Concerning invariance it has been argued for the usage of object centred models. Likewise the anticipation of a combinatorial explosion, caused by the multitude of imaginable views, lead to the rejection of viewer centred representations. These considerations gave for example rise to the generalized cones introduced by Marr and Nishihara (1978) and the geon structural descriptions (GSD) proposed by Biederman (1985,1987), which represent objects by volumetric primitives together with their relations. Another kind of models, such as those of Thompson and Mundy (1987) and Lowe (1986), is based on surface descriptions built upon vertices, edges, and surfaces in conjunction with their connection relation.
Object centred models have the one purpose in common: the description of objects by high level features which provide stability over all perspectives. These are in the case of generalized cones the lengths, radii, and angles of the cones together with their connection relation, in the case of geons the attributes symmetry, size, form of the axis, shape of the edges, and the positional relations of the primitives. Whereas Thompson and Mundy (1987) use the arrangements of so-called vertex pairs and Lowe (1986) introduces non-accidental features as for example collinearity of points or line segments, parallelity of curvatures, end points, and equidistance of parallels. This brief survey shows that object centred models involve already for their features a high degree of complexity and computational efforts. Furthermore the recognition processes are based on extensive operations requiring the adjustment of the current input view and the object representations stored in the model database. As a matter of fact, a great number of representation schemes is merely designed for theoretical treatments without being implemented, not even for testing purposes. Among the above mentioned references only the models of Lowe (1986) and Thompson and Mundy (1987) were realized as functioning image processing systems. Nevertheless, the implementations are restricted to relatively simple and completely specified objects within primitive environments so that their applicability to real world input scenes seems questionable.