Encoding 3D structural information using multiple self-organizing feature maps

This paper describes a system which encodes a free-form three-dimensional (3D) object using Artificial Neural Networks. The types of surface shapes which the system is able to handle include not only pre-defined surfaces such as simple piecewise quadric surfaces but also more complex free-form surfaces. The system utilizes two Self-Organizing Maps to encode surface parts and their geometrical relationships. Authors demonstrated the use of this encoding technique on "simple" 3D free-form object recognition systems [M. Takatsuka, R.A. Jarvis, Hierarchical neural networks for learning 3D objects from range images, Journal of Electronic Imaging 7 (1) (1998) 16-28]. This paper discusses the design and mechanism of the Multiple SOFMs for encoding 3D information in greater detail including an application to face ("complex" 3D free-form object) recognition. (C) 2001 Elsevier Science B.V. All rights reserved.