A depth map estimation approach for trinocular stereo

This paper presents an approach for dense depth estimation taking the input of a trinocular stereo. The approach works with a global energy minimization framework based on Markov Random Field models. The occlusion and spatial consistency constraints are explicitly considered in an iterative fashion. Depth maps are initialized by belief propagation using AD-Census metric, and then refined by Mean Shift Segments Fusion. We further incorporate the occlusion constraint for trinocular stereo to refine our depth map. In order to make the inference tractable, we implement our algorithm in an iterative scheme where the disparity map and occlusion map of each view are updated iteratively. Finally, three disparity maps corresponding to a trinocular stereo are estimated. Our approach addresses the problems such as untextured regions, occlusions, inconsistency in spatial region in a unified framework. The experimental results show our approach works well on trinocular stereo.