This is a PyTorch implementation of Multi-Label Out-of-Distribution Detection with Spectral Normalized Joint Energy by Yihan Mei, Xinyu Wang, Dell Zhang, Xiaoling Wang. Code is modified from JointEnergy, ODIN, Outlier Exposure, and deep Mahalanobis detector.
Our experimental configuration of in-distribution and out-of-distribution datasets are almost identical with JointEnergy.
Put PASCAL-VOC under ./Pascal/
folder, and put Texture under ./dtd/
folder.
Train the ResNet model for PASCAL-VOC dataset
python train.py --arch resnet101 --dataset pascal --save_dir ./save_models/
To reproduce the SNoJoE score for PASCAL-VOC dataset, please run:
python eval.py --arch resnet101 --dataset pascal --ood_data imagenet --ood energy --method sum
OoD detection performance comparison using SNoJoE vs. competitive baselines.
@misc{mei2024multilabel,
title={Multi-Label Out-of-Distribution Detection with Spectral Normalized Joint Energy},
author={Yihan Mei and Xinyu Wang and Dell Zhang and Xiaoling Wang},
year={2024},
eprint={2405.04759},
archivePrefix={arXiv},
primaryClass={cs.CV}
}