3D object recognition by neural trees

In this paper, a two stage method for 3D object recognition from range images is presented. The first stage extracts local surface features from the input range images. These features are used in the second stage to group image pixels into different surface patches according to the six surface classes proposed by the differential geometry. A neural tree architecture whose nodes are perceptrons without hidden layers and with sigmoidal activation functions is used. A new strategy is proposed to split the training set when it is not linearly separable in order to assure the convergence of the tree learning process. This method has been successfully applied to a large number of synthetic and real images, some of which are presented in the result section.