PU-LNGCN: Multi-scale Design of Point Cloud Upsampling Using Graph Convolutional Networks

Abstract

Learning and analyzing 3D point clouds with deep neural networks is challenging due to the irregular and unordered nature of the data. Therefore, as the task of converting sparse and unordered point clouds into dense and complete ones, point cloud upsampling has attracted extensive attention. In this paper, we specially design a structure-based graph convolutional network called Local Neighborhood Graph Convolutional Network (LNGCN) to fully exploit structural information of graph. We introduce the proposed LNGCN and further propose a novel multi-scale feature extraction block called Multiscale LNGCN to encode rich information of point cloud data at different granular levels. By aggregating features at multiple scales, this feature extractor enables further performance improvement in the final upsampled point clouds. We combine the Multiscale LNGCN block into current point upsampling pipelines and propose a new architecture called PU-LNGCN. By using extensive quantitative and qualitative experiments, we show that PU-LNGCN can handle noisy and non-uniformly distributed point clouds as well as real scanned data by LiDAR sensors very well. PU-LNGCN outperforms previous methods and achieves state-of-the-art performance.