This project is an unofficial implementation of "EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies", which is implemented step-by-step according to the pseudocode in the appendix

Datasets

./data

• ImageNet

  • n01440764
  • n01443537 ...

• MVTec_AD

  • bottle
    • ground_truth
    • test
    • train
  • cable
    • ground_truth
    • test
    • train ...

• result

Quick start

1. Install PyTorch environment

conda activate <your_env>
pip install -r requirements.txt

1. Distill a PDN architecture teacher network from wide_resnet101

python distillaion_training.py

2. train the student network and autoencoder network

python train_reduced_student.py

Some results

MVTec bottle

image_AUROC: 1.0
image_F1Score: 1.0
pixel_AUROC: 0.9876494407653809
pixel_F1Score: 0.7927650213241577