Temporally Consistent Depth Map Estimation Based on 3D-MRF

Mutiview plus associated depth information form a typical 3D video representation. Depth map of each frame in a video sequence is usually estimated by stereo matching approaches separately. As a result it has weak temporal consistency. In this paper, we propose a novel framework based on spatio-temporal Markov Random Fields. It enforces temporal correlation by employing additional state in the graphical model. Improved belief propagation-sequential algorithm is exploited as an efficient optimization scheme to minimize the energy function. The experimental results demonstrate that the proposed method produces dependable depth maps in both spatial and temporal domain.