Recognizing objects in a natural environment: a contextual vision system (CVS)

Abstract : Existing machine vision techniques are not competent to reliably recognize objects in unconstrained views of natural scenes. In this paper we identify a number of weaknesses in current recognition systems, including an inability to solve the partitioning problem or to effectively use context and other types of knowledge beyond that of immediate object appearance. We propose specific mechanisms for dealing with some of these problems and describe the design of a vision system that incorporates these new mechanisms. The system has been partially implemented and we include some experimental results indicative of its operation and performance.