面向三维点云的平衡泛化和特化的细粒度分类网络

doi:10.11996/JG.j.2095-302X.2025030602

图学学报 ›› 2025, Vol. 46 ›› Issue (3): 602-613.DOI: 10.11996/JG.j.2095-302X.2025030602

• 计算机图形学与虚拟现实 • 上一篇下一篇

面向三维点云的平衡泛化和特化的细粒度分类网络

刘鸿硕¹(), 白静¹^,²^,³(), 晏浩¹, 林淦¹

1.北方民族大学计算机科学与工程学院，宁夏银川 750021
2.国家民委图像图形智能处理实验室，宁夏银川 750021
3.六盘山实验室，宁夏银川 750021

收稿日期:2024-08-22 接受日期:2025-01-10 出版日期:2025-06-30 发布日期:2025-06-13
通讯作者:白静(1982-)，女，教授，博士。主要研究方向为CAD&CG、计算机视觉和机器学习等。E-mail：baijing@nun.edu.cn
第一作者:刘鸿硕(2001-)，男，硕士研究生。主要研究方向为三维点云细粒度分类。E-mail：1709671541@qq.com
基金资助:
国家自然科学基金(62162001);宁夏自然科学基金(2022AAC02041);宁夏优秀人才支持计划

BGS-Net: fine-grained classification networks with balanced generalization and specialization for 3D point clouds

LIU Hongshuo¹(), BAI Jing¹^,²^,³(), YAN Hao¹, LIN Gan¹

1. School of Computer Science and Engineering, North Minzu University, Yinchuan Ningxia 750021, China
2. The Key Laboratory of Images and Graphics Intelligent Processing of State Ethnic Affairs Commission, Yinchuan Ningxia 750021, China
3. Liupanshan Laboratory, Yinchuan Ningxia 750021, China

Received:2024-08-22 Accepted:2025-01-10 Published:2025-06-30 Online:2025-06-13
Contact: BAI Jing (1982-), professor, Ph.D. Her main research interests cover CAD&CG, computer vision, and machine learning, etc. E-mail：baijing@nun.edu.cn
First author：LIU Hongshuo (2001-), master student. His main research interest covers 3D point cloud fine-grained classification. E-mail：1709671541@qq.com
Supported by:
National Natural Science Foundation of China under Grant(62162001);Natural Science Foundation of Ningxia of China under Grant(2022AAC02041);Ningxia Excellent Talent Program

摘要/Abstract

摘要：

随着三维理解和计算机视觉技术的快速发展，点云数据因其精确的几何描述和丰富的空间信息在研究中备受关注。特别是在智能交通等实际应用中，利用点云数据识别不同型号的车辆需要准确分类细微差异和特征，因此开展点云细粒度分类尤为重要。现有方法多侧重设计任务特化的网络，通过提取点云的局部判别性特征提升分类性能，但往往忽视模型的泛化能力，导致在不同场景或未见类别中的表现下降。此外，复杂环境下的噪声、遮挡或数据分布变化亦会削弱模型的特化能力。为解决上述问题，提出了一种两阶段点云细粒度分类网络BGS-Net，旨在平衡模型的泛化性与特化性。在第一阶段，采用掩码蒸馏自监督学习结合耦合掩码方法，为2个学生模型分配互补掩码，引导其从教师模型中学习独立的特征表示，从而增强泛化能力。第二阶段设计了平衡泛化与特化的训练策略，通过冻结一个编码器以保留通用特征，同时调优另一个编码器以提取点云的局部判别性特征，实现任务特化。实验结果表明，BGS-Net在细粒度分类、元类别分类、少样本分类及真实场景分类任务中均表现优异，显著优于现有方法，验证了在保持高泛化能力的同时实现任务特化的有效性。该方法为点云细粒度分类提供了新的研究思路，提升了模型在实际应用中的适用性和鲁棒性。

关键词: 点云理解, 自监督学习, 耦合掩码, 特征学习, 模型泛化

Abstract:

