适用于无监督行人重识别的反向骨干网

doi:10.13229/j.cnki.jdxbgxb.20240916

摘要/Abstract

摘要：

行人重识别（re-id）的目的是在不同的摄像机上识别同一个人的图像。虽然无监督模型比有监督模型有更好的泛用性，但无监督的聚类会更容易受到噪声干扰。针对这一问题，本文提出了一个可以减少噪声干扰的模型反向骨干网（RBNet），利用反向骨干网学习姿态检测模型输出的人体关键点，调整局部空间信息并生成掩码，用生成掩码增强指定位置注意力。实验结果表明：对比baseline在Market-1501到DukeMTMC-reID的跨域实验结果，mAP提升了7.0%，Rank-1提升了6.4%。强化对不同局部信息的注意力，可有效提升模型准确率。

关键词: 人工智能, 行人重识别, 人体关键点, 无监督域自适应

Abstract:

The purpose of person re-identification （re-id） is to identify images of the same person on different cameras. Although unsupervised models have better generalization than supervised models， unsupervised clustering will be more susceptible to noise interference. To address this problem， this paper proposes a model reverse backbone net （RBNet） that can reduce noise interference， using RBNet to learn the keypoints of the human body output from the pose detection model， adjusting the local spatial information and generating masks， and augmenting the specified location attention with the generated masks. The experimental results show that comparing baseline's cross-domain experimental results from Market-1501 to DukeMTMC-reID， mAP is enhanced by 7.0% and Rank-1 by 6.4%. Strengthening the attention to different local information can effectively improve the model accuracy.

Key words: artificial intelligence, person re-identification, human keypoints, unsupervised domain adaptation

中图分类号:

TP391.4

于鹏,朴燕. 适用于无监督行人重识别的反向骨干网[J]. 吉林大学学报(工学版), 2024, 54(11): 3309-3317.

Peng YU,Yan PIAO. Reverse backbone net for unsupervised person re-identification[J]. Journal of Jilin University(Engineering and Technology Edition), 2024, 54(11): 3309-3317.

