基于注意力机制和循环域三元损失的域适应目标检测

周洋; 韩冰; 高新波; 杨铮; 陈玮铭

doi:10.16383/j.aas.c220938

基于注意力机制和循环域三元损失的域适应目标检测

doi: 10.16383/j.aas.c220938

周洋^1,,
韩冰^1,,
高新波^{1, 2,},
杨铮^1,,
陈玮铭^1,

1.
西安电子科技大学电子工程学院西安 710071
2.
重庆邮电大学图像认知重庆市重点实验室重庆 400065

基金项目: 国家自然科学基金(62076190, 41831072, 41874195, 62036007), 陕西省重点创新产业链(2022ZDLGY01-11)资助

详细信息

作者简介:
周洋：西安电子科技大学电子工程学院硕士研究生. 2020获西南石油大学电子信息工程学士学位. 主要研究方向为计算机视觉和域适应目标检测. E-mail: yzhou_6@stu.xidian.edu.cn

韩冰：博士, 西安电子科技大学电子工程学院教授. 目前主要研究领域为智能辅助驾驶、视觉感知与认知、空间物理与人工智能交叉研究等. 本文通信作者. E-mail: bhan@xidian.edu.cn

高新波：博士, 西安电子科技大学教授, 重庆邮电大学校长. 目前主要从事机器学习、图像处理、计算机视觉、模式识别和多媒体内容分析等领域的研究. E-mail: xbgao@ieee.org

杨铮：西安电子科技大学电子工程学院博士研究生. 2017获西安电子科技大学智能科学与技术学士学位. 主要研究方向为深度学习, 目标跟踪和强化学习. E-mail: zhengy@stu.xidian.edu.cn

陈玮铭：西安电子科技大学电子工程学院硕士研究生. 2019获西安电子科技大学机械设计制造及其自动化学士学位. 主要研究方向为计算机视觉, 目标检测和遥感技术. E-mail: wmchen@stu.xidian.edu.cn

计量
- 文章访问数: 520
- HTML全文浏览量: 218
- 被引次数: 0
出版历程
- 收稿日期: 2022-12-05
- 录用日期: 2023-05-18
- 网络出版日期: 2023-08-18

Domain Adaptive Object Detection Based on Attention Mechanism and Cycle Domain Triplet Loss

ZHOU Yang^1
,,
HAN Bing^1
,,
GAO Xin-Bo^{1, 2
,},
YANG Zheng^1
,,
CHEN Wei-Ming^1
,

1.
School of Electronic Engineering, XiDian University, Xian, 710071
2.
Chongqing Key Laboratory of Image Cognition, Chongqing University of Posts and Telecommunications, Chongqing, 400065

Funds: Supported by National Natural Science Foundation of China (62076190,41831072, 41874195, 62036007), The Key Industry Innovation Chain of Shaanxi under Grant (2022ZDLGY01-11)

More Information

Author Bio:
ZHOU Yang　Master student at the School of Electronic Engineering, Xidian University. He received his bachelor degree in electronic and information engineering from Southwest Petroleum University in 2020. His research interest covers computer vision and domain adaptive detection

HAN Bing　Professor at school of Electronic Engineering, Xidian University. Her research interest covers intelligent auxiliary drive syste, visual perception and cognition, and cross-disciplinary research between space physics and pattern recognition. Corresponding author of this paper

GAO Xin-bo　Professor of Xidian University and the president of Chongqing University. His current research interests include Machine Learning, Image processing, computer vision, pattern recognition and multimedia analysis

YANG Zheng　Ph. D.candidate at the School of Electronic Engineering, Xidian University. He received his bachelor degree in intelligent science and technology from Xidian University in 2019. His research interests include deep learning, object tracking, and reinforcement learning

CHEN Wei-Ming　Master student at the School of Electronic Engineering, Xidian University. He received his bachelor degree in mechanical design manufacture and automation from Xidian University in 2019. His research interests include computer vision, object detection, and remote sensing

