Most current two-dimensional recognition systems are model-based: the representation of a scene is compared with each model in a database of models, or with each of a collection of features belonging to a database of models, until the scene is completely analyzed. It has recently been recognized that if one has a large database of models, it is necessary to index this database so that, when the system attempts to recognize a scene, it only retrieves the most likely models for further analysis. This paper deals with such an indexing mechanism.

First, a polygonal approximation of the object or scene boundary is represented by an ordered set of vectors, each of which indicates the internal angle, the distance to the vertex, and the coordinates of the vertex. The system selects a few “privileged strings” (subsets of five continuous vectors starting from a sharp corner) from each model, uses a similarity measure to form clusters of similar strings, and generates a similarity hierarchy. This hierarchy becomes the iconic indexing mechanism for model retrieval. The scheme efficiently supports the insertion and deletion of models by a simple index tree modification.

The paper clearly describes model building, index design, and model retrieval and verification, but the reader could use a detailed example of how the indexing structures are constructed from the feature vectors. The authors briefly describe the results of experiments using a model database of 11 shapes, and they provide a complete set of references.