View article

Special-purpose vision systems have shown considerable success within their limited task domains (eg,[30, 56, 39, 42, 3]). To date, however, there have been no general purpose systems that work effectively across a variety of domains. Why do special-purpose systems succeed where general systems fail? We believe that special-purpose systems have succeeded because they are better able to define, structure, and apply knowledge that is relevant to their task domain. Knowledge in vision includes domain-independent knowledge about occlusion, perspective, physical support, etc., as well as domain-dependent object knowledge about attributes and relations, and object-specific control knowledge for recognizing objects in scenes. Object knowledge can encompass three-dimensional structure, two-dimensional appearance, and geometric and co-occurrence relationships with other objects and object parts …