Representing 3D shapes based on implicit surface functions learned from RBF neural networks

Abstract

We propose to represent the shape of 3D objects using a neural network classifier. The 3D shape is learned from a neural network, where Radial Basis Function (RBF) is applied as the activation function for each perceptron. The implicit functions derived from the neural network is a combination of radial basis functions, which can represent complex shapes. The use of RBF provides a rotation, translation and scaling invariant feature to represent the shape. We conduct experiments on a new prostate dataset and public datasets. Our testing results show that our neural network-based method can accurately represent various shapes.