Stereo Correspondence by Dynamic Programming on a Tree

摘要: Dynamic programming on a scanline is one of the oldest and still popular methods for stereo correspondence. While efficient, its performance far from state art because vertical consistency between scanlines not enforced. We re-examine use dynamic correspondence by applying it to tree structure, as opposed individual scanlines. The nodes this are all image pixels, but only "most important" edges 4 connected neighbourhood system included. Thus our algorithm truly global optimization method disparity estimate at pixel depends estimates other unlike based methods. evaluate benchmark Middlebury database. very fast; takes fraction second typical image. results considerably better than that art, offers good trade off in terms accuracy computational efficiency.