Novel method for occlusion reduction and disparity refinement

Estimating the depth of the scene using stereo image pair has been one of the important problems to solve in computer vision. Depth estimation is done through stereo matching for which many approaches are proposed in the literature, each having its own advantages and disadvantages. In this paper, we are addressing problems associated with local area based algorithms for stereo matching like an error due to occlusion and wrongly matched pixels. The idea proposed in this paper makes a use of low-rank sparse matrix completion algorithm for computing the disparity of occluded and wrongly matched pixels. The proposed algorithm can be used on top of any enhancement for computing the disparity map, but for the simulation purpose, we have used the sum of absolute difference (SAD) and the sum of squared difference (SSD) matching. The simulations results show a gain in the range of 5 to 6 percent in structural similarity index (SSIM) with respect to SAD and SSD.