OT-VP: Optimal Transport-guided Visual Prompting for Test-Time Adaptation

OT-VP: Optimal Transport-guided Visual Prompting for Test-Time Adaptation

Yunbei Zhang, Akshay Mehra, Jihun Hamm

Overview

Illustration of OT-VP

Requirements

pip install -r requirements.txt

ImageNet-C Experiments

We use the ImageNet pre-trained ViT model from timm. ImageNet-C can be downloaded here.
Corruption can be chosen from 0 to 14, corresponding to 'gaussian_noise', 'shot_noise', 'impulse_noise', 'defocus_blur', 'glass_blur', 'motion_blur', 'zoom_blur', 'snow', 'frost', 'fog', 'brightness', 'contrast', 'elastic_transform', 'pixelate', 'jpeg_compression' respectively.

python -m domainbed.scripts.adapt --dataset ImageNetC --data_dir [path/to/ImageNet-C] --algorithm OTVP --corruption [0-14]

Citation

Please cite our work if you find it useful.

@article{zhang2024ot,
  title={OT-VP: Optimal Transport-guided Visual Prompting for Test-Time Adaptation},
  author={Zhang, Yunbei and Mehra, Akshay and Hamm, Jihun},
  journal={arXiv preprint arXiv:2407.09498},
  year={2024}
}

Acknowlegdement

DomainBed code is heavily used.
DoPrompt is used to implement Visual Prompting.