计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (9): 272-282.DOI: 10.3778/j.issn.1002-8331.2212-0355

• 图形图像处理 • 上一篇    下一篇

双分支结构的多层级三维点云补全

邱云飞,王宜帆   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
  • 出版日期:2024-05-01 发布日期:2024-04-29

Multi-Level 3D Point Cloud Completion with Dual-Branch Structure

QIU Yunfei, WANG Yifan   

  1. College of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2024-05-01 Published:2024-04-29

摘要: 为了缓解现有点云补全方法在特征提取过程中很难平衡局部特征和全局特征的问题,提出了一种双分支结构的多层级点云补全算法。利用两个独立的分支网络分别提取出输入点云的局部特征信息和全局特征信息,再将两种特征信息进行拼接形成特征向量。使用五层联合感知机将特征向量映射成多个维度,进而提取多维特征信息并将其整合成最终特征向量。采用金字塔结构在256、512、1?024特征维度上对最终特征向量进行特征解码,预测三种不同分辨率的点云。引入鉴别器网络,通过联合训练鉴别器产生的对抗损失和分层重建点云产生的补全损失去优化网络。在ShapeNet数据集上进行实验,算法显著提升了点云补全精度,并且在缺失大面积点云时也能恢复出较为完善的物体形状。

关键词: 三维点云, 形状补全, 深度学习, 双分支结构, 鉴别器网络

Abstract: In order to alleviate the problem that the existing point cloud completion methods are difficult to balance local features and global features in the feature extraction process, this paper proposes a multi-level point cloud completion algorithm with double branch structure. Two independent branch networks are used to extract the local feature information and global feature information of the input point cloud respectively, and the two feature information are concatenated to form a feature vector. The five levels combinate perceptron is used to map the feature vector into multiple dimensions, and the multi-dimensional feature information is extracted and integrated into the final feature vector. Then, the pyramid structure is used to decode the final feature vector in 256, 512 and 1 024 feature dimensions, and the point clouds with three different resolutions are predicted. Finally, the discriminator network is introduced to optimize the network by jointly training the adversarial loss generated by the discriminator and the completion loss generated by the hierarchical reconstruction point cloud. Experiments on ShapeNet dataset show that the algorithm significantly improves the accuracy of point cloud completion. In addition, a relatively complete object shape can be recovered when a large area of point cloud is missing.

Key words: three-dimensional point cloud, shape completion, deep learning, dual-branch structure, discriminator network