Student-Teacher Feature Pyramid Matching for Unsupervised Anomaly Detection implementation (unofficial)

Unofficial pytorch implementation of
Student-Teacher Feature Pyramid Matching for Unsupervised Anomaly Detection (STPM)
- Guodong Wang, Shumin Han, Errui Ding, Di Huang (2021)
https://arxiv.org/abs/2103.04257v2

update(21/06/22) : I changed entire code to lightning version. previous code is in legacy directory.

Usage

# python 3.6, torch==1.8.1, torchvision==0.9.1
pip install -r requirements.txt
python train.py --phase train or test --dataset_path ...\mvtec_anomaly_detection --category bottle --project_path path\to\save\results

MVTecAD AUC-ROC score (mean of n trials)

Category	Paper (pixel-level)	This code (pixel-level)	Paper (image-level)	This code (image-level)
carpet	0.988	0.991(1)	-	0.989(1)
grid	0.990	0.992(1)	-	1.000(1)
leather	0.993	0.988(1)	-	0.999(1)
tile	0.974	0.968(1)	-	0.955(1)
wood	0.972	0.964(1)	-	0.992(1)
bottle	0.988	0.986(1)	-	1.000(1)
cable	0.955	0.952(1)	-	0.923(1)
capsule	0.983	0.984(1)	-	0.880(1)
hazelnut	0.985	0.988(1)	-	1.000(1)
metal nut	0.976	0.973(1)	-	1.000(1)
pill	0.978	0.969(1)	-	0.938(1)
screw	0.983	0.985(1)	-	0.882(1)
toothbrush	0.989	0.988(1)	-	0.878(1)
transistor	0.825	0.820(1)	-	0.937(1)
zipper	0.985	0.986(1)	-	0.936(1)
mean	0.970	0.969(1)	0.955	0.954(1)

Localization results