基于自蒸馏框架的点云分类及其鲁棒性研究  

Research on Point Cloud Classification and Its Robustness Based on Self-Distillation Framework

作　　者：李维刚[1,2] 厉许昌 田志强 李金灵 LI Weigang;LI Xuchang;TIAN Zhiqiang;LI Jinling(College of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China;Engineering Research Center for Metallurgical Automation and Measurement Technology,Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China)

机构地区：[1]武汉科技大学信息科学与工程学院,武汉430081 [2]武汉科技大学冶金自动化与检测技术教育部工程研究中心,武汉430081

出　　处：《计算机工程》2024年第9期72-81,共10页Computer Engineering

基　　金：湖北省揭榜制科技项目(2020BED003)。

摘　　要：与2D图像数据集相比,3D点云数据集的规模较小且表征性较差,容易导致神经网络出现过拟合和泛化能力差的问题。为此,提出一种点云自蒸馏(PointSD)框架,通过对表征形式不同的数据样本进行学习,使网络提取到原始点云数据中的更多特征信息,实现样本之间的知识交互,在不增加额外计算负荷的情况下提升网络的泛化能力,适用于不同规模的分类网络模型。基于该框架提出一种点云抗腐败训练方法TND-PointSD,解决了当前点云训练方法抗腐败能力不足的问题。实验结果表明:在ScanObjectNN数据集上,应用PointSD框架的PointNet++和RepSurf-U 2X基准网络的平均准确率(MA)相比于应用标准训练(ST)方法提高了8.22和4.86个百分点;在ModelNet40-C数据集上,在15种腐败类型上分类网络的平均整体准确率(MOA)均有所提升,证明了TND-PointSD方法能够有效地增强网络模型的腐败鲁棒性。Compared to that of a Two-Dimensional(2D)image dataset,a Three-Dimensional(3D)point cloud dataset is smaller in scale and poorly represented,which easily leads to problems of overfitting and poor generalization ability of neural networks.Accordingly,a Point cloud Self-Distillation(PointSD)framework is proposed.This framework enables the network to extract more feature information from the original point cloud data by learning data samples with different representation forms,thus realizing the knowledge interaction between samples,improving the generalization capabilities of the network without increasing the additional computational load,and making the network suitable for classification network models of different scales.Based on this framework,a point cloud anti-corruption training method,TND-PointSD,is proposed,which solves the problem of the insufficient anti-corruption capabilities of the current point cloud training methods.Experimental results show that the Mean Accuracy(MA)of the PointNet++and RepSurf-U 2X benchmark networks using the PointSD framework are 8.22 and 4.86 percentage points higher respectively,than those of the Standard Training(ST)method on the ScanObjectNN dataset.In addition,the Mean Overall Accuracy(MOA)of the classification networks on the ModelNet40-C dataset is improved for 15 corruption types.The study thus shows that the TND-PointSD method can effectively enhance the corruption robustness of the network model.

关 键 词：点云数据 点云分类 自蒸馏 数据增强 腐败鲁棒性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]