Laws@SRI-AI.ARPA (Ken Laws) (06/25/86)
I doubt that it makes much difference what inference engine Mark Nagel uses for his vision problem, as long as it allows calls to external routines. Since almost the entire vision problem must be handled by procedural attachment ("conventional image analysis techniques"), the inference engine need only provide the capabilities of a simple programming language. A probabilistic or fuzzy-reasoning system such as Prospector might have considerable advantage over logic-based approaches, but would have much the same flavor as the statistical techniques that Mark wishes to avoid. The real problems in visual pattern recognition are in computing robust descriptors (esp. if they must be computed quickly) and in the knowledge-representation (i.e., knowing what kind of descriptors to compute and how to store the answers). Very little of the problem has to do with logical reasoning, forward or backward chaining, etc. -- Ken Laws -------