Abstract:

Aiming at the problem that the graph learning-based methods in point cloud completion are easy to lead to the smoothing phenomenon of point features when the number of network layers is deepened， thereby causing local geometric detail loss， a point cloud completion method based on smooth-sharpen graph convolution was proposed， so as to improve the completion quality of irregular point clouds. Firstly， the SA （Set Abstraction） layer in PointNet++ was cross-stacked with the point Transformer module to extract multi-level features and global shape features of the incomplete point cloud. Secondly， the graph structure was constructed by generating the sub-point features through point-wise splitting， and the local similarity and coordinate difference information between point clouds were captured by using smooth graph convolution and sharpen graph convolution， respectively， so as to retain key geometric details and achieve coarse completion. Finally， global shape features and the results of coarse completion were fused in the refinement module to generate high-resolution point clouds gradually. Experimental results show that the proposed method outperforms comparison methods ODGNet （Orthogonal Dictionary Guided shape completion Network） and CRA-PCN （Cross-Resolution Aggregation Point Completion Network） on the ShapeNet and PCN （Point Completion Network） datasets. Specifically， the proposed method has the CD-T （Chamfer Distance-TopNet）， CD-P （Chamfer Distance-Point completion network） and F1-Score increased by at least 0.220， 0.285 and 0.009 on the PCN dataset， respectively， and the CD-P and F1-Score increased by at least 0.021 and 0.003 on the ShapeNet extended dataset， indicating that the proposed method has more advantages in geometric structure recovery. In addition， the completion results on the real-world KITTI （Karlsruhe Institute of Technology Toyota Technological Institute at Chicago） dataset further show that the proposed method can reconstruct the vehicle contour more completely， avoid blurring and artifacts effectively， and improve the effect of point cloud completion.

Key words: deep learning, point cloud completion, graph convolution, feature extraction, sharpen graph convolution, Graph Neural Network (GNN)

摘要：

针对点云补全中基于图学习的方法在网络层数加深时易出现点特征平滑现象所造成的局部几何细节丢失问题，提出一种基于平滑-锐化图卷积的点云补全方法，旨在提升不规则点云的补全质量。首先，将PointNet++中的SA（Set Abstraction）层与点Transformer模块交叉堆叠，以提取不完整点云的多层次特征和全局形状特征；其次，通过逐点分裂生成子点特征并以此构建图结构，同时利用平滑图卷积和锐化图卷积分别捕获点云间的局部相似性和坐标差异信息，从而保留关键几何细节并实现粗补全；最后，在细化模块中融合全局形状特征与粗补全的结果，逐步生成高分辨率点云。实验结果表明，本文方法在ShapeNet和PCN（Point Completion Network）数据集上均优于对比方法ODGNet（Orthogonal Dictionary Guided shape completion Network）与CRA-PCN（Cross-Resolution Aggregation PCN）。在PCN数据集上，本文方法的CD-T（Chamfer Distance-TopNet）、CD-P（Chamfer Distance-Point completion network）和F1值分别至少提升0.220、0.285和0.009；在ShapeNet扩展数据集上，本文方法的CD-P和F1值分别至少提升0.021和0.003，表明本文方法在几何结构恢复方面更具优势。此外，在真实场景KITTI（Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago）数据集上的补全结果进一步表明，本文方法能更完整地重建车辆轮廓，从而有效地避免模糊与伪影，并提升点云补全的效果。

关键词: 深度学习, 点云补全, 图卷积, 特征提取, 锐化图卷积, 图神经网络