SWAP-MOVE WITH LONGITUDINAL NEIGHBORING OPTIMIZATION AND MAXIMUM A POSTERIOR ESTIMATE FOR VISUAL CORRESPONDENCE

摘要

The stereo vision is based on accurate image calibration. In practice, the low-cost stereo vision systems have limited calibration accuracy. Therefore, the matching points will have disparities not only in the latitudinal neighborhood, but also in the longitudinal neighborhood, which will seriously affect the performance of existing algorithms. In this paper, we redesign the swap-move strategy with both longitudinal and latitudinal disparities, which expand the optimization method from one-dimensional to two-dimensional in the max-flow optimization procedure. We prove that it can be locally optimal from the perspective of maximum a posteriori (MAP) estimation. Finally, the effectiveness of this algorithm is verified by real data experiments.