PyTorch re-implementation of PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization

1. AUROC Scores

category	Paper	My Implementation
zipper	0.985	0.923
wood	0.949	0.992
transistor	0.975	0.998
toothbrush	0.988	0.883
tile	0.941	0.994
screw	0.985	0.815
pill	0.957	0.958
metal_nut	0.972	0.992
leather	0.992	1.000
hazelnut	0.982	0.985
grid	0.973	0.959
carpet	0.991	0.997
capsule	0.985	0.937
cable	0.967	0.930
bottle	0.983	1.000

2. PRO Scores

category	Paper	My Implementation
zipper	0.959	0.935
wood	0.911	0.891
transistor	0.845	0.949
toothbrush	0.931	0.915
tile	0.860	0.826
screw	0.944	0.936
pill	0.927	0.952
metal_nut	0.856	0.933
leather	0.978	0.978
hazelnut	0.926	0.937
grid	0.946	0.866
carpet	0.962	0.952
capsule	0.935	0.921
cable	0.888	0.918
bottle	0.948	0.951

3. Graphical Results

zipper

wood

transistor

toothbrush

tile

screw

pill

metal_nut

leather

hazelnut

grid

carpet

capsule

cable

bottle

4. Requirements

• CUDA 10.2
• nvidia-docker2

5. Usage

a) Download docker image and run docker container

docker pull taikiinoue45/mvtec:padim
docker run --runtime nvidia -it taikiinoue45/mvtec:padim /bin/bash

b) Run Experiment

python run.py ./config.yaml params.category=bottle params.tracking_uri=file:///app/PaDiM/mlruns

6. Contacts

• github: https://github.com/taikiinoue45/
• twitter: https://twitter.com/taikiinoue45/
• linkedin: https://www.linkedin.com/in/taikiinoue45/