Deformable Point Cloud Recognition Using Intrinsic Function and Deep Learning

Recognizing 3D point cloud shapes under isometric deformation is an interesting and challenging issue in the geometry field. This paper proposes a novel feature learning approach by using both the model-based intrinsic descriptor and the deep learning technique. Instead of directly applying deep convolutional neural networks (CNN) on point clouds, we first represent the isometric deformation by using a set of local intrinsic functions to grasp the invariant properties of the shape. Then, an effective point CNN network is developed to learn the parameters and perform semantic feature learning in an end-to-end fashion to link the local and global information together for discriminative shape representation and classification. To reduce the computational costs of our CNN network, some simple operations, like downsampling and fusion, are applied to decrease the number of points and the intrinsic dimensions based on our average heat function. The experimental results on multiple standard benchmarks have demonstrated that our proposed algorithm can achieve very competitive results on both the accuracy rates and the computational efficiency.