At the workshop that led me to write my "position paper," I was struck by the diversity of the opinions that the participants expressed. At the start of the workshop, as chairman, I presented an earlier version of the position paper published here, and called for comments. Everyone did indeed have something to say, but in many cases their comments were not directly related to my "positions." When the current version of the paper was circulated for written comments, a similar phenomenon occurred; not all of the responses are equally focused on the issues raised in my paper, and in any case I do not strongly disagree with any of them. I will therefore not discuss the responses in detail, but will only use this space for a few concluding remarks.
As several of the responders agree, spatial reasoning is a difficult and poorly understood process, and traditional expert systems techniques will not be very useful in this area. It may be wisest to work on "unnatural" visual domains in which trained experts are in fact needed. In any case, we need a theory of feature information, perhaps formulated in Bayesian terms. We also badly need architectures specialized for vision and for spatial reasoning. If we do attempt to build "expert" vision systems, we will need to develop methods of performance evaluation, and we will also need to develop effective user interfaces, on the graphic as well as the symbolic level.
Much of the best vision research during the last decade has concentrated on "low-" and "intermediate-" level aspects, notably on techniques for handling images of three-dimensional and time-varying scenes. Relatively little work has been done on the development of "high-level" (i.e., "expert"?) vision systems. I believe that we will see significant steps in that direction during the coming decade; but it will not be easy.