remmarp / padim-tf

[TF 2.x] PaDiM - unofficial tensorflow implementation of the paper 'a Patch Distribution Modeling Framework for Anomaly Detection and Localization'.

[TF 2.x] PaDiM - Anomaly Detection Localization

This repository contains an unofficial PaDiM implementation using tensorflow.

Paper

PaDiM: a patch distribution modeling framework for anomaly detection and localization. [Link]

Dependencies

• Windows 10, Python 3.8.8, Tensorflow 2.4.1 GPU
• Scikit-learn, Scikit-image, Matplotlib

Run

# options: seed, rd, target, batch_size, is_plot, net
python main.py

Dataset

MVTecAD dataset

Results (AU ROC)

Implementation results on MVTec

• Network Type:

  • PyTorch.: WideResNet50, Rd 550 (from PyTorch version) (WR50-Rd550)
  • Net 1: EfficientNetB7 [layer a_expand_activation 5, 6, 7], Rd 1000 (ENB7-Rd1000)
  • Net 2: EfficientNetB7 [layer a_expand_activation 4, 6, 7], Rd 1000 (ENB7-Rd1000)
  • Net 3: EfficientNetB7 [layer a_activation 5, 6, 7], Rd 1000 (ENB7-Rd1000)

• I observed that intermediate layers selection has some effects on detection performance.

• Besides, a high image-level au-roc does not guarantee a high level of au-roc on patch-level.

MvTec	PyTorch (Img)	Net 1 (Img)	Net 2 (Img)	Net 3 (Img)
carpet	0.999	0.950	0.982	0.996
grid	0.957	0.936	0.971	0.976
leather	1.000	0.999	1.000	1.000
tile	0.974	0.957	0.984	0.981
wood	0.988	0.948	0.954	0.990
bottle	0.998	0.983	0.996	0.999
cable	0.922	0.909	0.919	0.973
capsule	0.915	0.946	0.953	0.958
hazelnut	0.933	0.983	0.973	0.997
metal_nut	0.992	0.869	0.930	0.931
pill	0.944	0.882	0.879	0.925
screw	0.844	0.632	0.767	0.895
toothbrush	0.972	0.767	0.972	0.811
transistor	0.978	0.930	0.949	0.975
zipper	0.909	0.980	0.986	0.990
Avg. (tex.)	0.9840	0.9579	0.9781	0.9885
Avg. (obj.)	0.9410	0.8881	0.9323	0.9455
Avg. (all)	0.9550	0.9114	0.9476	0.9598

MvTec	org. (Patch)	Net 1 (Patch)	Net 2 (Patch)	Net 3(Patch)
carpet	0.990	0.973	0.854	0.829
grid	0.965	0.958	0.750	0.768
leather	0.989	0.986	0.902	0.831
tile	0.939	0.905	0.729	0.748
wood	0.941	0.946	0.831	0.814
bottle	0.982	0.971	0.861	0.831
cable	0.968	0.963	0.815	0.843
capsule	0.986	0.977	0.940	0.911
hazelnut	0.979	0.965	0.876	0.834
metal_nut	0.971	0.986	0.926	0.926
pill	0.961	0.955	0.893	0.903
screw	0.983	0.986	0.941	0.893
toothbrush	0.983	0.979	0.937	0.864
transistor	0.987	0.977	0.958	0.958
zipper	0.975	0.965	0.840	0.814
Avg. (tex.)	0.9650	0.9536	0.8131	0.7979
Avg. (obj.)	0.9780	0.9724	0.8987	0.8776
Avg. (all)	0.9730	0.9661	0.8702	0.8510

ROC Curve (Net 1) Bottle

PR Curve (Net 1) Bottle

Localization examples (Net 1) (cherry-picked)