摘要

摘要: 目前大多数深度学习算法都依赖于大量的标注数据并欠缺一定的泛化能力. 无监督域适应算法能提取到已标注数据和未标注数据间隐式共同特征, 从而提高算法在未标注数据上的性能. 目前域适应目标检测算法主要为两阶段目标检测器设计. 针对单阶段检测器中无法直接进行实例级特征对齐导致一定数量域不变特征的缺失, 提出结合通道注意力机制的图像级域分类器加强域不变特征提取. 此外对于域适应目标检测中存在类别特征的错误对齐引起的精度下降问题, 通过原型学习构建类别中心, 设计了一种基于原型的循环域三元损失函数, 从而实现原型引导的精细类别特征对齐. 以单阶段目标检测算法作为检测器, 在多种域适应目标检测公共数据集上进行实验. 实验结果证明该方法能有效提升原检测器在目标域的泛化能力达到更高的检测精度, 并且对于单阶段目标检测网络具有一定的通用性.
- 无监督领域自适应 /
- 注意力机制 /
- 循环域三元损失函数 /
- 目标检测
Abstract: Most current deep learning algorithms rely heavily on large amounts of annotated data and exist deficiency in generalization ability. The domain adaptation algorithm can extract the common implicit invariant features from the labeled data and unlabeled data, so that the algorithm can achieve good generalization performance on the unlabeled data. At present, domain adaptation algorithms have made excellent achievements for two-stage object detection. For the one-stage object detection, the difficulty of explicit aligning instance-level features leads to the absence of a number of domain invariant features. In this paper, a image-level domain classifier with channel attention mechanism is proposed to strengthen domain invariant feature extraction. In addition, by constructing the category centers through prototype learning, a cycle domain triplet loss function based on the category centers of different domains is designed, so as to realize the finer category feature alignment guided by the category. It can mitigate misalignment of category feature. One-stage object detection algorithms are used as detectors. Experiments are conducted on various domain object detection public datasets. The experimental results show that our method can improve the generalization ability on target domain and achieves the better results than other methods. Meanwhile, the experiment on different detector indicate ours method is universal for the one-stage object detection network.
- Unsupervised domain adaptation /
- attention mechanism /
- cycle domain triplet loss /
- object detection

HTML全文

图 1 基于注意力机制和循环域三元损失的域适应目标检测算法流程

Fig. 1 The pipeline of domain adaptive object detection based on attention mechanism and cycle domain triplet loss

下载: 全尺寸图片幻灯片

图 2 循环域三元损失函数原理图

Fig. 2 Principal of cycle domain adaptive tripleLoss

下载: 全尺寸图片幻灯片

图 3 本文方法在CityScapes→Foggy CityScapes上的主观检测结果

Fig. 3 The Subjective results of the ours method on CityScapes→Foggy CityScapes

下载: 全尺寸图片幻灯片

图 4 本文方法在SunnyDay→DuskRainy和SunnyDay→NightRainy上的主观检测结果

Fig. 4 The Subjective results of the ours method on SunnyDay→DuskRainy and SunnyDay→NightRainy

下载: 全尺寸图片幻灯片

图 5 本文方法在KITTI→CityScapes和Sim10k→CityScapes上的消融实验结果

Fig. 5 The Subjective results of the ablation experiment on KITTI→CityScapes and Sim10k→CityScapes

下载: 全尺寸图片幻灯片

图 6 本文方法在VOC→Clipart1k上的主观结果

Fig. 6 The Subjective results of ours on VOC→Clipart1k

下载: 全尺寸图片幻灯片

图 7 不同循环迭代训练次数在YOLOv3和YOLOv5检测器上的结果

Fig. 7 The result of different cycle iteration on YOLOv3 and YOLOv5

下载: 全尺寸图片幻灯片

表 1 CityScapes→Foggy CityScapes数据集上的对比实验, “−”代表该方法没有进行此实验

Table 1 The results of different methods on CityScapes→Foggy CityScapes, “−” represents the experiment is absent on this method

