/ET-OOD

CVPR2023:Uncertainty-Aware Optimal Transport for Semantically Coherent Out-of-Distribution Detection

Primary LanguagePythonMIT LicenseMIT

Uncertainty-Aware Optimal Transport for Semantically Coherent Out-of-Distribution Detection （CVPR 2023）

SCOOD benchmarks download link:
gdrive  onedrive

The codebase accesses the SCOOD benchmarks from the root directory in a folder named data/ by default, i.e.

├── ...
├── data
│   ├── images
│   └── imglist
├── scood
├── test.py
├── train.py
├── ...

Dependencies

  • Python >= 3.8
  • Pytorch = 1.8.1
  • CUDA >= 11.3
  • torchvision=0.9.1
  • faiss-gpu=1.7.1

Experimental Results

You can run the following script (specifying the output and data directories) which perform training & testing for CIFAR10/100 experimental results:

bash cifar10.sh output_dir data_dir
bash cifar100.sh output_dir data_dir

The information during training can be monitored in real-time in output_dir/log.txt and the results will be saved in output_dir/results.csv.

Acknowledgments

This paper follows the excellent work from SCOOD.

Cite

If our work is useful for your research, please consider citing our paper :

@inproceedings{lu2022etood,
  title={Uncertainty-Aware Optimal Transport for Semantically Coherent Out-of-Distribution Detection},
  author={Fan Lu, Kai Zhu, Wei Zhai, Kecheng Zheng and Yang Cao},
  booktitle={CVPR},
  year={2023}
}