基于自蒸馏框架的点云分类及其鲁棒性研究

doi:10.19678/j.issn.1000-3428.0068451

摘要/Abstract

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。

关键词: 点云数据, 点云分类, 自蒸馏, 数据增强, 腐败鲁棒性

Abstract:

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.

Key words: point cloud data, point cloud classification, self-distillation, data enhancement, corruption robustness

李维刚, 厉许昌, 田志强, 李金灵. 基于自蒸馏框架的点云分类及其鲁棒性研究[J]. 计算机工程, 2024, 50(9): 72-81.

LI Weigang, LI Xuchang, TIAN Zhiqiang, LI Jinling. Research on Point Cloud Classification and Its Robustness Based on Self-Distillation Framework[J]. Computer Engineering, 2024, 50(9): 72-81.

https://www.ecice06.com/CN/Y2024/V50/I9/72

图/表 8

图1 结合基准网络的PointSD框架

Fig.1 PointSD framework combined with benchmark network

图2 TND数据增强方法流程

Fig.2 Procedure of TND data enhancement method

图3 ModelNet40-C数据集各腐败类型变换结果

Fig.3 Transformation results of each corruption type on the ModelNet40-C dataset

图4 两种网络应用多种方法在不同腐败类型上的测试结果雷达图

Fig.4 Radar chart of test results on different corruption types for two networks applying multiple methods

参考文献 24

1	李朝, 兰海, 魏宪. 基于注意力的毫米波-激光雷达融合目标检测. 计算机应用, 2021, 41(7): 2137- 2144. URL
	LI C, LAN H, WEI X. Attention-based object detection with millimeter wave radar-lidar fusion. Journal of Computer Applications, 2021, 41(7): 2137- 2144. URL
2	李维刚, 陈婷, 田志强. 基于孪生自适应图卷积算法的点云分类与分割. 计算机应用, 2023, 43(11): 3396- 3402. URL
	LI W G, CHEN T, TIAN Z Q. Point cloud classification and segmentation based on siamese adaptive graph convolution algorithm. Journal of Computer Applications, 2023, 43(11): 3396- 3402. URL
3	逄晨曦, 李文辉. 基于注意力改进的自适应空间特征融合目标检测算法. 吉林大学学报(理学版), 2023, 61(3): 557- 566. URL
	PANG C X, LI W H. Adaptive spatial feature fusion object detection algorithm based on attention improvement. Journal of Jilin University(Science Edition), 2023, 61(3): 557- 566. URL
4	HILDEBRAND J, SCHULZ S, RICHTER R, et al. Simulating LiDAR to create training data for machine learning on 3D point clouds[EB/OL]. [2023-08-14]. https://www.semanticscholar.org/paper/SIMULATING-LIDAR-TO-CREATE-TRAINING-DATA-FOR-ON-3D-Hildebrand-Schulz/26c0f93658737bb749ae336eeeb65b162cd4a438.
5	CHARLES R Q, HAO S, MO K C, et al. PointNet: deep learning on point sets for 3D classification and segmentation[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Washington D. C., USA: IEEE Press, 2017: 77-85.
6	QI C R, YI L, SU H, et al. PointNet++: deep hierarchical feature learning on point sets in a metric space[C]//Proceedings of the 2017 International Conference on Neural Information Processing Systems. Cambridge, USA: MIT Press, 2017: 5105-5114.
7	RAN H X, LIU J, WANG C J. Surface representation for point clouds[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Washington D. C., USA: IEEE Press, 2022: 18920-18930.
8	LEE D, LEE J. Regularization strategy for point cloud via rigidly mixed sample[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Washington D. C., USA: IEEE Press, 2021: 15895-15904.
9	CHEN Y L, HU V T, GAVVES E, et al. PointMixup: augmentation for point clouds. Berlin, Germany: Springer International Publishing, 2020.
10	ZHANG J L, CHEN L J, OUYANG B, et al. PointCutMix: regularization strategy for point cloud classification. Neurocomputing, 2022, 505, 58- 67. doi: 10.1016/j.neucom.2022.07.049
11	XU T B, LIU C L. Data-distortion guided self-distillation for deep neural networks[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2019, 33: 5565-5572.
12	ZHAO B R, CUI Q, SONG R J, et al. Decoupled knowledge distillation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Washington D. C., USA: IEEE Press, 2022: 11943-11952.
13	SUN J C, ZHANG Q Z, KAILKHURA B, et al. Benchmarking robustness of 3D point cloud recognition against common corruptions[EB/OL]. [2023-08-14]. http://arxiv.org/abs/2201.12296.
14	UY M A, PHAM Q H, HUA B S, et al. Revisiting point cloud classification: a new benchmark dataset and classification model on real-world data[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Washington D. C., USA: IEEE Press, 2019: 1588-1597.
15	WU Z R, SONG S R, KHOSLA A, et al. 3D ShapeNets: a deep representation for volumetric shapes[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Washington D. C., USA: IEEE Press, 2015: 1912-1920.
16	PASZKE A, GROSS S, CHINTALA S, et al. Automatic differentiation in PyTorch[EB/OL]. [2023-08-14]. https://openreview.net/forum?id=BJJsrmfCZ.
17	WANG Y, SUN Y B, LIU Z W, et al. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics, 2019, 38(5): 1- 12.
18	LI Y, BU R, SUN M, et al. PointCNN: convolution on X-transformed points[EB/OL]. [2023-08-14]. https://arxiv.org/abs/1801.07791.
19	CATTS H W. The simple view of reading: advancements and false impressions. Remedial and Special Education, 2018, 39(5): 317- 323. doi: 10.1177/0741932518767563
20	CHENG S L, CHEN X W, HE X W, et al. PRA-Net: point relation-aware network for 3D point cloud analysis. IEEE Transactions on Image Processing, 2021, 30, 4436- 4448. doi: 10.1109/TIP.2021.3072214
21	BERG A, OSKARSSON M, O'CONNOR M. Points to patches: enabling the use of self-attention for 3D shape recognition[C]//Proceedings of the 26th International Conference on Pattern Recognition (ICPR). Washington D. C., USA: IEEE Press, 2022: 528-534.
22	MA X, QIN C, YOU H X, et al. Rethinking network design and local geometry in point cloud: a simple residual MLP framework[EB/OL]. [2023-08-14]. http://arxiv.org/abs/2202.07123.
23	WANG C Y, HAN X F, XIAO G Q. PointCMT: an MLP-Transformer network for contrastive learning of point representation[C]//Proceedings of the International Joint Conference on Neural Networks (IJCNN). Washington D. C., USA: IEEE Press, 2023: 1-6.
24	QIAN G C, LI Y C, PENG H W, et al. PointNeXt: revisiting PointNet++ with improved training and scaling strategies[EB/OL]. [2023-08-14]. http://arxiv.org/abs/2206.04670.

