Constrained Constructive Solid Geometry a Unique Representation of Scenes
James A. D. W. Anderson, G. D. Sullivan, Keith D. Baker · 1988
Constraints are described for Constructive Solid Geometry which ensure that a scene composed of solids is described uniquely, up to a choice of the decomposition of compound solids into primitive ones. The constraints can be applied more easily to an Additive Constructive Solid Geometry which is also better suited to implementation on a parallel architecture. Constructive Solid Geometry (CSG) has been used in several model-based vision programs, most notably in ACRONYM1. This paper examines some of the problems that arise with CSG and proposes constraints that make CSG a unique representation of scenes. It is also suggested that Additive Constructive Solid Geometry (ACSG) is more useful for vision because it leads to simpler computation of connectivity, within and between objects, and can be implemented more readily on a parallel architecture. The motivation for examining the properties of a formal modelling system arises from what we shall call the strong thesis of model-based vision. The principal tenet is that all image features required to verify a model can be derived automatically from a model and knowledge of the optics of image formation. That is, models should provide complete knowledge of visual form which can, in principle, be used to solve any visual problem. The subsidiary tenet is that there is a two-way mapping between image locations and instantiated models. Thus, in principle, it is possible for analysis to proceed both top-down from instantiated models to the image and bottom-up from the image to instantiated models. This latter property underlies the definition of visual knowledge as "knowledge which can be brought into a two-way, spatial mapping with an image " (compare with Sloman's similar definition2). Thus, in model-based vision, models provide a justification for the particular image processing techniques used. They provide a deep knowledge of image processing and mediate between image processing and the rest of the system's knowledge. If models are to be used automatically by an intelligent vision program then it is important that the models and modelling processes are well formed and do not require human intervention.