[CVPR2024] PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization

Link to our paper

Requirements

Python 3.7.11+
Pytorch 1.8.0+

Data Preparation

Download the dataset PACS, OfficeHome and DomainNet.

Arrange data with the following structure:

Path/To/Dataset
├── Domain1
      ├── cat
      ├── ......
├── Domain2
      ├── cat
      ├── ......
├── Domain3
      ├── cat
      ├── ......
├── Domain4
      ├── cat
      ├── ......
├── image_list
      List for each dataset is provided in ./image_list 

Modify the file path in the scripts.

Train and inference

For the training and inference process, please simply execute:

bash scripts/run.sh

Change test_envs to different values (e.g., 0,1,2,3) to conduct leave-one-domain-out protocol.

Acknowledgment

We thank the authors from OpenDG-Eval for reference. We modify their code to implement Hybrid Domain Generalization.

Citation

@inproceedings{chen2024practicaldg,
  title={PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization},
  author={Chen, Zining and Wang, Weiqiu and Zhao, Zhicheng and Su, Fei and Men, Aidong and Meng, Hongying},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={23501--23511},
  year={2024}
}