方法	检测器	person	rider	car	truck	bus	motor	bike	train	mAP	mGP
DAF^[10]	Faster-RCNN	25.0	31.0	40.5	22.1	35.3	20.0	27.1	20.2	27.7	38.8
SWDA^[11]	Faster-RCNN	29.9	42.3	43.5	24.5	36.2	30.0	35.3	32.6	34.3	70.0
C2F^[14]	Faster-RCNN	34.0	46.9	52.1	30.8	43.2	34.7	37.4	29.9	38.6	79.1
CAFA^[16]	Faster-RCNN	41.9	38.7	56.7	22.6	41.5	24.6	35.5	26.8	36.0	81.9
ICCR-VDD^[21]	Faster-RCNN	33.4	44.0	51.7	33.9	52.0	34.2	36.8	34.7	40.0	—
MeGA^[20]	Faster-RCNN	37.7	49.0	52.4	25.4	49.2	34.5	39.0	46.9	41.8	91.1
DAYOLO^[28]	YOLOv3	29.5	27.7	46.1	9.1	28.2	12.7	24.8	4.5	36.1	61.0
本文方法(v3)	YOLOv3	34.0	37.2	55.8	31.4	44.4	22.3	30.8	50.7	38.3	83.9
MS-DAYOLO^[31]	YOLOv4	39.6	46.5	56.5	28.9	51.0	27.5	36.0	45.9	41.5	68.6
A-DAYOLO^[32]	YOLOv5	32.8	35.7	51.3	18.8	34.5	11.8	25.6	16.2	28.3	—
S-DAYOLO^[34]	YOLOv5	42.6	42.1	61.9	23.5	40.5	24.4	37.3	39.5	39.0	69.9
本文方法(v5)	YOLOv5	30.9	37.4	53.3	23.8	39.5	24.2	29.9	35.0	34.3	83.8

下载: 导出CSV

表 2 SunnyDay→DuskRainy数据集上的对比实验

Table 2 The results of different methods on SunnyDay→DuskRainy

方法	检测器	bus	bike	car	motor	person	rider	truck	mAP	$\Delta$mAP
DAF^[10]	Faster-RCNN	43.6	27.5	52.3	16.1	28.5	21.7	44.8	33.5	5.2
SWDA^[11]	Faster-RCNN	40.0	22.8	51.4	15.4	26.3	20.3	44.2	31.5	3.2
ICCR-VDD^[21]	Faster-RCNN	47.9	33.2	55.1	26.1	30.5	23.8	48.1	37.8	9.5
本文方法(v3)	YOLOv3	50.1	24.9	70.7	24.2	39.1	19.0	53.2	40.2	7.4
本文方法(v5)	YOLOv5	46.2	22.1	68.2	16.5	34.8	17.5	50.5	36.5	9.4

下载: 导出CSV

表 3 SunnyDay→NightRainy数据集上的对比实验

Table 3 The results of different methods on SunnyDay→NightRainy

方法	检测器	bus	bike	car	motor	person	rider	truck	mAP	$\Delta$mAP
DAF^[10]	Faster-RCNN	23.8	12.0	37.7	0.2	14.9	4.0	29.0	17.4	1.1
SWDA^[11]	Faster-RCNN	24.7	10.0	33.7	0.6	13.5	10.4	29.1	17.4	1.1
ICCR-VDD^[21]	Faster-RCNN	34.8	15.6	38.6	10.5	18.7	17.3	30.6	23.7	7.4
本文方法(v3)	YOLOv3	45.0	8.2	51.1	4.0	20.9	9.6	37.9	25.3	5.1
本文方法(v5)	YOLOv5	40.7	9.3	45.0	0.6	12.8	9.2	32.5	21.5	4.7

下载: 导出CSV

表 4 KITTI→CityScapes和Sim10k→CityScapes数据集上的对比实验, “−”代表该方法没有进行此实验

Table 4 The results of different methods on KITTI→CityScapes and Sim10k→CityScapes, “−” represents the experiment is absent on this method

方法	KITTI→CityScapes		Sim10k→CityScapes
方法	AP	GP	AP	GP
DAF^[10]	38.5	21.0	39.0	22.5
SWDA^[11]	37.9	19.5	42.3	30.8
C2F^[14]	—	—	43.8	35.3
CAFA^[16]	43.2	32.9	49.0	47.7
MeGA^[20]	43.0	32.4	44.8	37.0
DAYOLO^[28]	54.0	82.2	50.9	39.5
本文方法(v3)	61.1	29.4	60.8	37.1
A-DAYOLO^[32]	37.7	—	44.9	—
S-DAYOLO^[34]	49.3	52.9	—	—
本文方法(v5)	60.0	50.4	60.3	56.3

