Combining local belief from low-level primitives for perceptual grouping

International audienceSegmentation is usually unable to cope with artifacts due to slight change in lighting conditions or object occlusion for instance. That iswhy perceptual grouping is often used to overcome segmentation’s lacks. This refers to the ability of human visual system to impose structureand regularity over signal-based data. Gestalt psychologists have exhibited some properties which are used during perceptual grouping, such asproximity, continuity, or symmetry. Then, some implementations of these have been proposed in computer vision. However, most of these worksrely on contour-based primitives. Besides, they often use one single property to merge close regions, which may not be sufficiently robust. Wepropose a new framework for bottom-up perceptual grouping, which relies on a region-based segmentation. It allows us to use region or contourinformation, when it is the most suitable. Besides, we propose to trigger a grouping when several Gestalt properties support it. This couldincrease the robustness of perceptual grouping. We use Dempster–Shafer theory to combine the influence of several Gestalt properties overeach grouping, as it is especially designed for this. We also present numerous promising results, which show the efficiency of our approach