With the rapid advancement of 3D understanding and computer vision technologies, point cloud data have gained significant attention for their precise geometry and rich spatial information. In applications such as intelligent transportation, accurately classifying subtle differences in vehicle models is crucial, rendering fine-grained point cloud classification essential. However, existing methods focus heavily on task-specific networks that enhance classification by extracting local discriminative features, often neglecting the model’s generalization abilities. This resulted in decreased performance in diverse scenarios and unseen categories, especially in environments with noise, occlusions, or data distribution changes. To address these challenges, the balanced generalization specialization network (BGS-Net) was proposed as a two-stage framework that balanced generalization and specialization. In the first stage, BGS-Net employed mask distillation self-supervised learning with coupled masks to guide two student models in learning independent feature representations from a teacher model, thereby enhancing generalization. In the second stage, a balanced training strategy was implemented by freezing one encoder to preserve general features while fine-tuning the other encoder to extract locally discriminative features for task specialization. Experimental results demonstrated that BGS-Net significantly outperformed existing methods in fine-grained, meta-category, few-shot, and real-world classification tasks, thereby confirming its effectiveness in maintaining high generalization while achieving task specialization. This approach enhanced the applicability and robustness of point cloud classification in practical applications.

Key words: point cloud understanding, self-supervised learning, coupling mask, feature learning, model generalization

中图分类号:

TP391
TP18

刘鸿硕, 白静, 晏浩, 林淦. 面向三维点云的平衡泛化和特化的细粒度分类网络[J]. 图学学报, 2025, 46(3): 602-613.

LIU Hongshuo, BAI Jing, YAN Hao, LIN Gan. BGS-Net: fine-grained classification networks with balanced generalization and specialization for 3D point clouds[J]. Journal of Graphics, 2025, 46(3): 602-613.

