Hybrid tree guided patch matching

Either any of the current global or non-local stereo matching algorithms cannot be good enough to show both matching accuracy and calculation efficiency during the matching processing, especially while there are less texture regions or the images are captured from real scene. Therefore, the goal of our research is to break current bottleneck of stereo matching in aspects of the precision and speed, then get a relatively perfect method compared with other current stereo matching algorithms. Based on this ambitious goal, we proposed hybrid tree guided patch matching algorithm to get a dense and accurate depth image in fast processing speed. We utilize (1) pixel-level and region-level minimum spanning tree to achieve an initial disparity value searching constraint by using hybrid tree cost aggregation, and then (2) apply a robust guided patch matching method to calculate the final accurate disparity of each pixel efficiently by using a cost aggregation restricted through the hybrid tree generated disparity value. In the experience, we demonstrate that our proposed algorithm can generate a high quality depth images and better efficiency compared with recent new stereo matching algorithms. In the Middlebury evaluation, our algorithm also got top ten ranking and better performance among the most of the global stereo matching algorithms in both accuracy and efficiency.