aodongli / zero-shot-ad-via-batch-norm

Official code repository for paper Zero-Shot Anomaly Detection via Batch Normalization (NeurIPS 2023).

Zero-Shot Anomaly Detection via Batch Normalization

Official code repository for NeurIPS 2023 paper Zero-Shot Anomaly Detection via Batch Normalization.

Code for different datasets is shown in the folder names. Refer to each folder for the datasets of interest.

• MVTec AD
• AnoShift, CIFAR100-C, Omniglot

Package requirements are listed in each folder's requirements.txt. Run pip install -r requireme.txt to install all packages.

Brief Overview

We introduce a straightforward yet powerful approach to train an out-of-box deep anomaly detector into a zero-shot anomaly detector. This method requires minimal configurations: 1) ensure that the deep model is set for batch-level prediction and 2) maintain all batch normalization layers in the training mode during inference. Below, you'll find a step-by-step comparison with the traditional stationary anomaly detection framework. Key configurations are color-highlighted for clarity. More details can be found in the full paper.

---

@inproceedings{acr,
  title={Zero-Shot Anomaly Detection via Batch Normalization},
  author={Li, Aodong and Qiu, Chen and Kloft, Marius and Smyth, Padhraic and Rudolph, Maja and Mandt, Stephan},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}