• �� Mar 10, 2025. TailedCore code released.

TL;DR: We suggest a novel practical challenging anomaly detection task, noisy long-tailed anomaly detection where tail classes are unknown and head classes are contaminated. We suggest TailSampler, which first tail classes with class size estimation and denoise head classes seprately.

Performance comparison with baselines

Pipeine of TailedCore

Noise discriminative models remove tail classes(left) while greedy sampling samples both tail and noise

Ablation with noise ratio comparing with baselines

Install the required packages with the command below

bash install_packages.sh

Download MVTecAD dataset from the link, and place it at, for example, ./datasets/mvtecad. Then run the following

bash make_all_mvtecad_nlt.sh

Download VisA dataset from the link, and place it at, for example, ./datasets/visa. Then run the following

bash make_all_mvtecad_nlt.sh

After generating the noisy long-tailed dataset, run the code to train model. The configuration file for training or testing should be saved in ./configs directory.

python main.py --dataset --mvtec --noisy_lt_dataset paretno_nr0.1_seed42 --config tailedcore_mvtec

Refer the files coreset_model for the code of each models sampler for the code of each samplers

which are the core codes of our method.

The following is a BibTeX reference:

@inproceedings{jung2025tailedcore,
  title={{TailedCore}: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection},
  author={Yoon Gyo Jung and Jaewoo Park and Jaeho Yoon and Kuan-Chuan Peng and Wonchul Kim and Andrew Beng Jin Teoh and Octavia Camps},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  url={https://arxiv.org/abs/2504.02775},
  year={2025}
}

The code is based on the repository of PatchCore