Practical Aspect-graph Derivation Incorporating Feature Segmentation Performance

A procedure is described for the automatic derivation of aspect graphs of surface-based geometric models. The object models are made of finite, typed, second order surface patches allowing the representation of a large number of complex curved objects while retaining ease of recognition. A new representation, the detectability sphere, is developed to encode feature detectability constraints. The detectability metric is directly related to the performance of the imaging system, allowing the generated aspect graph to more truthfully represent the scene's relationship with the vision system. An algorithm is described which fuses information from several views of the object to produce a small number of characteristic views which cover some desired portion of the viewsphere, and annotates these fundamental views with pose-verification hints. The procedure is compared with previous analytic and approximate solutions to the aspect-graph problem regarding relevance to the vision process, range of applicability, and computational complexity.