图/表 8

图 1

表 1

图2

表 2

图 3

图 4

表 3

表4

参考文献 31

1	Chiat P T, Sharmili R, Kim H Y. AANet: attribute attention network for person re-identifications[C]∥32nd IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 7127-7136.
2	Lin Y T, Zheng L, Zheng Z D, et al. Improving person re-identification by attribute and identity learning[J]. Pattern Recognition, 2019, 95: 151-161.
3	Sun Y F, Zheng L, Yang Y, et al. Beyond part models: person retrieval with refined part pooling (and a strong convolutional baseline)[C]∥15th European Conference on Computer Vision, Munich, Germany, 2018: 501-518.
4	Miao J X, Wu Y, Yang Y. Identifying visible parts via pose estimation for occluded person re-identification[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 33(9): 4624-4634.
5	Cai H L, Wang Z G, Cheng J X. Multi-scale body-part mask guided attention for person re-identification[C]∥32nd IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 1555-1564.
6	Qiao S Y, Chen L C, Alan Y. Detectors: detecting objects with recursive feature pyramid and switchable atrous convolution[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 10208-10219.
7	Hou S Q, Yin K N, Liang J, et al. Gradient-supervised person re-identification based on dense feature pyramid network[J]. Complex & Intelligent Systems, 2022, 8(6I): 5329-5342.
8	Zhang Z Z, Lan G L, Zeng W J, et al. Relation-aware global attention for person re-identification[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020: 3183-3192.
9	Zhou K Y, Yang Y X, Cavallaro A, et al. Omni-scale feature learning for person re-identification[C]∥IEEE/CVF International Conference on Computer Vision, Seoul, South Korea, 2019: 3701-3711.
10	Divyansh G, Netra P, Thomas T, et al. Towards explainable person re-identification[C]∥IEEE Symposium Series on Computational Intelligence, Orlando, USA, 2021: 9660071.
11	Niki M, Gian L F, Christian M. Deep pyramidal pooling with attention for person re-identification[J]. IEEE Transactions on Image Processing, 2020, 29: 7306-7316.
12	Yang F X, Zhong Z, Luo Z M, et al. Joint noise-tolerant learning and meta camera shift adaptation for unsupervised person re-identification[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 4853-4862.
13	Antti T, Harri V. Mean teachers are better role models: weight-averaged consistency targets improve semi-supervised deep learning results[C]∥31st Annual Conference on Neural Information Processing Systems, Long Beach, USA, 2017: 1195-1204.
14	Ge Y X, Chen D P, Li H S. Mutual mean-teaching: pseudo label refinery for unsupervised domain adaptation on person re-identification[C]∥8th International Conference on Learning Representations, Addis Ababa, Ethiopia, 2020: 200101526.
15	He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]∥ IEEE Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2016: 770-778.
16	Sun K, Xiao B, Liu D, et al. Deep high-resolution representation learning for human pose estimation[C]∥32nd IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 5686-5696.
17	Martin E, Hans P K, Jorg S, et al. A density-based algorithm for discovering clusters in large spatial databases with noise[C]∥2nd International Conference on Knowledge Discovery and Data Mining, Portland, USA, 1996: 226-231.
18	Zheng L, Shen L Y, Tian L, et al. Scalable person re-identification: a benchmark[C]∥IEEE International Conference on Computer Vision, Santiago, USA, 2015: 1116-1124.
19	Ergys R, Francesco S, Roger Z, et al. Performance measures and a data set for multi-target, multi-camera tracking[C]∥14th European Conference on Computer Vision, Amsterdam, Netherlands, 2016: 17-35.
20	Wei L H, Zhang S L, Gao W, et al. Person transfer GAN to bridge domain gap for person re-identification[C]∥31st IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 79-88.
21	Park H J, Ham B. Relation network for person re-identification[C]∥34th AAAI Conference on Artificial Intelligence, New York, USA, 2020: 11839-11847.
22	He L X, Liao X Y, Liu W, et al. Fastreid: a pytorch toolbox for general instance re-identification[J]. Proceedings of the 31st ACM International Conference on Multimedia, 2020, 10: 9662-9667.
23	Tan H C, Liu X P, Bian Y H, et al. Incomplete descriptor mining with elastic loss for person re-identification[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 32(1): 160-171.
24	Li Y L, He J F, Zhang T Z, et al. Diverse part discovery: occluded person re-identification with part-aware transformer[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 2897-2906.
25	Li H J, Wu G J, Zheng W S. Combined depth space based architecture search for person re-identification[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 6725-6734.
26	Zheng K C, Liu W, He L X, et al. Group-aware label transfer for domain adaptive person re-identification[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 5306-5315.
27	Wang D K, Zhang S L. Unsupervised person re-identification via multi-label classification[J]. International Journal of Computer Vision, 2022, 130(12): 2924-2939.
28	Zou Y, Yang X D, Yu Z D, et al. Joint disentangling and adaptation for cross-domain person re-identification[C]∥16th European Conference on Computer Vision, Glasgow, UK, 2020: 87-104.
29	Jin X, Lan C L, Zeng W J, et al. Global distance-distributions separation for unsupervised person re-identification[C]∥16th European Conference on Computer Vision, Glasgow, UK, 2020: 735-751.
30	Zhong Z, Zheng L, Luo Z M, et al. Learning to adapt invariance in memory for person re-identification[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(8): 2723-2738.
31	Chen G Y, Gu T P, Lu J W, et al. Person re-identification via attention pyramid[J]. IEEE Transactions on Image Processing, 2021, 30: 7663-7676.

相关文章 15

[1]	曹毅,夏宇,高清源,叶培涛,叶凡. 基于超连接图卷积网络的骨架行为识别方法[J]. 吉林大学学报(工学版), 2025, 55(2): 731-740.
[2]	井佩光,田雨豆,汪少初,李云,苏育挺. 基于动态扩散图卷积的交通流量预测算法[J]. 吉林大学学报(工学版), 2024, 54(6): 1582-1592.
[3]	曲福恒,潘曰涛,杨勇,胡雅婷,宋剑飞,魏成宇. 基于加权空间划分的高效全局K-means聚类算法[J]. 吉林大学学报(工学版), 2024, 54(5): 1393-1400.
[4]	孙帅帅,冯春晓,张良. 基于离散采样的多模态四足机器人路径规划[J]. 吉林大学学报(工学版), 2024, 54(4): 1120-1128.
[5]	金志刚,苏仁鋆,赵晓芳. 基于异质图网络的心理评估方法[J]. 吉林大学学报(工学版), 2024, 54(4): 1078-1085.
[6]	高敬鹏,王国轩,高路. 基于异步合作更新的LSTM-MADDPG多智能体协同决策算法[J]. 吉林大学学报(工学版), 2024, 54(3): 797-806.
[7]	刘浏,丁鲲,刘姗姗,刘茗. 基于机器阅读理解的事件检测方法[J]. 吉林大学学报(工学版), 2024, 54(2): 533-539.
[8]	李健,熊琦,胡雅婷,刘孔宇. 基于Transformer和隐马尔科夫模型的中文命名实体识别方法[J]. 吉林大学学报(工学版), 2023, 53(5): 1427-1434.
[9]	于鹏,朴燕. 基于多尺度特征的行人重识别属性提取新方法[J]. 吉林大学学报(工学版), 2023, 53(4): 1155-1162.
[10]	刘春晖,王思长,郑策,陈秀连,郝春蕾. 基于深度学习的室内导航机器人避障规划算法[J]. 吉林大学学报(工学版), 2023, 53(12): 3558-3564.
[11]	白天,徐明蔚,刘思铭,张佶安,王喆. 基于深度神经网络的诉辩文本争议焦点识别[J]. 吉林大学学报(工学版), 2022, 52(8): 1872-1880.
[12]	王生生,姜林延,杨永波. 基于最优传输特征选择的医学图像分割迁移学习[J]. 吉林大学学报(工学版), 2022, 52(7): 1626-1638.
[13]	田皓宇,马昕,李贻斌. 基于骨架信息的异常步态识别方法[J]. 吉林大学学报(工学版), 2022, 52(4): 725-737.
[14]	刘勇,徐雷,张楚晗. 面向文本游戏的深度强化学习模型[J]. 吉林大学学报(工学版), 2022, 52(3): 666-674.
[15]	侯春萍,杨庆元,黄美艳,王致芃. 基于语义耦合和身份一致性的跨模态行人重识别方法[J]. 吉林大学学报(工学版), 2022, 52(12): 2954-2963.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

