Multilayer RGBD-Video Completion for Hole Filling in 3D-View Synthesis

摘要

Recent advances in video completion have made it possible to plausibly fill in large missing regions for a wide variety of unconstrained scenes captured with a freely moving camera. These advances enable automatic generation of high-quality background plates for filling holes during virtual-view synthesis. Most hole-filling approaches, however, still tend to impose restrictive assumptions, such as approximating the background as a single plane, or they forgo background reconstruction by copying from other input-video frames altogether. Moreover, all existing approaches either directly or indirectly rely on binary separation of the scene into background and foreground, which is sometimes not possible. In this paper we propose a multilayer extension of the video-completion algorithm in [5] to RGBD video and demonstrate its advantages compared with alternative approaches when filling holes for 3D-view synthesis.