Aspect graphs and their use in object recognition

Abstract

Previous researchers have described several different approaches to 3-D object recognition based on using an iterative technique to control the matching of features from the 2-D projection of a 3-D model to observed image features. The major problem encountered with such approaches is how to automatically choose starting parameter estimates in a manner which both avoids recognition errors due to local minima and is still reasonably efficient. This paper investigates the use of theaspect graph to address this problem. The basic idea is quite simple — an iterative solution is generated for each of a set of candidate aspects and the best of these is chosen as the recognized view. Two assumptions are required in order for this approach to be valid: (1) the iterative search for the correct candidate aspect must converge to the correct answer, and (2) the solution found for the correct aspect must be better than that found for any of the incorrect candidate aspects. In order to explore the validity of these assumptions, a simple aspect graph-based recognition system was implemented. Experiments were carried out using both real and simulated data. The results indicate that the underlying assumptions are generally valid, and that this approach has advantages over previous techniques which incorporated an iterative search.