Perceptual grouping for generic recognition

摘要

We address the problem of recognition of generic objects from a single intensity image. This precludes the use of purely geometric methods which assume that models are geometrically and precisely designed. Instead, we propose to use descriptions in terms of features and their qualitative geometric relationships. To succeed, it is clear that these features need to be high level, rather than points or lines. We propose to detect groups using perceptual organization criteria such as proximity, symmetry, parallelism, and closure. The detection of these features is performed in an efficient way using proximity indexing. Since many groups are created, we also perform selection of relevant groups by organizing them into sets of similar perceptual content. Finally we present an implementation of a recognition system using these sets as primitives. It is an efficient colored graph matching algorithm using the adjacency matrix representation of a graph. Using indexing, we retrieve matching hypotheses, which are verified against each other with respect to topological constraints. Groups of consistent hypotheses represent detected model instances in a scene. The complete system is illustrated on real images. We also discuss further extensions.