In this paper, a new neural tree architecture whose nodes are generalized perceptrons without hidden layers is applied to segment range images into surface patches, according to the six models of differential geometry, e.g., peak, ridge, valley, saddle, pit and flat. A new learning scheme which improves upon the standard neural tree algorithms in terms of convergence is proposed. Splitting nodes are introduced into the neural tree architecture to divide the training set when the current perceptron node repeats the same classification of the parent node: such a strategy is able to assure in any case the convergence of the tree building process and to reduce misclassifications. Significant results on synthetic and real 3D range images are presented and compared with conventional approaches. (C) 1998 Elsevier Science B.V. All rights reserved.
