A 3D shape classifier with neural network supervision

The task of 3D shape classification is to assign a set of unordered shapes into pre-tagged classes with class labels, and find the most suitable class for a newly given shape. In this paper, we present a 3D shape classifier approach based on supervision of the learning of point spatial distributions. In this classifier, we first extract the low-level features by characterizing the point spatial density distributions of 3D shapes, and train one feedforward neural network to learn these features by examples. The Konstanz shape database was chosen as the test database, and we grouped the classified objects into two sets, the training set and the test set, which each had an approximately equal number of shapes. We trained the network with the training set, and evaluated the accuracy rate on the test set. We also compared this classifier to k nearest neighbors classifier for 3D shapes. This approach can be used to classify 3D shapes and enhance the performance of the existent 3D shape retrieval methods.