Model Based (intermediate-level) Computer Vision

A system for computer vision is presented, which is based on two-dimensional prototypes, and which uses a hierarchy of features for mapping purposes. More specifically, one is dealing with scenes composed of planar faced, convex objects. Extensions to the general planar faced case are discussed. The visual input is provided by a TV-camera, and the problem is to interpret that input by computer, as a projection of a three-dimensional scene. The system proposed and demonstrated in this paper uses perspectively consistent two-dimensional models (prototypes) of views of three-dimensional objects, and interpretations of scene-representations are based on the establishment of mapping relationships from conglomerates of scene-elements (line-constellations) to prototypes templates. The prototypes are learned by the program through analysis of - and generalization on - ideal instances. (Modified author abstract).