High accuracy correspondence field estimation via MST based patch matching

This paper presents an effective framework for correspondence field estimation. The core idea is to construct pixel-level and superpixel-level patch matching to achieve high accuracy estimation as well as fast speed computation. To this end, a hybrid edge-preserving supported weighting approach is first developed, which contributes to better performance on the pixel level, especially on those in the regions of fine structures. Then, a local Minimum Spanning Tree (MST) is constructed to describe regions and develop the adaptive smooth penalty weights, so that the over-patching in large textureless regions can be effectively avoided. In addition, the MST is further extended to handle occlusions in way of edge preserving strategy. Finally, all the above treatments are collected into an optimization model where the objective function is developed in terms of Markov Random Filed (MRF). In computation, a fast yet efficient iterative optimization strategy is developed. Our approach achieves favorable place on optical flow benchmark, which locates at the top two and top four for endpoint error and angular error evaluations among more than 130 approaches listed in the webpage.