This project is an unofficial implementation of "EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies".
./data
-
ImageNet
- n01440764
- n01443537 ...
-
MVTec_AD
- bottle
- ground_truth
- test
- train
- cable
- ground_truth
- test
- train ...
- bottle
-
result
conda activate <your_env>
pip install -r requirements.txt
python distillaion_training.py
python train_reduced_student.py -c configs/mvtec_train.yaml
Download pretrain weights from release page.
Model | Dataset | Official Paper | ours |
---|---|---|---|
EfficientAD-M | VisA | 98.1 | 97.54 |
EfficientAD-M | Mvtec LOCO | 90.7 | pending |
EfficientAD-M | Mvtec AD | 99.1 | 99.36 |
EfficientAD-S | VisA | 97.5 | 97.20 |
EfficientAD-S | Mvtec LOCO | 90.0 | pending |
EfficientAD-S | Mvtec AD | 98.8 | 98.51 |
MVTec bottle