Studying control of selective perception using T-world and TEA

The author hypothesizes that selective perception allows more accurate solutions to visual tasks to be found in less wall-clock time than non-selective techniques. The best way to assess the practical truth of this hypothesis is by studying, designing and building complete vision systems-the issues are fundamentally systems issues. On the other hand, special-case systems are not convincing: he presents the T-world problem as an abstraction of an interesting class of real-world vision problems. T-world has enough structure to support basic study of fundamental tradeoffs inherent in selective computer perception. The complete system is called TEA-1: it is a purposive and sufficing vision system that solves a version of the T-world problem. TEA-1 is a fully implemented system, and extensive experiments in the laboratory and simulation have explored the key factors that make the selective perception approach appealing, analyzing how each factor affects the overall performance of TEA-1 when solving a set of automatically generated (in simulation) T-world domains and tasks