STPM-Anomaly-Detection-Localization-master

This is an implementation of the paper Student-Teacher Feature Pyramid Matching for Unsupervised Anomaly Detection.

Datasets

MVTec AD datasets : Download from MVTec website

Environment

pip install -r requirements.txt

Usage

python main.py --phase 'train or test' --dataset_path 'D:/dataset/mvtec_anomaly_detection' --save_path 'path\to\save\results' --obj 'class name'

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	0.988(1)	-	0.999(1)
grid	0.990	0.980(1)	-	0.925(1)
leather	0.993	0.989(1)	-	1.0(1)
tile	0.974	0.919(1)	-	0.979(1)
wood	0.972	0.926(1)	-	0.988(1)
bottle	0.988	0.973(1)	-	0.993(1)
cable	0.955	0.971(1)	-	0.995(1)
capsule	0.983	0.963(1)	-	0.818(1)
hazelnut	0.985	0.971(1)	-	0.975(1)
metal nut	0.976	0.963(1)	-	0.995(1)
pill	0.978	0.934(1)	-	0.887(1)
screw	0.983	0.961(1)	-	0.806(1)
toothbrush	0.989	0.978(1)	-	0.989(1)
transistor	0.825	0.921(1)	-	0.978(1)
zipper	0.985	0.969(1)	-	0.899(1)
mean	0.970	0.960(1)	0.955	0.948(1)

Visualization examples

Acknowledgement

The code is partially adapted from STPM_anomaly_detection