Occlusion-Aware Stereo Matching

摘要

Stereo vision systems with additional flash/no-flash cues have been demonstrated to be robust to depth discontinuities. The ratio of a flash and no-flash image pair naturally provides additional scene depth information and thus can serve as a strong cue for preserving depth discontinuities. However, existing solution simply uses ratio as the guidance to perform matching cost aggregation and thus is still vulnerable to occlusions. Inevitable misalignment of flash and no-flash images due to camera and/or scene motion remains unsolved as well. This paper investigates into these two problems. An occlusion detection approach is derived based on foreground/background extraction. Matching cost computed in the occluded regions (which is useless and harmful) is thus discarded so that reliable information from non-occluded regions can be easily propagated in. The foreground, occlusion and depth estimation is modeled in a uniform framework base on Expectation-Maximum. The proposed solution is evaluated using both indoor and outdoor data sets, showing clear improvement over the state-of-the-art methods.