图/表 10

参考文献 50

[1]	白静, 邵会会, 姬卉, 等. 面向三维点云的端到端细粒度分类网络[J]. 计算机辅助设计与图形学学报, 2023, 35(1): 128-134.
	BAI J, SHAO H H, JI H, et al. An end-to-end fine-grained classification network for 3D point clouds[J]. Journal of Computer-Aided Design & Computer Graphics, 2023, 35(1): 128-134 (in Chinese).
[2]	WU R S, BAI J, LI W J, et al. DCNet: exploring fine-grained vision classification for 3D point clouds[J]. The Visual Computer, 2024, 40(2): 781-797.
[3]	SHAO H H, BAI J, WU R S, et al. FGPNet: a weakly supervised fine-grained 3D point clouds classification network[J]. Pattern Recognition, 2023, 139: 109509.
[4]	LI Y, LIU Z, CHANG X J, et al. Diversity-boosted generalization-specialization balancing for zero-shot learning[J]. IEEE Transactions on Multimedia, 2023, 25: 8372-8382.
[5]	JIANG S W, XU T F, GUO J, et al. Tree-CNN: from generalization to specialization[J]. EURASIP Journal on Wireless Communications and Networking, 2018, 2018: 216.
[6]	FINN C, ABBEEL P, LEVINE S. Model-agnostic meta-learning for fast adaptation of deep networks[EB/OL]. [2024-06-22]https://dblp.uni-trier.de/db/conf/icml/icml2017.html#FinnAL17.
[7]	WANG H C, LIU Q, YUE X Y, et al. Unsupervised point cloud pre-training via occlusion completion[C]// The IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 9762-9772.
[8]	YU X M, TANG L L, RAO Y M, et al. Point-BERT: pre-training 3D point cloud transformers with masked point modeling[C]// The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 19291-19300.
[9]	PANG Y T, WANG W X, TAY F E H, et al. Masked autoencoders for point cloud self-supervised learning[C]// The 17th European Conference on Computer Vision. Cham: Springer, 2022: 604-621.
[10]	ZHANG Y B, LIN J H, HE C H, et al. Masked surfel prediction for self-supervised point cloud learning[EB/OL]. [2024-06-22]https://arxiv.org/abs/2207.03111.
[11]	ZEID K A, SCHULT J, HERMANS A, et al. Point2Vec for self-supervised representation learning on point clouds[C]// The 45th DAGM German Conference on Pattern Recognition. Cham: Springer, 2024: 131-146.
[12]	XIE Q X, LUONG M T, HOVY E, et al. Self-training with noisy student improves ImageNet classification[C]// The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 10684-10695.
[13]	QI C R, SU H, MO K, et al. Pointnet: deep learning on point sets for 3d classification and segmentation[C]// IEEE conference on computer vision and pattern recognition. New York: IEEE Press, 2017: 652-660.
[14]	QI C R, YI L, SU H, et al. PointNet++: deep hierarchical feature learning on point sets in a metric space[C]// The 31st International Conference on Neural Information Processing Systems. New York: ACM, 2017:, 5105-5114.
[15]	ZHAO H S, JIANG L, JIA J Y, et al. Point transformer[C]// The IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 16239-16248.
[16]	GUO M H, CAI J X, LIU Z N, et al. PCT: point cloud transformer[J]. Computational Visual Media, 2021, 7(2): 187-199.
[17]	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]. [2024-06-22]https://dblp.uni-trier.de/db/conf/iclr/iclr2022.html#MaQYR022.
[18]	FU J L, ZHENG H L, MEI T. Look closer to see better: recurrent attention convolutional neural network for fine-grained image recognition[C]// The IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2017: 4476-4484.
[19]	ZHENG H L, FU J L, MEI T, et al. Learning multi-attention convolutional neural network for fine-grained image recognition[C]// The IEEE International Conference on Computer Vision. New York: IEEE Press, 2017: 5219-5227.
[20]	LI M, LEI L, SUN H, et al. Fine-grained visual classification via multilayer bilinear pooling with object localization[J]. The Visual Computer, 2022, 38(3): 811-820.
[21]	ZHUANG P Q, WANG Y L, QIAO Y. Learning attentive pairwise interaction for fine-grained classification[C]// The AAAI Conference on Artificial Intelligence. Palo Alto: AAAI, 2020: 13130-13137.
[22]	CHANG D L, DING Y F, XIE J Y, et al. The devil is in the channels: mutual-channel loss for fine-grained image classification[J]. IEEE Transactions on Image Processing, 2020, 29: 4683-4695.
[23]	XIE S N, GU J T, GUO D M, et al. PointContrast: unsupervised pre-training for 3D point cloud understanding[C]// The 16th European Conference on Computer Vision. Cham: Springer, 2020: 574-591.
[24]	ZHANG Z W, GIRDHAR R, JOULIN A, et al. Self-supervised pretraining of 3D features on any point-cloud[C]// The IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 10232-10243.
[25]	HUANG S Y, XIE Y C, ZHU S C, et al. Spatio-temporal self-supervised representation learning for 3D point clouds[C]// The IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 6515-6525.
[26]	MERSCH B, CHEN X, BEHLEY J, et al. Self-supervised point cloud prediction using 3D spatio-temporal convolutional networks[EB/OL]. [2024-06-22]https://dblp.uni-trier.de/db/conf/corl/corl2021.html#MerschCBS21.
[27]	ZHANG R R, GUO Z Y, ZHANG W, et al. PointCLIP: point cloud understanding by CLIP[C]// The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 8542-8552.
[28]	MITTAL H, OKORN B, HELD D. Just go with the flow: self-supervised scene flow estimation[C]// The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 11174-11182.
[29]	LI R B, ZHANG C, LIN G S, et al. RigidFlow: self-supervised scene flow learning on point clouds by local rigidity prior[C]// The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 16938-16947.
[30]	VASWANI A. Attention is all you need[C]// The 31st International Conference on Neural Information Processing Systems. New York: ACM, 2017: 6000-6010.
[31]	HE K M, FAN H Q, WU Y X, et al. Momentum contrast for unsupervised visual representation learning[C]// The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 9726-9735.
[32]	CHANG A X, FUNKHOUSER T, GUIBAS L, et al. ShapeNet:an information-rich 3D model repository[EB/OL]. [2024-06-22]https://arxiv.org/abs/1512.03012.
[33]	LIU X H, HAN Z Z, LIU Y S, et al. Fine-grained 3D shape classification with hierarchical part-view attention[J]. IEEE Transactions on Image Processing, 2021, 30: 1744-1758. DOI PMID
[34]	SUN J C, ZHANG Q Z, KAILKHUNRA B, et al. ModelNet40-C: a robustness benchmark for 3D point cloud recognition under corruption[EB/OL]. [2024-06-22]https://iclr.cc/virtual/2022/8408.
[35]	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]// The IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2019: 1588-1597.
[36]	SHARMA C, KAUL M. Self-supervised few-shot learning on point clouds[C]// The 34th International Conference on Neural Information Processing Systems. New York: ACM, 2020: 605.
[37]	LOSHCHILOV I, HUTTER F. Decoupled weight decay regularization[EB/OL]. [2024-06-22]https://dblp.uni-trier.de/db/conf/iclr/iclr2019.html#LoshchilovH19.
[38]	LOSHCHILOV I, HUTTER F. GDR: stochastic gradient descent with warm restarts[EB/OL]. [2024-06-22]https://dblp.uni-trier.de/db/conf/iclr/iclr2017.html#LoshchilovH17.
[39]	蒋杰, 熊昌镇. 一种数据增强和多模型集成的细粒度分类算法[J]. 图学学报, 2018, 39(2): 244-250. DOI
	JIANG J, XIONG C Z. Data augmentation with multi-model ensemble for fine-grained category classification[J]. Journal of Graphics, 2018, 39(2): 244-250 (in Chinese). DOI
[40]	常东良, 尹军辉, 谢吉洋, 等. 面向图像分类的基于注意力引导的Dropout[J]. 图学学报, 2021, 42(1): 32-36.
	CHANG D L, YIN J H, XIE J Y, et al. Attention-guided Dropout for image classification[J]. Journal of Graphics, 2021, 42(1): 32-36 (in Chinese). DOI
[41]	韩亚振, 尹梦晓, 马伟钊, 等. DGOA: 基于动态图和偏移注意力的点云上采样[J]. 图学学报, 2024, 45(1): 219-229. DOI
	HAN Y Z, YIN M X, MA W Z, et al. DGOA: point cloud upsampling based on dynamic graph and offset attention[J]. Journal of Graphics, 2024, 45(1): 219-229 (in Chinese). DOI
[42]	LI H L, LIN B, ZHANG C, et al. Mask-point: automatic 3D surface defects detection network for fiber-reinforced resin matrix composites[J]. Polymers, 2022, 14(16): 3390.
[43]	LI J X, CHEN B M, LEE G H. SO-Net: self-organizing network for point cloud analysis[C]// The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2018: 9397-9406.
[44]	LIU X H, HAN Z Z, LIU Y S, et al. Point2Sequence: learning the shape representation of 3D point clouds with an attention-based sequence to sequence network[C]// The 33rd AAAI Conference on Artificial Intelligence. Palo Alto: AAAI, 2019: 8778-8785.
[45]	LI Y Y, BU R, SUN M C, et al. PointCNN: convolution on X-transformed points[C]// The 32nd International Conference on Neural Information Processing Systems. New York: ACM, 2018: 828-838.
[46]	WANG Y, SUN Y B, LIU Z W, et al. Dynamic graph CNN for learning on point clouds[J]. ACM Transactions on Graphics, 2019, 38(5): 146.
[47]	白静, 徐浩钧. MSP-Net: 多尺度点云分类网络[J]. 计算机辅助设计与图形学学报, 2019, 31(11): 1917-1924.
	BAI J, XU H J. MSP-Net: multi-scale point cloud classification network[J]. Journal of Computer-Aided Design & Computer Graphics, 2019, 31(11): 1917-1924 (in Chinese).
[48]	PAN L, CHEW C M, LEE G H. PointAtrousGraph: deep hierarchical encoder-decoder with point atrous convolution for unorganized 3D points[C]// 2020 IEEE International Conference on Robotics and Automation. New York: IEEE Press, 2020: 1113-1120.
[49]	WEN X, HAN Z Z, LIU X H, et al. Point2SpatialCapsule: aggregating features and spatial relationships of local regions on point clouds using spatial-aware capsules[J]. IEEE Transactions on Image Processing, 2020, 29: 8855-8869.
[50]	ZHANG R R, GUO Z Y, GAO P Y, et al. Point-M2AE: multi-scale masked autoencoders for hierarchical point cloud pre-training[C]// The 36th International Conference on Neural Information Processing Systems. New York: ACM, 2022: 1962.

