Knowledge About Interfacing Descriptions

Abstract

This concentrates on interactions between knowledge stated in diverse representations. It proposes a vision program that classifies any complicated object as an elaborated instance of a simple on it already understands. The resulting global-local connections facilitate evaluation of overall properties, such as visual shape and ability to support other objects. Flexibility is achieved through simultaneous use of multiple equivalent representations. These are coordinated via interfacing rules for giving hints, constraining choices, and filling in missing detail, making use of the great redundancy in most visual scenes. An important feature of the system consists of domain-dependent rules for guiding the flow of control and choosing hypothesis.