Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation

Subhankar Roy; Evgeny Krivosheev; Zhun Zhong; Nicu Sebe; Elisa Ricci

doi:10.48550/arXiv.2104.00808

Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation

Subhankar Roy, Evgeny Krivosheev, Zhun Zhong^* (Corresponding Author), Nicu Sebe, Elisa Ricci

^*Corresponding author for this work

Computing Science

Research output: Chapter in Book/Report/Conference proceeding › Published conference contribution

72 Citations (Scopus)

Abstract

In this paper we address multi-target domain adaptation (MTDA), where given one labeled source dataset and multiple unlabeled target datasets that differ in data distributions, the task is to learn a robust predictor for all the target domains. We identify two key aspects that can help to alleviate multiple domain-shifts in the MTDA: feature aggregation and curriculum learning. To this end, we propose Curriculum Graph Co-Teaching (CGCT) that uses a dual classifier head, with one of them being a graph convolutional network (GCN) which aggregates features from similar samples across the domains. To prevent the classifiers from over-fitting on its own noisy pseudo-labels we develop a co-teaching strategy with the dual classifier head that is assisted by curriculum learning to obtain more reliable pseudo-labels. Furthermore, when the domain labels are available, we propose Domain-aware Curriculum Learning (DCL), a sequential adaptation strategy that first adapts on the easier target domains, followed by the harder ones. We experimentally demonstrate the effectiveness of our proposed frameworks on several benchmarks and advance the state-of-the-art in the MTDA by large margins (e.g. +5.6% on the DomainNet).

Original language	English
Title of host publication	Proceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021
Publisher	IEEE Computer Society
Pages	5347-5356
Number of pages	10
ISBN (Electronic)	9781665445092
ISBN (Print)	9781665445092, 9781665445108
DOIs	https://doi.org/10.48550/arXiv.2104.00808 https://doi.org/10.1109/CVPR46437.2021.00531
Publication status	Published - 2 Nov 2021
Event	2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021 - Virtual, Online, United States Duration: 19 Jun 2021 → 25 Jun 2021

Publication series

Name	Conference on Computer Vision and Pattern Recognition (CVPR)
Publisher	IEEE
ISSN (Print)	1063-6919
ISSN (Electronic)	2575-7075

Conference

Conference	2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021
Country/Territory	United States
City	Virtual, Online
Period	19/06/21 → 25/06/21

Bibliographical note

Funding Information:
To address multi-target domain adaptation (MTDA), we proposed Curriculum Graph Co-Teaching (CGCT) that uses a graph convolutional network to perform robust feature aggregation across multiple domains, which is then trained with a co-teaching and curriculum learning strategy. To better exploit domain labels of the target we presented a Domain-aware curriculum (DCL) learning strategy that adapts easier target domains first and harder later, enabling a smoother feature alignment. Through extensive experiments we demonstrate that our proposed contributions handsomely outperform the state-of-the-art in the MTDA. Acknowledgements This work is supported by the EU H2020 SPRING No. 871245 and AI4Media No. 951911 projects; the Italy-China collaboration project TALENT:2018YFE0118400; and the Caritro Deep Learning Lab of the ProM Facility of Rovereto.

Funding

This work is supported by the EU H2020 SPRING No. 871245 and AI4Media No. 951911 projects; the Italy-China collaboration project TALENT:2018YFE0118400; and the Caritro Deep Learning Lab of the ProM Facility of Rovereto.

Funders	Funder number
H2020 European Research Council	871245
AI4Media	951911

Access to Document

10.48550/arXiv.2104.00808Licence: CC BY
10.1109/CVPR46437.2021.00531Licence: Unspecified

Cite this

Roy, S., Krivosheev, E., Zhong, Z., Sebe, N., & Ricci, E. (2021). Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation. In Proceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021 (pp. 5347-5356). (Conference on Computer Vision and Pattern Recognition (CVPR)). IEEE Computer Society. https://doi.org/10.48550/arXiv.2104.00808, https://doi.org/10.1109/CVPR46437.2021.00531

Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation. / Roy, Subhankar; Krivosheev, Evgeny; Zhong, Zhun (Corresponding Author) et al.
Proceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021. IEEE Computer Society, 2021. p. 5347-5356 (Conference on Computer Vision and Pattern Recognition (CVPR)).