[1]	张玉鑫, 张雷, 欧冬秀. 面向磁浮轨道的多源点云数据的混合滤波方法[J]. 计算机工程, 2024, 50(9): 54-62.
[2]	蔡俊民, 梁正友, 孙宇, 陈子奥. 基于可变形三维图卷积的轻量级点云分类研究[J]. 计算机工程, 2024, 50(9): 255-265.
[3]	张溢文, 蔡满春, 陈咏豪, 朱懿, 姚利峰. 融合空间特征的多尺度深度伪造检测方法[J]. 计算机工程, 2024, 50(7): 240-250.
[4]	林芷薇, 杨祖元, 王斯秋, 杨超. 基于多尺度线性全局注意力的运动员检测算法[J]. 计算机工程, 2024, 50(7): 352-359.
[5]	张宝鑫, 杨丹, 聂铁铮, 寇月. 基于自监督的多视角图协同过滤推荐方法[J]. 计算机工程, 2024, 50(5): 100-110.
[6]	宫阿娟, 潘天荣. 多病种眼底疾病诊断的深度学习策略讨论[J]. 计算机工程, 2024, 50(5): 363-372.
[7]	侯钰涛, 阿布都克力木·阿布力孜, 史亚庆, 马依拉木·木斯得克, 哈里旦木·阿布都克里木. 面向"一带一路"的低资源语言机器翻译研究[J]. 计算机工程, 2024, 50(4): 332-341.
[8]	安峰民, 张冰冰, 董微, 张建新. 面向视频行为识别深度模型的数据预处理方法[J]. 计算机工程, 2024, 50(2): 281-287.
[9]	江雨燕, 陶承凤, 李平. 数据增强和自适应自步学习的深度子空间聚类算法[J]. 计算机工程, 2023, 49(8): 96-103, 110.
[10]	王禹博, 陈利锋, 许卫霞. 结合多解码器与两阶段通道选择的异常检测方法[J]. 计算机工程, 2023, 49(3): 37-48.
[11]	席荣康, 蔡满春, 芦天亮. 基于数据增强与流数据处理的Tor流量分析模型[J]. 计算机工程, 2023, 49(3): 177-184.
[12]	曹健, 陈怡梅, 李海生, 蔡强. 基于深度学习的道路小目标检测综述[J]. 计算机工程, 2023, 49(10): 1-12.
[13]	孙伟, 常鹏帅, 戴亮, 张小瑞, 陈旋, 代广昭. 基于注意力引导数据增强的车型识别[J]. 计算机工程, 2022, 48(7): 300-306.
[14]	曹瑞阳, 郭佑民, 牛满宇. 基于最大最小距离的多中心数据综合增强方法[J]. 计算机工程, 2022, 48(6): 174-181.
[15]	佘朝阳, 严馨, 徐广义, 陈玮, 邓忠莹. 融合数据增强与半监督学习的药物不良反应检测[J]. 计算机工程, 2022, 48(6): 314-320.

选择文件类型/文献管理软件名称

选择包含的内容