KNOWLEDGE-DIRECTED VISION - CONTROL, LEARNING, AND INTEGRATION

The knowledge-directed approach to image interpretation, popular in th e 1980's, sought to identify objects in unconstrained two-dimensional (2-D) images and to determine the three-dimensional (3-D) relationship s between these objects and the camera by applying large amounts of ob ject- and domain-specific knowledge to the interpretation problem. Amo ng the primary issues faced by these systems were variations among ins tances of an object class and differences in how object classes were d efined in terms of shape, color, function, texture, size, and/or subst ructures. This paper argues that knowledge-directed vision systems typ ically failed for two reasons. The first is that the low- and mid-leve l vision procedures that were relied upon to perform the basic tasks o f vision were too immature at the time to support the ambitions interp retation goals of these systems. This problem, we conjecture, has been largely solved by recent advances in the field of 3-D computer vision , particularly in stereo and shape reconstruction from multiple views. The other impediment was that the control problem for vision procedur es was never properly addressed as an independent problem. This paper reviews the issues confronted by knowledge-directed vision systems, an d concludes that inadequate vision procedures and the lack of a contro l formalism blocked their further development. We then briefly introdu ce several new projects which, although still in the early stage of de velopment, are addressing the complex control issues that continue to obstruct the development of robust knowledge-directed vision systems.