Temporally Consistent Depth Maps Recovery from Stereo Videos

Dense depth maps provide significant geometry information for 3D video and free viewpoint video systems. Traditional stereo correspondence methods usually deal with each stereo image pair separately. As a result, the generated depth sequence is temporally inconsistent. This paper presented a novel approach to recover spatio-temporally consistent depth maps. The proposed method first applied sequential belief propagation algorithm to achieve an approximate minimum of spatial energy on Markov random fields. Then in temporal domain, a smoothness cost along optical flow was incorporated between consecutive frames. The combined cost which determined the disparity value was passed forward and temporal consistency was enforced during the process. In addition, the streamlined implementation was time and memory efficient. In experimental validation, quantitative evaluation as well as subject assessment was performed on several test datasets. The results showed that the proposed method yielded temporally consistent depth sequence and reduced flickering artifacts in the synthesized view while maintaining visual quality.