Research output: Chapter in Book/Report/Conference proceeding › Published conference contribution

Roy, S, Krivosheev, E, Zhong, Z, Sebe, N & Ricci, E 2021, Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation. in Proceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021. Conference on Computer Vision and Pattern Recognition (CVPR), IEEE Computer Society, pp. 5347-5356, 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021, Virtual, Online, United States, 19/06/21. https://doi.org/10.48550/arXiv.2104.00808, https://doi.org/10.1109/CVPR46437.2021.00531

Roy S, Krivosheev E, Zhong Z, Sebe N, Ricci E. Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation. In Proceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021. IEEE Computer Society. 2021. p. 5347-5356. (Conference on Computer Vision and Pattern Recognition (CVPR)). Epub 2021 Nov 2. doi: 10.48550/arXiv.2104.00808, 10.1109/CVPR46437.2021.00531

@inproceedings{50f58f12d21748a5a9fa901b4467e35f,

title = "Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation",

abstract = "In this paper we address multi-target domain adaptation (MTDA), where given one labeled source dataset and multiple unlabeled target datasets that differ in data distributions, the task is to learn a robust predictor for all the target domains. We identify two key aspects that can help to alleviate multiple domain-shifts in the MTDA: feature aggregation and curriculum learning. To this end, we propose Curriculum Graph Co-Teaching (CGCT) that uses a dual classifier head, with one of them being a graph convolutional network (GCN) which aggregates features from similar samples across the domains. To prevent the classifiers from over-fitting on its own noisy pseudo-labels we develop a co-teaching strategy with the dual classifier head that is assisted by curriculum learning to obtain more reliable pseudo-labels. Furthermore, when the domain labels are available, we propose Domain-aware Curriculum Learning (DCL), a sequential adaptation strategy that first adapts on the easier target domains, followed by the harder ones. We experimentally demonstrate the effectiveness of our proposed frameworks on several benchmarks and advance the state-of-the-art in the MTDA by large margins (e.g. +5.6% on the DomainNet).",

author = "Subhankar Roy and Evgeny Krivosheev and Zhun Zhong and Nicu Sebe and Elisa Ricci",

note = "Funding Information: To address multi-target domain adaptation (MTDA), we proposed Curriculum Graph Co-Teaching (CGCT) that uses a graph convolutional network to perform robust feature aggregation across multiple domains, which is then trained with a co-teaching and curriculum learning strategy. To better exploit domain labels of the target we presented a Domain-aware curriculum (DCL) learning strategy that adapts easier target domains first and harder later, enabling a smoother feature alignment. Through extensive experiments we demonstrate that our proposed contributions handsomely outperform the state-of-the-art in the MTDA. Acknowledgements This work is supported by the EU H2020 SPRING No. 871245 and AI4Media No. 951911 projects; the Italy-China collaboration project TALENT:2018YFE0118400; and the Caritro Deep Learning Lab of the ProM Facility of Rovereto. ; 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021 ; Conference date: 19-06-2021 Through 25-06-2021",

