An Unsupervised Detail-Preserving Point Cloud Completion Network Guided by Projection Views

Traditional supervised point cloud completion methods always require the complete point cloud data as a prior information, which lead to their poor generalization and low robustness. Meanwhile, the completion results generated by the existing unsupervised learning approaches often deviate from the input shapes, making it difficult to recover the fine details of the original shapes. Owing to the framework of generative adversarial network (GAN), an unsupervised detail-preserving point cloud completion network is proposed which is guided by the feature information of three projection views obtained from the underlying shape. The proposed network consists of the branch of point cloud shape repairing and the branch of projection image completion. Firstly, the point cloud shape repairing branch employs a tree-structured graph convolutional generator to create a coarse completion point cloud, aiming to recover its overall shape. The coarse completion result is then fed into a DGCNN network to extract its features. Secondly, the projection image completion branch projects the input model to obtain three projection views of the input point cloud. Next, an image generator based on cycle consistency is adopted to repair these projection images, and a ResNet-18 network is employed to extract features from the complete projection views. The feature distance between these aligned image features and the shape features extracted from the generated point cloud can thus be calculated, and also be added to the loss function of the discriminator which is employed to judge the truth or falsehood of the generated shape. Finally, network parameters of the point cloud generator can be optimized to learn the global shape and fine details. The proposed completion network is trained and tested on ShapeNet dataset for shape completion task and also validated on the KITTI and ModelNet40 datasets. Compared with the existing unsupervised completion networks, the average CD error of our proposed network is reduced by 11.0% to 41.0%, and the average F1-Score is improved by 0.8% to 14.0%, which can demonstrate its effectiveness for repairing shape structure of the input point cloud data and also recovering its shape details. In addition, our point completion network is robust to different degrees of data incompleteness or model noise, and also shows its generalization performance on unseen objects.