SmoothNet: Smooth Point Cloud Up-sampling

The 3D point cloud collected by LIDAR (Light Detection and Ranging) is usually sparse. However, for example, in the field of digitization of cultural relics, a denser, more uniform, and smoother point cloud is often required when analyzing and displaying models. Therefore, this paper proposes a novel deep learning structure for the field of point cloud up-sampling. Concretely, we use multi-layer GCN to extract point cloud features, in addition, introduce shuffle module to achieve multi-feature expansion. Besides, a new smoothing loss function is designed to enhance the local smoothness of point clouds. Under the PU600 dataset, our method outperforms other existing methods and performs better on building or cultural relic point clouds.