Visual Recognition from Spatial Correspondence and Perceptual Organization

Abstract

Depth reconstruction from the two-dimensional image plays an important roJe in certain visual tasks and has been a major focus of of computer vision research. However, in this paper we argue that most instances of recognition in human and machine vision can best be performed without the preliminary reconstruction of depth. Three other mechanisms are described that can be used to bridge the gap between the two-dimensional image and knowledge of three-dimensional objects. First, a process of perceptual organization can be used to form groupings and structures in the image that are likely to be invariant over a wide range of viewpoints. Secondly, evidential reasoning can be used to combine evidence from these groupings and other sources of information to reduce the size of the search-space during model-based matching. Finally, a process of spatial correspondence can be used to bring the projections of three-dimensional models into direct correspondence with the image by solving for unknown viewpoint and model parameters. These methods have been combined in an experimental computer vision system named SCERPO. This system has demonstrated the use of these methods for the recognition of objects from unknown viewpoints in single gray-scale images.