On recognition of features in polyhedral scenes

A system that incorporates both high level representation of image data and a recognition and interpretation capability is described. Both competences are based on a structural, rather than metric, description of objects and image features and the corresponding models of object categories. The current implementation in the domain of polyhedral scenes transforms a set of 3D segments, our basic data representation, into a set of shape primitives such as planes and convex polyhedra (boxes) or planar configurations of segments (patterns). The interpretation module then uses its knowledge of the application domain to interrogate the representation in search for the expected evidence of a particular object category. In this paper we investigate recognition of general object categories based on a structural, symbolic description of objects and features using simple shape primitives as proposed by Connell and Brady [8] and other authors. Even a small number of highly symbolic primitives can lead to a complex matching problem if one attempts to match all the scene primitives to all the model primitives in our world domain. In our work we adopt a task driven approach and, instead of investigating all the possible matches, our recognition module interrogates the scene representation in search for evidence of a single desired object or feature. Some aspects of our system ( COMPACT ) have been described earlier [9, 10, 11]. In this paper we shall give a brief outline of the whole structure and describe in more detail the recognition process.