下载: 导出CSV

表 5 CityScapes→FoggyCityscapes数据集上基于YOLOv3的消融实验

Table 5 The results of ablation experiment on CityScapes→FoggyCityscapes based on YOLOv3

方法	person	rider	car	truck	bus	motor	bike	train	mAP
SO	29.8	35.0	44.7	20.4	32.4	14.8	28.3	21.6	28.4
CADC	34.4	38.0	54.7	24.4	45.0	21.2	32.1	49.1	37.2
CDTL	31.1	38.0	46.7	28.9	34.5	23.4	27.8	13.7	30.5
CADC+CDTL	34.0	37.2	55.8	31.4	44.4	22.3	30.8	50.7	38.3
Oracle	34.9	38.8	55.9	25.3	45.0	22.6	33.4	49.1	40.2

下载: 导出CSV

表 6 CityScapes→FoggyCityscapes数据集上基于YOLOv5的消融实验

Table 6 The results of ablation experiment on CityScapes→FoggyCityscapes based on YOLOv5

方法	person	rider	car	truck	bus	motor	bike	train	mAP
SO	26.9	33.1	39.9	8.9	21.1	11.3	24.8	4.9	21.4
CADC	32.6	37.1	52.7	26.8	38.1	23.0	38.1	32.6	34.1
CDTL	29.7	36.7	43.2	13.1	25.5	17.1	28.7	13.1	26.2
CADC+CDTL	30.9	37.4	53.3	23.8	39.5	24.2	29.9	35.0	34.3
Oracle	34.8	37.9	57.5	24.4	42.7	23.1	33.2	40.8	36.8

下载: 导出CSV

表 7 SunnyDay→DuskRainy数据集上基于YOLOv3的消融实验

Table 7 The results of ablation experiment on SunnyDay→DuskRainy based on YOLOv3

方法	bus	bike	car	motor	person	rider	truck	mAP
Source Only	43.7	14.3	68.4	12.0	31.5	10.9	48.7	32.8
CADC	50.0	22.6	70.8	23.2	38.4	18.7	53.5	39.6
CDTL	45.4	20.1	69.2	15.2	34.8	17.2	47.8	35.7
CADC+ CDTL	50.1	24.9	70.7	24.2	39.1	19.0	53.2	40.2

下载: 导出CSV

表 8 SunnyDay→DuskRainy数据集上基于YOLOv5的消融实验

Table 8 The results of ablation experiment on SunnyDay→DuskRainy based on YOLOv5

方法	bus	bike	car	motor	person	rider	truck	mAP
Source Only	37.2	8.4	63.8	5.5	23.7	7.9	43.4	27.1
CADC	45.6	22.1	68.2	16.6	34.5	15.4	50.1	35.9
CDTL	41.6	13.1	65.5	7.6	29.7	10.2	44.9	30.4
CADC+ CDTL	46.2	22.1	68.2	16.5	34.8	17.5	50.5	36.5

下载: 导出CSV

表 9 SunnyDay→NightRainy数据集上基于YOLOv3的消融实验

Table 9 The results of ablation experiment on SunnyDay→NightRainy based on YOLOv3

方法	bus	bike	car	motor	person	rider	truck	mAP
Source Only	39.2	5.1	44.2	0.2	14.8	6.9	30.7	20.2
CADC	44.4	8.1	50.9	0.6	20.2	11.3	38.3	24.8
CDTL	40.4	8.2	45.8	0.6	16.2	7.2	33.4	21.7
CADC+ CDTL	45.0	8.2	51.1	4.0	20.9	9.6	37.9	25.3

下载: 导出CSV

表 10 SunnyDay→NightRainy数据集上基于YOLOv5的消融实验

Table 10 The results of ablation experiment on SunnyDay→NightRainy based on YOLOv5