方法分类	算法	年份	ModelNet40/%	FG3D/%
方法分类	算法	年份	ModelNet40/%	Airplane	Car	Chair
自监督	Point-BERT^[8]	2020	92.70	-	-	-
	PointGLR^[37]	2021	93.00	-	-	-
	OcCo^[7]	2020	93.00+voting	-	-	-
	MaskPoint^[42]	2020	93.80+voting	-	-	-
	Point-MAE^[9]	2021	93.20	-	-	-
	Point2Vec^[11]	2023	94.00	-	-	-
元类别	SO-Net^[43]	2018	90.90	82.92	59.32	70.05
	Point2Sequence^[44]	2018	92.60	92.76	73.54	79.12
	PointCNN^[45]	2018	91.70	90.30	68.37	74.87
	PointNet^[13]	2017	89.20	89.34	73.00	75.44
	DGCNN^[46]	2019	92.20	93.60	72.10	79.53
	MSP-Net^[47]	2019	91.73	93.03	74.25	68.69
	poinAtrousGraph^[48]	2020	93.10	95.22	74.77	79.20
	Point2SpatialCapsule^[49]	2020	93.40	95.19	75.92	79.53
	PointNet++(MSG)^[14]	2018	90.70	95.96	77.87	81.23
	PointTransformer^[15]	2020	93.70	91.53	67.88	71.73
	PCT^[16]	2020	93.20	95.16	78.89	81.37
	PointMLP^[17]	2022	94.50	95.76	76.35	81.81
细粒度	FGP-Net^[1]	2023	-	95.77	77.94	80.88
	FGPNet^[3]	2023	91.18	96.07	79.46	82.49
	DC-Net^[2]	2023	92.41	97.31	79.15	83.67
	Ours	2024	94.03	97.68	79.62	84.04

