Abstract
Most current industrial vision systems are designed to recognize known objects seen from standard view points in high contrast scenes. Their performance and reliability are marginal; many tasks including such as bin picking, recognition of parts on overhead conveyors, and implicit inspection of surface flaws are beyond current competence. Recent image understanding research suggests that the limitations of current industrial vision systems stem from inadequate representations for describing scenes; physical attributes (reflectance, texture, etc.) and three-dimensional pictorial features and object models. This paper builds a case for needed additional levels of representation and outlines the design of a general-purpose computer-vision system capable of high performance in a wide variety of industrial vision tasks. This paper was originally presented at the General Motors Symposium on “Computer Vision and Sensor-Based Robots,?? September 1978, the proceedings of which will be published by Plenum Press in 1979.
Share this
