Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (22): 202-212.DOI: 10.3778/j.issn.1002-8331.2301-0188

• Graphics and Image Processing • Previous Articles     Next Articles

MF-Net:Multi-Scale Feature Point Cloud Completion Network Combined with Residual Network

QIU Yunfei, ZHAO Jing, FANG Li   

  1. 1.School of Software, Liaoning Technical University, Huludao, Liaoning 125100, China
    2.Laboratory of Remote Sensing and Information Engineering, Quanzhou Institute of Equipment Manufacturing, Haixi Institute, Chinese Academy of Sciences, Quanzhou, Fujian 362216, China
  • Online:2023-11-15 Published:2023-11-15

MF-Net:结合残差网络的多尺度特征点云补全网络

邱云飞,赵静,方立   

  1. 1.辽宁工程技术大学 软件学院,辽宁 葫芦岛 125100
    2.中国科学院海西研究院泉州装备制造研究中心 遥感信息工程实验室,福建 泉州 362216

Abstract: Aiming at the problem of semantic information loss caused by the current point cloud completion network focusing only on global features, a point cloud completion network based on multi-scale feature extraction of residual network is proposed. The network adopts an end-to-end idea. Firstly, to avoid the problem of incomplete single feature, the original input is sampled into three different scale point clouds. Then, the global features of the low-resolution point cloud extracted by different methods and the local features of the original point cloud are fused recursively in a cascade way to form feature vectors, and are put into the fully connected network to realize the prediction of coarse point cloud. Next, the splicing original point cloud and coarse point cloud are sent into the fine reconstruction unit, and then the attention mechanism is integrated into the fine reconstruction unit, and the residual network is used to complete from rough to fine. Finally, the joint loss function among coarse point clouds, dense point clouds and ground truth point clouds is calculated to improve the completion performance. Experiments on ShapeNet and KITTI data sets show that the proposed method has good completion effect on the incomplete point cloud, both in qualitative and quantitative comparison, and also show its generalization ability.

Key words: 3D point cloud, point cloud completion, multi-scale feature, attention mechanism, residual network

摘要: 针对目前点云补全网络只关注全局特征造成的语义信息丢失问题,提出了一个基于残差网络的多尺度特征提取的点云补全网络。网络采用端到端的思想,为避免单一特征不全面问题,将原始输入采样为三种不同尺度的点云;利用级联方式递归式融合不同方法提取的低分辨率点云的全局特征和原始点云的局部特征,形成特征向量并输入全连接网络,实现粗点云的预测;将拼接后的原始点云和粗点云送入精细重构单元,再在精细重构单元中融合注意力机制并利用残差网络进行由粗略到精细的补全;通过计算粗点云、稠密点云与真实点云之间的联合损失函数以提高补全性能。在ShapeNet数据集和KITTI数据集上的实验证明,无论是定性比较还是定量比较,提出的方法对残缺点云均具有较好的补全效果,同时也体现了该方法具有泛化能力。

关键词: 3D点云, 点云补全, 多尺度特征, 注意力机制, 残差网络