方法分类	算法	年份	ModelNet40/%	FG3D/%
方法分类	算法	年份	ModelNet40/%	Airplane	Car	Chair
自监督	Point-BERT^[8]	2020	92.70	-	-	-
	PointGLR^[37]	2021	93.00	-	-	-
	OcCo^[7]	2020	93.00+voting	-	-	-
	MaskPoint^[42]	2020	93.80+voting	-	-	-
	Point-MAE^[9]	2021	93.20	-	-	-
	Point2Vec^[11]	2023	94.00	-	-	-
元类别	SO-Net^[43]	2018	90.90	82.92	59.32	70.05
	Point2Sequence^[44]	2018	92.60	92.76	73.54	79.12
	PointCNN^[45]	2018	91.70	90.30	68.37	74.87
	PointNet^[13]	2017	89.20	89.34	73.00	75.44
	DGCNN^[46]	2019	92.20	93.60	72.10	79.53
	MSP-Net^[47]	2019	91.73	93.03	74.25	68.69
	poinAtrousGraph^[48]	2020	93.10	95.22	74.77	79.20
	Point2SpatialCapsule^[49]	2020	93.40	95.19	75.92	79.53
	PointNet++(MSG)^[14]	2018	90.70	95.96	77.87	81.23
	PointTransformer^[15]	2020	93.70	91.53	67.88	71.73
	PCT^[16]	2020	93.20	95.16	78.89	81.37
	PointMLP^[17]	2022	94.50	95.76	76.35	81.81
细粒度	FGP-Net^[1]	2023	-	95.77	77.94	80.88
	FGPNet^[3]	2023	91.18	96.07	79.46	82.49
	DC-Net^[2]	2023	92.41	97.31	79.15	83.67
	Ours	2024	94.03	97.68	79.62	84.04

算法	5-way		10-way
算法	10-shot	20-shot	10-shot	20-shot
OcCo^[7]	91.9±3.6	93.9±3.1	86.4±5.4	91.3±4.6
Transf.-OcCo^[8]	94.0±3.6	95.9±2.3	89.4±5.1	92.4±4.6
Point-BERT^[9]	94.6±3.1	96.3±2.7	91.0±5.4	92.7±5.1
MaskPoint^[42]	95.0±3.7	97.2±1.7	91.4±4.0	93.4±3.5
Point-MAE^[10]	96.3±2.5	97.8±1.8	92.6±4.1	95.0±3.0
Point-M2AE^[50]	96.8±1.8	98.3±1.4	92.3±4.5	95.0±3.0
Point2Vec^[11]	97.0±2.8	98.7±1.2	93.9±4.1	95.8±3.1
Ours	97.2±2.8	98.9±1.1	95.05±4.95	96.7±2.3

算法	5-way		10-way
算法	10-shot	20-shot	10-shot	20-shot
OcCo^[7]	91.9±3.6	93.9±3.1	86.4±5.4	91.3±4.6
Transf.-OcCo^[8]	94.0±3.6	95.9±2.3	89.4±5.1	92.4±4.6
Point-BERT^[9]	94.6±3.1	96.3±2.7	91.0±5.4	92.7±5.1
MaskPoint^[42]	95.0±3.7	97.2±1.7	91.4±4.0	93.4±3.5
Point-MAE^[10]	96.3±2.5	97.8±1.8	92.6±4.1	95.0±3.0
Point-M2AE^[50]	96.8±1.8	98.3±1.4	92.3±4.5	95.0±3.0
Point2Vec^[11]	97.0±2.8	98.7±1.2	93.9±4.1	95.8±3.1
Ours	97.2±2.8	98.9±1.1	95.05±4.95	96.7±2.3

