Official code for our CVPR 2024 paper
pip install -r requirements.txt- Download segmentation maps for VisA.
- Download anomaly maps for EfficientAD.
- Download anomaly maps for RD4AD.
- Data structure should look as following:
data
|-- visa_segm
|-- anomaly_maps
|-----|--efficient_ad
|-----|--rd4adOnly VisA dataset is available for now. List of available models: ["efficient_ad", "rd4ad", "all_ad"]. "all_ad" includes both EfficientAD and RD4AD.
python main.py --model efficient_adCl. AUROC (image-level) for SegAD with different sources of anomaly maps.
| model | mean | candle | capsules | cashew | chewinggum | fryum | macaroni1 | macaroni2 | pcb1 | pcb2 | pcb3 | pcb4 | pipe_fryum |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RD4AD + SegAD | 95.3 | 98.5 | 80.2 | 98.9 | 99.4 | 96.1 | 97.4 | 90.7 | 96.4 | 96.3 | 94.1 | 99.9 | 95.8 |
| EfficientAD + SegAD | 98.3 | 98.7 | 89.7 | 98.6 | 99.9 | 98.6 | 99.5 | 98.1 | 99.5 | 99.7 | 98.4 | 99.3 | 99.2 |
| All AD + SegAD | 98.4 | 99.0 | 90.7 | 99.0 | 99.9 | 98.5 | 99.4 | 98.1 | 99.2 | 99.7 | 98.3 | 99.8 | 99.1 |
We use EfficientAD and Anomalib for baseline anomaly detection models. We are thankful for their amazing work!
Please cite this paper if it helps your project:
@inproceedings{baitieva2024supervised,
title={Supervised Anomaly Detection for Complex Industrial Images},
author={Aimira Baitieva and David Hurych and Victor Besnier and Olivier Bernard},
booktitle={CVPR},
year={2024}
}