Fast hierarchical cost volume aggregation for stereo-matching

Some of the best performing local stereo-matching approaches use cross-bilateral filters for proper cost aggregation. The recent attempts have been directed toward efficient approximations of such filter aimed at higher speed. In this paper, we suggest a simple yet efficient coarse-to-fine cost volume aggregation scheme, which employs pyramidal decomposition of the cost volume followed by edge-avoiding reconstruction and aggregation. The scheme substantially reduces the computational complexity while providing fair quality of the estimated disparity maps compared to other approximated bilateral filtering schemes. In fact, the speed of the proposed technique is comparable with the speed of fixed kernel aggregation implemented through integral images.