方法	bus	bike	car	motor	person	rider	truck	mAP
Source Only	25.4	3.2	36.3	0.2	9.1	4.4	20.8	14.2
CADC	38.7	8.3	42.7	0.3	12.3	6.4	32.0	20.1
CDTL	34.3	6.2	44.2	0.5	11.2	8.7	30.3	19.3
CADC+ CDTL	40.7	9.3	45.0	0.6	12.8	9.2	32.5	21.5

下载: 导出CSV

表 11 KITTI→CityScapes和Sim10k→CityScapes数据集上的对比实验

Table 11 The results of different methods on KITTI→CityScapes and Sim10k→CityScapes

	KITTI	Sim10k
YOLOv3
Source Only	59.6	58.5
CADC	60.5	59.6
CDTL	60.5	60.8
CADC+ CDTL	61.1	59.8
Oracle	64.7	64.7
YOLOv5
Source Only	54.0	53.1
CADC	59.5	58.6
CDTL	59.0	60.3
CADC+ CDTL	60.0	59.0
Oracle	65.9	65.9

下载: 导出CSV

表 12 本文方法在VOC→Clipart1k上的实验

Table 12 The experiment on VOC→Clipart1k

方法	aero	bcycle	bird	boat	bottle	bus	car	cat	chair	cow	table	dog	hrs	bike	prsn	plnt	sheep	sofa	train	tv	mAP
I3Net	23.7	66.2	25.3	19.3	23.7	55.2	35.7	13.6	37.8	35.5	25.4	13.9	24.1	60.3	56.3	39.8	13.6	34.5	56.0	41.8	35.1
I3Net+CDTL	23.3	61.6	27.8	17.1	24.7	54.3	39.8	12.3	41.4	34.1	32.2	15.5	27.6	77.9	57.0	37.4	5.50	31.3	51.8	47.8	36.0
I3Net+CDTL+$CADC^*$	31.2	60.4	31.8	19.4	27.0	63.3	40.7	13.7	41.1	38.4	27.2	18.0	25.5	67.8	54.9	37.2	15.5	36.4	54.8	47.8	37.6

下载: 导出CSV

表 13 本文方法在VOC→Comic2k上的实验

Table 13 The experiment on VOC→Comic2k

方法	bike	bird	car	cat	dog	person	mAP
I3Net	44.9	17.8	31.9	10.7	23.5	46.3	29.2
I3Net+CDTL	43.7	15.1	31.5	11.7	18.6	46.9	27.9
I3Net+CDTL+CADC^*	47.8	16.0	33.8	15.1	24.4	43.5	30.1

下载: 导出CSV

表 14 本文方法在VOC→Watercolor2k上的实验

Table 14 The experiment on VOC→Watercolor2k

方法	bike	bird	car	cat	dog	person	mAP
I3Net	81.3	49.6	43.6	38.2	31.3	61.7	51.0
I3Net+CDTL	79.5	47.2	41.7	33.5	35.4	60.3	49.6
I3Net+CDTL+CADC^*	84.1	45.3	46.6	32.9	31.4	61.4	50.3

下载: 导出CSV

表 15 像素级对齐对网络的影响

Table 15 The impact of pixel alignment to network

方法	检测器	C→F	K→C	S→C
CDTL+CADC	YOLOv3	35.9	59.8	58.4
CDTL+CADC+$D_{pixel}$	YOLOv3	37.2	60.5	59.6
CDTL+CADC	YOLOv5	32.7	58.9	56.8
CDTL+CADC+$D_{pixel}$	YOLOv5	34.1	59.5	58.6

下载: 导出CSV

表 16 通道注意力域分类器中损失函数的选择

Table 16 The choice of loss function in channel attention domain classifier(CADC)

检测器	$F_1$	$F_2$	$F_3$	mAP
YOLOv3/v5	CE	CE	CE	35.8/32.7
YOLOv3/v5	CE	CE	FL	36.4/33.2
YOLOv3/v5	CE	FL	FL	37.2/34.1
YOLOv3/v5	FL	FL	FL	37.0/33.5

下载: 导出CSV

参考文献(48)

