This repository contains the implementation code for paper:

Domain-Specific Bias Filtering for Single Labeled Domain Generalization

Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, Lanfen Lin

International Journal of Computer Vision (IJCV), 2022

[arXiv]

Domain generalization (DG) utilizes multiple labeled source datasets to train a generalizable model for unseen target domains. However, due to expensive annotation costs, the requirements of labeling all the source data are hard to be met in real-world applications.

We investigate a Single Labeled Domain Generalization (SLDG) task with only one source domain being labeled, which is more practical and challenging than the Conventional Domain Generalization (CDG). A major obstacle in the SLDG task is the discriminability-generalization bias: discriminative information in the labeled source dataset may contain domain-specific bias, constraining the generalization of the trained model.

To tackle this challenging task, we propose Domain-Specific Bias Filtering (DSBF), which initializes a discriminative model with the labeled source data and filters out its domain-specific bias with the unlabeled source data for generalization improvement. We divide the filtering process into (1) feature extractor debiasing using k-means clustering-based semantic feature re-extraction and (2) classifier calibrating using attention-guided semantic feature projection.

You may need to build suitable Python environment by installing the following packages (Anaconda is recommended).

• python 3.6
• pytorch 1.7.1 (with cuda 11.0 and cudnn 8.0)
• torchvision 0.8.2
• tensorboardx 2.1
• numpy 1.19

Device:

• GPU with VRAM >5GB (strictly).
• Memory >8GB.

We list the adopted datasets in the following.

Please note:

• Office-Caltech-Home dataset is constructed by choosing the common classes from Office-Caltech [3] and Office-Home [2] datasets, please see our paper for more details.
• Our dataset split follows previous works like RSC (Code) [4].
• Although these datasets are open-sourced, you may need to have permission to use the datasets under the datasets' license.
• If you're a dataset owner and do not want your dataset to be included here, please get in touch with us via a GitHub issue. Thanks!

1. Prepare the datasets.
2. Update the .txt files under folder DSBF/dataset/pthList/ with your dataset path.
3. Run the code with command:

nohup sh run.sh > run.txt 2>&1 &

4. Check results in DSBF/dataset-task-target-data.txt.

• [11/08 2022] Our paper is accepted by International Journal of Computer Vision (IJCV).
• [12/27/2021] We uploaded a new arXiv version. See new arXiv version.

If you find our code or idea useful for your research, please cite our work.

@article{yuan2021domain,
  title={Domain-Specific Bias Filtering for Single Labeled Domain Generalization},
  author={Yuan, Junkun and Ma, Xu and Chen, Defang and Kuang, Kun and Wu, Fei and Lin, Lanfen},
  journal={arXiv preprint arXiv:2110.00726},
  year={2021}
}

If you have any questions, feel free to contact us through email (yuanjk@zju.edu.cn or maxu@zju.edu.cn) or GitHub issues. Thanks!

[1] Li, Da, et al. "Deeper, broader and artier domain generalization." Proceedings of the IEEE international conference on computer vision. 2017.

[2] Venkateswara, Hemanth, et al. "Deep hashing network for unsupervised domain adaptation." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.

[3] Saenko, Kate, et al. "Adapting visual category models to new domains." European conference on computer vision. Springer, Berlin, Heidelberg, 2010.

[4] Huang, Zeyi, et al. "Self-challenging improves cross-domain generalization." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. Springer International Publishing, 2020.