| Bounding contours of physical objects are often fragmented by other occluding objects. This creates the need for perceptual grouping, or the association of those fragments corresponding to the same object. Perceptual grouping is currently a bottleneck to computer vision, since previous approaches either were heuristic, assumed the object belonged to restricted classes in order to exploit geometric invariants, or hypothesized boundary continuations. I introduce an alternative by invoking the principle that those fragments should be grouped whose fragmentation can be accounted for by a virtual occluder, and introduce the gap skeleton as a representation of this occluder. Properties of gap skeleton are proved, and an algorithm for computing it is given. Finally, a different perspective on gap skeleton is obtained in the domain of shape decomposition, where ligature emerges as the analog to gap skeleton. This leads to a new definition of a limb as a formal part, and suggests that both grouping and shape decomposition share a common basis. |
