problem in stereo matching. The initial disparities obtained in these regions tend to be regarded as outliers that must be detected and addressed. In this paper, two outlier detection methods are proposed, i.e., the efficient approach and the accurate approach. The efficient approach detects outliers by exploring the disparity map for the left image only and reduces runtime and memory costs. First, the match fixed point jumps (MFPJ) algorithm is proposed as an initial solution to detect outliers. Then, a high-probability outlier detection algorithm is proposed to accomplish denser outlier detection with less noise. The accurate approach first classifies outliers as occlusions or mismatches. Then, 3D label assignment is performed for occlusion outliers and normal-based plane fitting is conducted for mismatch outliers to refine the disparities of the outliers and to achieve an accurate stereo matching result. Evaluations of the Middlebury datasets demonstrate that the proposed methods effectively improve the stereo matching performance. (C) 2019 Elsevier Inc. All rights reserved.