Recovering Depth Map from Video with Moving Objects

In this paper, we propose a novel approach to reconstructing depth map from a video sequence, which not only considers geometry coherence but also temporal coherence. Most of the previous methods of reconstructing depth map from video are based on the assumption of rigid motion, thus they cannot provide satisfactory depth estimation for regions with moving objects. In this work, we develop a depth estimation algorithm that detects regions of moving objects and recover the depth map in a Markov Random Field framework. We first apply SIFT matching across frames in the video sequence and compute the camera parameters for all frames and the 3D positions of the SIFT feature points via structure from motion. Then, the 3D depths at these SIFT points are propagated to the whole image based on image over-segmentation to construct an initial depth map. Then the depth values for the segments with large reprojection errors are refined by minimizing the corresponding re-projection errors. In addition, we detect the area of moving objects from the remaining pixels with large re-projection errors. In the final step, we optimize the depth map estimation in a Markov random filed framework. Some experimental results are shown to demonstrate improved depth estimation results of the proposed algorithm.