Window-Based Three-Dimensional Aggregation for Stereo Matching

This paper presents a window-based three-dimens-ional (3-D) aggregation technique, which can approximate the surfaces of all kinds of objects, in stereo matching. The 3-D aggregation, which means to aggregate in 3-D surfaces, is implemented by decomposing the adaptive support window into horizontal segments; we allow the disparity to change smoothly in or between segments. Compared to traditional local stereo methods, the 3-D aggregation greatly improves the accuracy of results in slanted surfaces and occlusion areas while keeping excellent performance near depth discontinuities. We also propose an acceleration method that surprisingly improves the accuracy at the same time. The evaluation experiments confirm our achievements.