Tinker Toy Recognition from 2D Connectivity

摘要

An end-to-end system has been built as a preliminary investigation into structure based object recognition in the tinker toy world. The system successfully performs constant time topological matching of TV input to a multiple-object model base. Object recognition can occur without reconstruction of the component surfaces, as evidenced by the recognizability of line drawings and silhouettes. We are investigating the use of object structure as a recognition cue. For example, many people will first view Figure 1 as an arbitrary collection of line segments. But once having been told that this is a sketch of a bicycle, the structural information attached to bicycles is sufficient to make this a recognizable rendition. Having accepted the premise that object recognition is based on structural cues, the need arises for a more precise definition of the term 'structure'. For the purposes of this document, we define object structure to be the spatial relations between the primitive components making up the object. The need also arises for a domain in which the primative components making up the object. The need also arises for a domain in which the primitive components of an object are easily extracted. One such domain is the tinker toy world. Tinker toys consist of wheels connected by rods that fit into eight radial holes and tow axial holes in the wheels. This paper describes an end-to-end system that takes a picture of a tinker toy, performs a structural analysis of the picture and matches the resulting structure to a model base of tinker toy principal views.