A practical view based 3D object recognition system

This paper discusses the application of a self-organizing neural network to the problem of constructing adequate, 2D appearance-based models of 3D objects. The problem of basing recognition on such models is characterised as that of approximating the multivariate function mapping shapes of an object to its identity. The domain of this function is described as a set of hypersurfaces in the space of possible shape representations. A novel representational scheme is introduced based on storing values of geometric relationships between pairs of shape features in a histogram. A network architecture capable of constructing function approximations based on such shape representations is presented. Finally, methods for assessing the accuracy of approximation are detailed