[NeurIPS 24] MetaUAS

Official PyTorch Implementation of MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning, NeurIPS 2024.

Updates

• [2026-02-05]: Added 1-shot MetaUAS results on Real-IAD and Real-IAD-Variety datasets.

Datasets	Methods	I-ROC	I-PR	I-F1	P-ROC	P-PR	P-F1	P-PRO
Real-IAD	MetaUAS (1-shot)	80.0 ± 0.4	77.9 ± 0.4	72.4 ± 0.4	95.6 ± 0.2	36.6 ± 1.1	39.7 ± 1.0	83.5 ± 0.7
Real-IAD-Variety	MetaUAS (1-shot)	81.9 ± 0.1	96.3 ± 0.1	94.1 ± 0.0	92.0 ± 0.1	48.2 ± 0.4	48.3 ± 0.3	76.5 ± 0.1

Introduction

MetaUAS unifies anomaly segmentation into change segmentation and provides a pure visual foundation model, which requires only one normal image prompt and no additional training, and effectively and efficiently segments any visual anomalies. MetaUAS significantly outperforms most zero-shot, few-shot, and even full-shot anomaly segmentation methods.

MetaUAS Framework

Main Results

Demo

You can use our Online Demo to test your custom data for a quick start. Note that the online demo is currently based on CPU. You could also deploy the demo application to your local CPU/GPU server using the following command:

pip install -r requirements.txt
python app.py

Evaluation on MVTec/VisA/GoodsAD

bash test.sh

ToDo List

• release pre-trained MetaUAS models
• deploy online MetaUAS Demo
• open testing code
• open training code

Citing

If you find this code useful in your research, please consider citing us:

@inproceedings{gao2024metauas,
  title  = {MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning},
  author = {Gao, Bin-Bin},
  booktitle = {Advances in Neural Information Processing Systems},
  pages = {39812--39836},
  year = {2024}
}

Star History