[1]	Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems (NIPS), Lake Tahoe, Nevada, USA: IEEE, 2012. 1097−1105.
[2]	Bottou L, Bousquet O. The tradeoffs of large scale learning[J]. Advances in neural information processing systems, 2007, 20.
[3]	Shen J, Qu Y, Zhang W, et al. Wasserstein distance guided representation learning for domain adaptation[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2018, 32.
[4]	皋军, 黄丽莉, 孙长银. 一种基于局部加权均值的领域适应学习框架. 自动化学报, 2013, 39(7): 1037-1052. doi: 10.3724/SP.J.1004.2013.01037. GAO Jun, HUANG Li-Li, SUN Chang-Yin. A Local Weighted Mean Based Domain Adaptation Learning Framework. ACTA AUTOMATICA SINICA, 2013, 39(7): 1037-1052. doi: 10.3724/SP.J.1004.2013.01037
[5]	Ganin Y, Ustinova E, Ajakan H, et al. Domain-adversarial training of neural networks[J]. The journal of machine learning research, 2016, 17(1): 2096-2030.
[6]	Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial networks[J]. Communications of the ACM, 2020, 63(11): 139-144. doi: 10.1145/3422622
[7]	郭迎春, 冯放, 阎刚, 郝小可. 基于自适应融合网络的跨域行人重识别方法. 自动化学报, 2022, 48(11): 2744-2756 doi: 10.16383/j.aas.c220083. Guo Ying-Chun, Feng Fang, Yan Gang, Hao Xiao-Ke. Cross-domain person re-identification on adaptive fusion network. Acta Automatica Sinica, 2022, 48(11): 2744?2756 doi: 10.16383/j.aas.c220083
[8]	梁文琦, 王广聪, 赖剑煌. 基于多对多生成对抗网络的非对称跨域迁移行人再识别. 自动化学报, 2022, 48(1): 103-120 doi: 10.16383/j.aas.c190303. Liang Wen-Qi, Wang Guang-Cong, Lai Jian-Huang. Asymmetric cross-domain transfer learning of person re-identification based on the many-to-many generative adversarial network. Acta Automatica Sinica, 2022, 48(1): 103?120 doi: 10.16383/j.aas.c190303.
[9]	Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks[J]. Advances in neural information processing systems, 2015, 28.
[10]	Chen Y, Li W, Sakaridis C, et al. Domain adaptive faster r-cnn for object detection in the wild[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 3339−3348.
[11]	Saito K, Ushiku Y, Harada T, et al. Strong-Weak distribution alignment for adaptive object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 6956−6965.
[12]	Lin T Y, Goyal P, Girshick R, et al. Focal Loss for Dense Object Detection[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, PP(99): 2999-3007.
[13]	Shen Z, Maheshwari H, Yao W, et al. Scl: Towards accurate domain adaptive object detection via gradient detach based stacked complementary losses[J]. arXiv preprint arXiv: 1911.02559, 2019.
[14]	Zheng Y, Huang D, Liu S, et al. Cross-domain Object Detection through Coarse-to-Fine Feature Adaptation: IEEE, 10.1109/CVPR42600.2020.01378[P]. 2020.
[15]	Xu C D, Zhao X R, Jin X, et al. Exploring categorical regularization for domain adaptive object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 11724−11733.
[16]	HSU C C, TSAI Y H, LIN Y Y, et al. Every Pixel Matters: Center-Aware Feature Alignment for Domain Adaptive Object Detector[C]//VEDALDI A, BISCHOF H, BROX T, et al. Computer Vision – ECCV 2020. Cham: Springer International Publishing, 2020: 733−748.
[17]	Chen C, Zheng Z, Ding X, et al. Harmonizing transferability and discriminability for adapting object detectors[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 8869−8878.
[18]	Zhu J Y, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2223−2232.
[19]	Deng J, Li W, Chen Y, et al. Unbiased mean teacher for cross-domain object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021: 4091−4101.
[20]	Xu M, Wang H, Ni B, et al. Cross-domain detection via graph-induced prototype alignment[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 12355−12364.
[21]	Wu A, Liu R, Han Y, et al. Vector-decomposed disentanglement for domain-invariant object detection[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021: 9342−9351.
[22]	Chen C, Zheng Z, Huang Y, et al. I3net: Implicit instance-invariant network for adapting one-stage object detectors[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021: 12576−12585.
