Discrimination vision

Abstract

Listen

Translate article

Model-based vision systems must be capable of dealing with a large number of interpretations that are hypothetical and ambiguous. This paper is concerned with the representation of such interpretations. In order to suppress the number of interpretations many model-based vision systems use specialization hierarchies. We argue that such hierarchies cause a problem with uniformity in representation. In addition, they cannot prevent the occurrence of a combinatorial explosion in hypothetical interpretations. We propose to use discrimination graphs as a solution to both problems. Such graphs represent classes of objects with similar image features. The leaves of these graphs represent classes of elementary objects which can describe an image unambiguously. All other nodes represent abstract classes of such objects. Rather than invoking elementary object classes directly, as is done in most model-based vision systems, each image feature is mapped to the interpretation class representing the complete set of possible (elementary) object classes. As more and more constraints are discovered in the image, constraint propagation techniques force each interpretation class to become more specific. Discrimination graphs are closely involved in this process. This approach is therefore referred to as discrimination vision. The first part of this paper describes the discrimination vision approach to image interpretation and points out its advantages over more traditional approaches. The second part discusses the implementation of discrimination graphs in Mapsee-3, a sketch map interpretation program.