Fast temporal filtering of depth maps

Abstract

In this paper, we propose a method of filtering depth maps that are automatically generated from video sequences using optical flow, 3D reconstruction and scene analysis methods. To attain better quality, information from both the source video and depth map is used. The proposed algorithm uses motion estimation to take into account temporal information, but the algorithm’s structure permits use of optical flow to improve quality, but at the expense of greater computation time. The method can be applied as a preprocessing stage for enhancement of multi-view or stereo video. Joint temporal and spatial processing can yield further improvements in quality. A comparison of the results with test ground-truth sequences using the BI-PSNR metric is presented.