PU-FPG: Point cloud upsampling via form preserving graph convolutional networks

Point cloud upsampling can improve the resolutions of point clouds and maintain the forms of point clouds, which has attracted more and more attention in recent years. However, upsampling networks sometimes generate point clouds with unclear contours and deficient topological structures, i.e., the problem of insufficient form fidelity of upsampled point clouds. This paper focuses on the above problem. Firstly, we manage to find the points located at contours or sparse positions of point clouds, i.e., the form describers, and make them multiply correctly. To this end, 3 statistics of points, i.e., local coordinate difference, local normal difference and describing index, are designed to estimate the form describers of the point clouds and rectify the feature aggregation of them with reliable neighboring features. Secondly, we divide points into disjoint levels according to the above statistics and apply K nearest neighbors algorithm to the points of different levels respectively to build an accurate graph. Finally, cascaded networks and graph information are fused and added to the feature aggregation so that the network can learn the topology of objects deeply, enhancing the perception of model toward graph information. Our upsampling model PU-FPG is obtained by combining these 3 parts with upsampling networks. We conduct abundant experiments on PU1K dataset and Semantic3D dataset, comparing the upsampling effects of PU-FPG and previous works in multiple metrics. Compared with the baseline model, the Chamfer distance, the Hausdorff distance and the point-to-surface distance of PU-FPG are reduced by 0.159 x 10(-3), 2.892 x 10(-3) and 0.852 x 10(-3), respectively. This shows that PU-FPG can improve the form fidelity and raise the quality of upsampled point clouds effectively. Our code is publicly available at https://github.com/SATURN2021/PU-FPG.