An ATN Model for 3-D Recognition of Solids in Single Images

Abstract

Augmented Transition Network (ATN) models have been developed in the field of natural language understanding. In this paper, an ATN model for image analysis is discussed; it is described for the 3-D recognition of randomly oriented sol ids with known geometry. Due to its perspicuity and generative power, ATN aopears to be well suited to incorporate the geometrical knowledge at hand. Mechanisms for relating 2-D images to 3-D objects are handled by using the inverse representation: calculating 2-D model instantiations of a 3-D geometric model. The ATN controls an analysis-by-synthesis search procedure that is based on hypothesis generation and verification. A state-space search technique is used. Identical sub spaces are recognized, the same is true for subspaces containing similar hypothetical model instantiations. The procedure copes with degraded image data. Introducing general semantic knowledge and knowledge of the domain fits harmonically into the ATN framework.