Control of Selective Perception Using Bayes Nets and Decision Theory

A selective vision system sequentially collects evidence to answer a specic question with a desired level of condence. EÆciency comes from processing the scene only where necessary, to the level of detail necessary, and with only the necessary operators. Knowledge representation and sequential decision-making are central issues for selective vision, which takes advantage of prior knowledge of a domain's abstract and geometrical structure (e.g. \part-of " and \adjacent " relationships), and also uses information from a scene instance gathered during analysis. The TEA-1 selective vision system uses Bayes nets for representation, benet-cost analysis for control of visual and non-visual actions, and its data structures and decision-making algorithms provide a general, reusable framework. TEA-1 solves the T-world problem, an abstraction of a large set of scene domains and tasks. Some factors that aect the success of selective perception are analyzed by using TEA-1 to solve ensembles of randomly produced, simulated T-world problems. Experimental results with a real-world T-world problem, dinner table scenes, are also presented