Plane-Constrained Efficient Large-scale Stereo Disparity Estimation Algorithm

As a light-weight and efficient stereo disparity estimation algorithm, ELAS (Efficient Large-scale Stereo) has gained a lot of attention in the community of vision-based robotics. However, ELAS directly generate the disparity map from the Delaunay triangulation mesh of detected support points, which ignores the fact that these points may reside on different physical planes and consequently results in big errors on planar surfaces. In this paper we introduce the plane constraint into ELAS for better performance and robustness. We first cluster the selected support points using mean shift segmentation before triangulation, and then use the classified point sets belonging to different planes to triangulate separately, which avoids the errors caused by the cross-plane triangulation neglecting possible depth discontinuities in ELAS algorithm. Moreover, our algorithm also gets parametric plane models by fitting the large flat planes to rectify some mismatched support points. We evaluated our algorithm on both the Middlebury dataset and our own captured indoor scenes, and the results show that our approach behaves better in occlusion and weak texture areas with faster speed.