Representing and using functional definitions for visual recognition

This thesis addresses a deficiency in artificial intelligence research on visual recognition. Visual recognition of objects or functional relationships between objects usually relies on visual information available in a static scene. Motivated by psychological theories of concepts, this thesis argues for the importance of function for visual recognition and demonstrates how functional definitions may be represented and used for recognition. For some time, some psychologists have proposed a core vs identification procedure distinction between the various components of a concept. Translated into concepts of our concern, this means a functional definition vs symptomatic perceptual condition distinction. Symptomatic perceptual conditions specify what visual-featural or spatial-relational information available in a static scene must be detected for the recognition of an instance of a concept. Functions, however, are characterized in terms of interactions between objects. Functional definitions hence specify what behaviour of objects involved in the interaction over time must be observed for certain functions to be present. Visual instances of a typical concrete concept, such as chair, are usually specified by a set of disjunctive symptomatic perceptual conditions. This specification, however, does not allow generalizations to novel visual instances. A single functional definition, on the other hand, can admit a much wider range of visual instances than any single symptomatic perceptual condition and it can also allow the recognition of visually novel instances. Observation over time is not always necessary to determine function. Functional definitions coupled with knowledge of physical behaviour allow the system to infer function based on information available in a static scene. The emphasis of this research is in the representation and use of functional definitions for inferring function from static scenes. Two concepts are studied in detail in a computational and representational framework. One is a concept that describes a functional relationship between two objects--support. The other is an artifactual concept--chair. The generalization of the basic computational and representational framework to other related concepts are also discussed, but these concepts are not studied in computational and representational detail.