Fast Stereo Matching with Recursive Refinement and Depth Upsizing for Estimation of High Resolution Depth

In this paper, a fast and precise stereo matching system, which is mainly composed of adaptive stereo matching, recursive refinement and depth upsizing subsystems, is proposed to estimate high resolution depth maps. The adaptive stereo matching subsystem, which adaptively uses color gradient, color intensity and texture census costs, can reduce the most of computation and achieve good primary depth maps. The recursive refinement subsystem consists of median filtering, left-right check, hole filling, and bilateral filtering functions. The refinement subsystem is performed recursively to reduce the most errors occurred in occlusion regions such that we can obtain high-precision depth maps. Finally, the depth upsizing subsystem by referring original high-resolution images performs depth value upscaling, texture-selected interpolation, and weighted depth voting processes to finally obtain the high-precision and high-resolution depth maps. To evaluate the proposed algorithms in three subsystems, the performances of the proposed stereo matching system are exhibited in step-by-step approaches. In comparison of the existing methods, the accuracy performance and computation time of the proposed stereo matching system are finally demonstrated. For future 3D ultra-high-resolution videos, the proposed stereo matching system with low computation will be a good solution to extract high-precision and high-resolution depth information.