[23]	李威, 王蒙. 基于渐进多源域迁移的无监督跨域目标检测[J]. 自动化学报, 2022, 48(8): 1-15. doi: 10.16383/j.aas.c190532 Li Wei, Wang Meng. Unsupervised cross-domain object detection based on progressive multi-source transfer. Acta Automatica Sinica, 2022, 48(9): 2337?2351 doi: 10.16383/j.aas.c190532.
[24]	A. L. Rodriguez and K. Mikolajczyk, "Domain adaptation for object detection via style consistency," British Machine Vision Conference, 2019.
[25]	Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]//European conference on computer vision. Springer, Cham, 2016: 21−37.
[26]	Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 779−788.
[27]	Yolov8 [Online], available: https://github.com/ultralytics/yolov8, Feb 15, 2023
[28]	Zhang S, Tuo H, Hu J, et al. Domain Adaptive YOLO for One-Stage Cross-Domain Detection[C]//Asian Conference on Machine Learning. PMLR, 2021: 785−797.
[29]	Redmon J, Farhadi A. Yolov3: An incremental improvement[J]. arXiv preprint arXiv: 1804.02767, 2018.
[30]	HNEWA M, RADHA H. Integrated Multiscale Domain Adaptive YOLO[J]. arXiv: 2202.03527, 2022
[31]	Bochkovskiy A, Wang C Y, Liao H Y M. Yolov4: Optimal speed and accuracy of object detection[J]. arXiv preprint arXiv: 2004.10934, 2020.
[32]	Vidit V, Salzmann M. Attention-based domain adaptation for single-stage detectors[J]. Machine Vision and Applications, 2022, 33(5): 65. doi: 10.1007/s00138-022-01320-y
[33]	Yolov5[Online], available: https://github.com/ultralytics/yolov5, Nov 28, 2022
[34]	LI G, JI Z, QU X, et al. Cross-Domain Object Detection for Autonomous Driving: A Stepwise Domain Adaptative YOLO Approach[J]. IEEE Transactions on Intelligent Vehicles, 2022: 1-1.
[35]	Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 7132−7141
[36]	Wang Q, Wu B, Zhu P, et al. ECA-Net: Efficient channel attention for deep convolutional neural networks[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 11534−11542.
[37]	Lee H J, Kim H E, Nam H. Srm: A style-based recalibration module for convolutional neural networks[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 1854−1862.
[38]	Wang M, Wang W, Li B, et al. Interbn: Channel fusion for adversarial unsupervised domain adaptation[C] //Proceedings of the 29th ACM international conference on multimedia. 2021: 3691−3700.
[39]	Ding S, Lin L, Wang G, et al. Deep feature learning with relative distance comparison for person re-identification[J]. Pattern Recognition, 2015, 48(10): 2993-3003. doi: 10.1016/j.patcog.2015.04.005
[40]	Snell J, Swersky K, Zemel R. Prototypical networks for few-shot learning[J]. Advances in neural information processing systems, 2017, 30.
[41]	He K, Fan H, Wu Y, et al. Momentum contrast for unsupervised visual representation learning[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 9729−9738.
[42]	Cordts M, Omran M, Ramos S, et al. The cityscapes dataset for semantic urban scene understanding[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 3213−3223.
[43]	Sakaridis C, Dai D, Van Gool L. Semantic foggy scene understanding with synthetic data[J]. International Journal of Computer Vision, 2018, 126(9): 973-992. doi: 10.1007/s11263-018-1072-8
[44]	Yu F, Chen H, Wang X, et al. Bdd100k: A diverse driving dataset for heterogeneous multitask learning[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 2636−2645.
[45]	A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11): 1231-1237, 2013. doi: 10.1177/0278364913491297
[46]	Johnson-Roberson M, Barto C, Mehta R, et al. Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?[J]. arXiv preprint arXiv: 1610.01983, 2016.
[47]	Everingham M, Van Gool L, Williams C K I, et al. The pascal visual object classes (voc) challenge[J]. International journal of computer vision, 2009, 88: 303-308.
[48]	Inoue N, Furuta R, Yamasaki T, et al. Cross-domain weakly-supervised object detection through progressive domain adaptation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 5001−5009.