Fast Binocular Depth Inference via Bidirectional Motion Based Interpolation

Depth information provides fundamental supports to multimedia applications for both images and videos. Depth acquisition for stereo images has drawn much attention while few approaches are proposed for stereo videos. Conducting stereo matching frame-by-frame is time consuming and the result is temporally inconsistent. As a matter of fact, the redundancy shared by frame sequences may cause extra computational cost. Inspired by rapidly acquiring stereo video depth for some specific applications, we propose a novel bidirectional motion-based interpolation framework, which avoids frame-by-frame matching through making use of the motion estimation and the redundancy between frames. Firstly, comparable accurate depth maps are generated for self-adaptive selected frames via stereo matching. Then rough depth sequences inbetween are calculated using bidirectional motion-based interpolation. To improve the depth accuracy for non-selected frames, we propose a refinement approach to handle cracks and holes. The evaluation on both computer rendered and real world captured datasets show that our approach is competent for fast and accurate binocular video depth acquisition.