year = "2021",

month = nov,

day = "2",

doi = "10.48550/arXiv.2104.00808",

language = "English",

isbn = "9781665445092",

series = "Conference on Computer Vision and Pattern Recognition (CVPR)",

publisher = "IEEE Computer Society",

pages = "5347--5356",

booktitle = "Proceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021",

}

TY - GEN

T1 - Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation

AU - Roy, Subhankar

AU - Krivosheev, Evgeny

AU - Zhong, Zhun

AU - Sebe, Nicu

AU - Ricci, Elisa

N1 - Funding Information: To address multi-target domain adaptation (MTDA), we proposed Curriculum Graph Co-Teaching (CGCT) that uses a graph convolutional network to perform robust feature aggregation across multiple domains, which is then trained with a co-teaching and curriculum learning strategy. To better exploit domain labels of the target we presented a Domain-aware curriculum (DCL) learning strategy that adapts easier target domains first and harder later, enabling a smoother feature alignment. Through extensive experiments we demonstrate that our proposed contributions handsomely outperform the state-of-the-art in the MTDA. Acknowledgements This work is supported by the EU H2020 SPRING No. 871245 and AI4Media No. 951911 projects; the Italy-China collaboration project TALENT:2018YFE0118400; and the Caritro Deep Learning Lab of the ProM Facility of Rovereto.

PY - 2021/11/2

Y1 - 2021/11/2

N2 - In this paper we address multi-target domain adaptation (MTDA), where given one labeled source dataset and multiple unlabeled target datasets that differ in data distributions, the task is to learn a robust predictor for all the target domains. We identify two key aspects that can help to alleviate multiple domain-shifts in the MTDA: feature aggregation and curriculum learning. To this end, we propose Curriculum Graph Co-Teaching (CGCT) that uses a dual classifier head, with one of them being a graph convolutional network (GCN) which aggregates features from similar samples across the domains. To prevent the classifiers from over-fitting on its own noisy pseudo-labels we develop a co-teaching strategy with the dual classifier head that is assisted by curriculum learning to obtain more reliable pseudo-labels. Furthermore, when the domain labels are available, we propose Domain-aware Curriculum Learning (DCL), a sequential adaptation strategy that first adapts on the easier target domains, followed by the harder ones. We experimentally demonstrate the effectiveness of our proposed frameworks on several benchmarks and advance the state-of-the-art in the MTDA by large margins (e.g. +5.6% on the DomainNet).

AB - In this paper we address multi-target domain adaptation (MTDA), where given one labeled source dataset and multiple unlabeled target datasets that differ in data distributions, the task is to learn a robust predictor for all the target domains. We identify two key aspects that can help to alleviate multiple domain-shifts in the MTDA: feature aggregation and curriculum learning. To this end, we propose Curriculum Graph Co-Teaching (CGCT) that uses a dual classifier head, with one of them being a graph convolutional network (GCN) which aggregates features from similar samples across the domains. To prevent the classifiers from over-fitting on its own noisy pseudo-labels we develop a co-teaching strategy with the dual classifier head that is assisted by curriculum learning to obtain more reliable pseudo-labels. Furthermore, when the domain labels are available, we propose Domain-aware Curriculum Learning (DCL), a sequential adaptation strategy that first adapts on the easier target domains, followed by the harder ones. We experimentally demonstrate the effectiveness of our proposed frameworks on several benchmarks and advance the state-of-the-art in the MTDA by large margins (e.g. +5.6% on the DomainNet).

UR - http://www.scopus.com/inward/record.url?scp=85121333970&partnerID=8YFLogxK

UR - https://openaccess.thecvf.com/content/CVPR2021/papers/Roy_Curriculum_Graph_Co-Teaching_for_Multi-Target_Domain_Adaptation_CVPR_2021_paper.pdf

UR - https://openaccess.thecvf.com/content/CVPR2021/supplemental/Roy_Curriculum_Graph_Co-Teaching_CVPR_2021_supplemental.pdf

U2 - 10.48550/arXiv.2104.00808

DO - 10.48550/arXiv.2104.00808

M3 - Published conference contribution

AN - SCOPUS:85121333970

SN - 9781665445092

SN - 9781665445108

T3 - Conference on Computer Vision and Pattern Recognition (CVPR)

SP - 5347

EP - 5356

BT - Proceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021

PB - IEEE Computer Society

T2 - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021

Y2 - 19 June 2021 through 25 June 2021

ER -

Curriculum Graph Co-Teaching for Multi-Target Domain Adaptation

Abstract

Publication series

Conference

Bibliographical note

Funding

Access to Document

Other files and links

Fingerprint

Cite this