Method	Market-1501		DukeMTMC-reID		MSMT17
	mAP	R-1	mAP	R-1	mAP	R-1
GCP ^［21］	88.9	95.2	78.6	89.7	—	—
MMGA ^［5］	87.2	95.0	78.1	89.5	—	—
RGA-SC ^［8］	88.4	96.1	—	—	57.5	80.3
AANet-50 ^［1］	82.45	93.89	72.56	86.42	—	—
PCB ^［3］	81.6	93.8	69.2	83.3	—	—
OSNet ^［9］	84.9	94.8	73.5	88.6	52.9	78.7
FastReID ^［22］	88.2	95.4	79.8	89.6	59.9	83.3
CBDB-Net ^［23］	85.0	94.4	74.3	87.7	—	—
PAT ^［24］	88.0	95.4	78.2	88.8	—	—
CDNet ^［25］	86.0	95.1	76.8	88.6	54.7	78.9
RBNet（conv 1）	88.5	95.6	80.0	89.7	60.8	84.9
RBNet（relayer 1）	89.1	96.1	81.1	90.0	62.7	85.1
RBNet（relayer 2）	84.3	94.0	73.9	83.3	49.4	77.1

M-> D
基于模型	model	mAP	Rank-1	Rank-5	Rank-10
ResNet-50	MMCL ^［27］	51.4	72.4	—	—
E_app （ResNet 50）	DG-Net++ ^［28］	63.8	78.9	—	—
ResNet-50	GDS ^［29］	55.1	73.1	—	—
ResNet 50	GPP ^［30］	54.4	74.0	—	—
ResNet 50	MMT-500 ^［14］	63.1	76.8	88.0	92.2
MMT （ResNet 50）	MetaCam ^［12］	65.0	79.5
fast-reid	GLT ^［26］	69.2	82.0	90.2	92.8
MMT （ResNet 50）	MMT+RBNet	70.1	83.2	90.8	92.9
D-> M
基于模型	model	mAP	Rank-1	Rank-5	Rank-10
ResNet-50	MMCL	60.4	84.4	—	—
E_app（ResNet-50）	DG-Net++	61.7	82.1	—	—
ResNet-50	GDS	61.2	81.1	—	—
ResNet 50	GPP+	63.8	84.1	—	—
ResNet 50	MMT-500	71.2	87.7	94.9	96.9
MMT （ResNet 50）	MetaCam	76.5	90.1	—	—
fast-reid	GLT	79.5	92.2	96.5	97.8
MMT （ResNet 50）	MMT+RBNet	79.8	92.4	96.7	97.9

M->D
Method	mAP	Rank-1	Rank-5	Rank-10
APNet-c3 ^［31］	22.8	37.7	52.4	59.0
OSNet ^［9］	26.7	48.5	62.3	67.4
resnet 50	17.3	32.7	47.1	53.4
resnet 50+RBNet	27.4	49.4	62.9	67.7
D->M
Method	mAP	Rank-1	Rank-5	Rank-10
APNet-c3	23.7	50.9	66.6	72.6
OSNet	26.1	57.7	73.7	80.0
resnet 50	17.7	43.9	50.9	67.5
resnet 50+RBNet	27.5	60.6	75.7	80.8