MRF-Based Stereo Correspondence and Virtual View Interpolation

Stereo correspondence and virtual view synthesis play a crucial role in three-dimensional video systems. In this paper, the traditional stereo correspondence problem is modeled on Markov random fields (MRF) where the matching consistency is optimized by minimizing the global energy function. Both of the left-to-right disparity map and right-to-left disparity map are estimated simultaneously with the max-product belief propagation algorithm. The experimental results show that high quality intermediate views with significantly reduced hole regions can be obtained based on the bidirectional disparity maps.