Depth information recovery based on monocular videos

Depth information plays an important role in3-d modeling area. In this paper, a depth recovery algorithm for monocular videos is proposed. With the help of results from camera self-calibration, the matching cost between the pixel in the current frame and its projection pixel in neighbor frame, can be calculated by multi-view geometry theory, and based on a cost function, which is consisted of color consistent constraint, geometric consistent constraint and local optimization result, we use belief-propagation algorithm to get a global optimization result. With lots of experiments on a variety of challenging videos, the results show that the boundaries of objects are kept well, and the area where the depth changes smoothly is recovered very well.