Recognition Cones That Perceive and Describe Scenes that Move and Change Over Time

Abstract

This paper describes a "recognition cone" model of the perceptual system that can input and process a continuously changing image of a scene (e.g. the successive frames of a movie or television camera). It recognizes and describes the things (mixtures of words and objects) in the scene, including their parts, qua1ities, and interrelations. It uses a hierarchy of configurational characterizing transforms, where a transform can imply a)other transforms to apply, as well as b)possible names and other descriptive information to assign to the scene, and c)triggers to choose among possibilities in subregions of the scene. Choices of the components of the description are made from the set of implied possibilities, where each can be imp1ied by a variety of contextually interrelated characteristics. The system will output either a "complete" or a stylized description, or a description that is the dynamic consequence of a user's conversational sequence of interactive requests for more information. This system is actually the perceptual "front end" of a wholistic cognitive system that calls for and uses its internal descriptions to help choose and carry out sequences of actions, both internal (e.g. to deduce, remember) and external (e.g. to move, manipulate, glance about).