PaDiM-EfficientNetV2

Both PaDiM-EfficientNetV2 and PaDiM-EfficentNet_NS outperform PaDiM-WR50-Rd550 from original implementation on image-level ROCAUC significantly by over 2% !!!

Features

• The current SOTA from EfficentNet family, EfficientNetV2 and EfficentNet_NS(Noisy Student), are used as the feature extractor
• The cov_inv instead of cov is saved to speed up the calculation of mahalanobis distance
• Optimize both torch and numpy methods to calculate mahalanobis distance
• Save train params, mean and cov_inv, in hdf5 format instead of pickle in case of OOM issue when memory is limited
• Add PCA and NPCA to reduce feature dimensionality
• Pretrained feature distribution of efficientnetv2_m model is shared below

Requirement

• Packages
  • python == 3.8.5
  • pytorch == 1.12 (lower version might work but not verified)
  • tqdm
  • sklearn
  • matplotlib
  • timm
• Both EfficientNetV2 and EfficientNet_NS models are imported from timm (pytorch-image-models)

Datasets

MVTec AD datasets : Download from MVTec website

Usage

python main.py --data_path mvtec_anomaly_detection/ --save_path ./mvtec_result --arch efficientnetv2_m_in21ft1k --use_gpu
python main.py --data_path mvtec_anomaly_detection/ --save_path ./mvtec_result --arch efficientnet_b7_ns --use_gpu

Download pre-trained feature distributions

Download pretrained feature distribution of efficientnetv2_m model from dropbox link efficientnetv2_m feature distribution

Results

Implementation results on MVTec

• Image-level anomaly detection accuracy (ROCAUC)

MvTec	R18-Rd100	WR50-Rd550	Effi-B7-NS	EffiV2-M
Carpet	0.984	0.999	1.0	1.0
Grid	0.898	0.957	0.997	0.984
Leather	0.988	1.0	1.0	1.0
Tile	0.959	0.974	0.991	0.991
Wood	0.990	0.988	0.995	0.996
All texture classes	0.964	0.984	0.997	0.994
Bottle	0.996	0.998	1.0	1.0
Cable	0.855	0.922	0.961	0.966
Capsule	0.870	0.915	0.913	0.936
Hazelnut	0.841	0.933	0.960	0.992
Metal nut	0.974	0.992	0.988	0.995
Pill	0.869	0.944	0.974	0.974
Screw	0.745	0.844	0.956	0.944
Toothbrush	0.947	0.972	0.981	0.936
Transistor	0.925	0.978	0.999	0.995
Zipper	0.741	0.909	0.946	0.958
All object classes	0.876	0.941	0.968	0.970
All classes	0.905	0.955	0.977	0.978

• Pixel-level anomaly detection accuracy (ROCAUC)

MvTec	R18-Rd100	WR50-Rd550	Effi-B7-NS	EffiV2-M
Carpet	0.988	0.990	0.984	0.985
Grid	0.936	0.965	0.967	0.965
Leather	0.990	0.989	0.986	0.985
Tile	0.917	0.939	0.923	0.941
Wood	0.940	0.941	0.928	0.924
All texture classes	0.953	0.965	0.958	0.96
Bottle	0.981	0.982	0.971	0.976
Cable	0.949	0.968	0.973	0.975
Capsule	0.982	0.986	0.979	0.986
Hazelnut	0.979	0.979	0.977	0.984
Metal nut	0.967	0.971	0.954	0.958
Pill	0.946	0.961	0.944	0.965
Screw	0.972	0.983	0.989	0.983
Toothbrush	0.986	0.987	0.979	0.986
Transistor	0.968	0.975	0.981	0.983
Zipper	0.976	0.984	0.968	0.962
All object classes	0.971	0.978	0.972	0.976
All classes	0.965	0.973	0.967	0.971

ROC Curve

• ResNet18

• Wide_ResNet50_2

• Efficientnet-b7_ns

• EfficientnetV2_m

Typical Localization examples

Reference

[1] Thomas Defard, Aleksandr Setkov, Angelique Loesch, Romaric Audigier. PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization. https://arxiv.org/pdf/2011.08785

[2] Base code. https://github.com/xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master

[3] The distance of Mahalanobis is replaced by a tensor (attach GPU) base to reduce the inference time. https://github.com/ingbeeedd/PaDiM-EfficientNet