Accurate dense stereo matching of slanted surfaces using 2D integral images

This paper presents an advanced algorithm providing accurate stereo correspondences of two frames through concise disparity gradient estimation at the per-pixel level and 2D integral images. The key contributions of this novel algorithm are twofold: First, combining an upright cross-based support region with disparity gradient estimation realizes the implicit construction of a 3D support region for each anchor pixel. This approach yields the disparity accuracy for slanted surfaces as well as fronto-parallel surfaces. Second, the 2D integral image technique leads to a speedup of matching cost aggregation in the implicit 3D support regions. The experimental results show that the proposed algorithm can successfully convey the correspondences of actual sequence of outdoor stereo images and Middlebury stereo images with high accuracy in near real time.