Abstract: This Issue of The International Journal of Computer Vision contains four papers from the University of Rochester. The papers reflect an underlying theme that has guided much of the vision research at Rochester for the last few years, namely, that understanding the phenomenon of intelligence and discovering how to produce an artificial one must proceed in the context of behavior. Applied to machine vision, the theme states that the full potential of vision as evidenced by its operation in biological systems can be achieved only by considering it in terms of intelligent behavior.
One reason for advocating a behavioral approach to artificial intelligence is that most ideas for machine intelligence are inspired by the abilities of animals, and human abilities must have derived from pre-existing capabilities in the ancestral forms. Until we can understand (and demonstrate our understanding by implementation) how a bee flies, how a rabbit runs, how a wolf hunts, we cannot hope to understand how a human plays chess, composes a symphony, or washes the dishes. Another reason for taking a behavioral approach to vision is that the total quantity of information contained in a visual signal is far greater than any system needs or can handle. In most proposed applications of vision, all but a tiny fraction of this information is irrelevant. Despite this fact, many approaches attempt to preserve as much of the original information as possible on the grounds that one never knows what one might need. We believe that such a least commitment strategy is exactly the the wrong approach. The strategy that should be followed is one of "most commitment" where we throw out as information as quickly as possible on the grounds that what is thrown out does not have to be processed. The behavioral approach provides a mechanism for deciding what can be thrown out via the use of a-priori knowledge about the functionality of the system. Implementing vision in terms of behavior may also drastically reduce the amount of representation that is required. Finally, there is the issue of technological timeliness. With recent developments in sensors, parallel and special purpose processors, and robot hardware, it is now becoming possible to explore through synthesis the complex behaviors that constitute intelligence.