LV-GCNN: A lossless voxelization integrated graph convolutional neural network for surface reconstruction from point clouds

Abstract

Surface reconstruction from a 3D point cloud is a long-standing problem in the field of computer graphics and vision, especially for point clouds that are sparse and noisy. To solve this problem, a novel method, termed LV-GCNN, that combines lossless voxelization and a graph convolutional neural network is proposed in this paper. Firstly, a lossless voxelization method for point clouds, which achieves lossless conversion from disordered point clouds to ordered volumetric maps, is proposed. Secondly, with the generated voxel, the feature pyramid is extracted through a 3D convolutional neural network. Thirdly, based on the similarity between the graph and surface mesh, the graph neural network is initialized as a unit spherical mesh. The coordinate relationship between the graph vertices and feature voxel units is used to match the corresponding hierarchical features for graph vertices. The initial spherical mesh is progressively deformed and upsampled via graph convolution and graph unpooling to achieve coarse-to-fine optimization of the surface mesh. Finally, because the deformation-based approach cannot reconstruct objects with a genus greater than zero, a genus optimization method is designed. Experiments show that the surface meshes generated by the LV-GCNN are comparable to or better than the results of state-of-the-art methods under several evaluation criteria. In addition, the reconstruction effect under different situations is discussed. Ablation experiments show the importance of several applied modules in LV-GCNN. Extra experiments show that the proposed method can achieve impressive results for point sets with diverse densities or that contain different levels of noise. The LV-GCNN results outperform other methods when the input point cloud is exceptionally sparse (256 points) or contains Gaussian noise with a standard deviation of 0.05. The chamfer distance (CD), Hausdorff distance (HD), voxel difference (VD), and depth difference (DD) of the reconstruction results of the LV-GCNN are 0.0494, 0.1457, 0.7390, and 18,402 when 256 points are used as input and are 0.0654, 0.1582, 0.6735, and 21,150 when points with noise with a standard deviation of 0.05 are used as input.

Keywords

• Point cloud
• Surface reconstruction
• Lossless voxelization
• Feature pyramid
• Graph convolutional neural network