Deep Neural Network for 3D Point Cloud Completion with Multistage Loss Function

3D shape completion by deep neural networks has been arousing increasing interest among research community. In this paper, A novel neural network architecture with multistage loss function that directly works on point clouds is proposed to map partial input point clouds to complete point clouds. Specifically, our network architecture not only works like an autoencoder that preserves the partial input. but also learns the global feature and fills the missing region. Experiments demonstrate the effectiveness of our approach on producing dense complete output point clouds with realistic structures in missing areas.