Detecting Out-of-Distribution through the Lens of Neural Collapse

This repository contains code for the paper Detecting Out-of-Distribution through the Lens of Neural Collapse (CVPR 2025) by Litian Liu and Yao Qin. The codebase is adapted from and integrated into the OpenOOD Benchmark.

Explore related work:

Setup

Please follow OpenOOD official instruction to complete the setup.

pip install git+https://github.com/Jingkang50/OpenOOD

Demo

CIFAR-10 ResNet-18 Benchmark

python scripts/eval_ood.py \
     --id-data cifar10 \
     --root ./results/cifar10_resnet18_32x32_base_e100_lr0.1_default \
     --postprocessor nci \
     --save-score --save-csv

ImageNet ResNet-50 Benchmark

python scripts/eval_ood_imagenet.py \
     --tvs-pretrained \
     --arch resnet50 \
     --postprocessor nci \
     --save-score --save-csv

Citation

Please cite our paper if you find this codebase helpful!

@article{liu2025detecting,
  title={Detecting Out-of-Distribution through the Lens of Neural Collapse},
  author={Liu, Litian and Qin, Yao},
  journal={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}