Performance of Stereo Methods in Cluttered Scenes

This paper evaluates the performance of different stereo formulations in the context of cluttered scenes with large number of binocular-monocular boundaries (i.e. occlusion boundaries). Three stereo methods employing three different constraints are considered. These are basic (Basic), uniqueness (KZ-uni), and visibility (KZ-vis). Scenes for the experiments are synthetically generated and some are shown to have significantly more occlusion boundaries than the Middlebury scenes. This allows evaluating the methods with different types of scenes to understand the efficacy of different constraints for cluttered scenes. The evaluation considers mislabeled pixels of different types (binocular/monocular) in different regions (on or away from occlusion boundary). We have found that for sparse scenes (fewer occlusion boundaries) all three methods have similar performance. For dense scenes the performance is dominated by pixels on the boundary. For binocular pixels Basic always does better but for monocular pixels KZ-vis has the lowest error. If binary occlusion labeling is considered then the cross-checked version of basic constraint Basic-cc performs best followed by KZ-uni.