Adaptive differential refinement of block-matching based correspondence vector fields

In this paper we propose a post-processing method for block-matching based correspondence vector fields. We combine image adaptive filtering for object accurate vector field refinement with differential floating point precise optical flow estimation. More precisely, image warping and large cross-adaptive filter kernels are integrated into the Horn and Schunck optical flow estimation approach to break the block structure of the input correspondence vector fields and compute flow field updates fulfilling the smoothness constraint inside object borders. To reduce the influence of occlusion areas we integrate in-loop occlusion detection and adjust the adaptive filter weights after every iteration.