算法	OBJ-BG	OBJ-ONLY	PB-T50-RS
Transf.-OcCo^[8]	84.90	85.50	78.80
Point-BERT^[9]	87.40	88.10	83.10
MaskPoint^[42]	89.30	89.70	84.60
Point-MAE^[10]	90.00	88.30	85.20
Point2Vec^[11]	91.20	90.40	87.50
Ours	91.02	90.33	85.98

面向三维点云的平衡泛化和特化的细粒度分类网络

BGS-Net: fine-grained classification networks with balanced generalization and specialization for 3D point clouds

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 10

参考文献 50

相关文章 7

编辑推荐

Metrics

本文评价

实验	上游学生机1	上游学生机2	下游学生机1	下游学生机2	FG3D/%			Modelnet40/%
实验	上游学生机1	上游学生机2	下游学生机1	下游学生机2	Airplane	Car	Chair	Modelnet40/%
①	随机掩码	-	非冻结	-	97.27	78.48	82.90	94.00
②	随机掩码	随机掩码	非冻结	-	97.32	79.01	83.09	93.66
③	随机掩码	随机掩码	冻结	非冻结	97.40	79.40	83.73	93.52
④	随机掩码	中心变换	冻结	非冻结	97.27	79.43	83.78	93.23
⑤	随机掩码	中心周围点掩码	冻结	非冻结	96.99	79.01	82.75	93.92
⑥	耦合掩码	耦合掩码	冻结	非冻结	97.68	79.62	84.04	94.03

实验	掩码率/%	掩码重合度/%	共享权重	FG3D/%			Modelnet40/%
实验	掩码率/%	掩码重合度/%	共享权重	Airplane	Car	Chair	Modelnet40/%
①	65	15	×	96.72	79.40	83.32	93.52
②	50	0	×	95.65	78.71	79.59	91.82
③	50	0	√	97.68	79.62	84.04	94.03

实验	任务不可知编码器掩码率	任务特化编码器掩码率	FG3D			Modelnet40
实验	任务不可知编码器掩码率	任务特化编码器掩码率	Airplane	Car	Chair	Modelnet40
①	-	-	97.33	78.94	83.31	93.65
②	65	65	97.35	78.99	83.68	93.72
④	随机掩码	-	97.45	79.22	83.45	93.63
⑤	-	随机掩码	97.68	79.62	84.04	94.03

实验	损失函数	FG3D			Modelnet40
实验	损失函数	Airplane	Car	Chair	Modelnet40
①	Null	97.31	78.87	83.47	92.98
②	L1	97.27	79.42	83.77	93.75
③	L2	97.13	78.56	83.16	93.84
④	SL1	97.68	79.62	84.04	94.03

[1]	王雪婷, 郭新, 汪松, 陈恩庆. 基于变分自编码器掩蔽重建的骨骼点动作识别方法[J]. 图学学报, 2025, 46(2): 270-278.
[2]	林晓, 张秋阳, 郑晓妹, 杨启哲. 基于自监督的主动标签清洗[J]. 图学学报, 2024, 45(3): 495-504.
[3]	安峰 , 戴军 , 韩振 , 严仲兴 . 引入注意力机制的自监督光流计算[J]. 图学学报, 2022, 43(5): 841-848.
[4]	李妮妮, 王夏黎, 付阳阳, 郑凤仙, 何丹丹, 袁绍欣. 一种优化 YOLO 模型的交通警察目标检测方法[J]. 图学学报, 2022, 43(2): 296-305.
[5]	李农勤 1，杨维信 2,3 . 基于生成式对抗神经网络的手写文字图像补全[J]. 图学学报, 2019, 40(5): 878-884.
[6]	朱文佳 1，陈宇红 2，冯瑜瑾 2，王俊 3，余烨 3 . 一种基于目标优化学习的车标识别方法[J]. 图学学报, 2019, 40(4): 689-696.
[7]	杨世强，乔丹，弓逯琦，李小莉，李德信 . 基于 Laplace 逼近 Gaussian 过程的指节图像中层偏移测度特征学习[J]. 图学学报, 2019, 40(3): 574-582.