SAAE-DFR

This is a code implemention for paper "Self-Attention Autoencoder for Anomaly Segmentation"

Dataset

The MVTec AD dataset is available at:

https://www.mvtec.com/company/research/datasets/mvtec-ad

Pretrained-ViT

We use pretrained ViT-B16 in the paper. You can get it at https://github.com/asyml/vision-transformer-pytorch.

Train

python main.py --mode train --device cuda:0 --batch_size 32 --epochs 700

Evaluation

python main.py --mode evaluation

Results

Class	ROC-AUC	PRO-AUC
Carpet	97.9	93.1
Grid	98.6	96.4
Leather	99.6	98.7
Tile	97.3	92.7
Wood	97.6	95.4
Mean textures	98.2	95.3
Bottle	97.9	94.3
Cable	96.8	89.0
Capsule	98.2	92.9
Hazelnut	98.5	96.6
Metal Nut	97.6	91.7
Pill	98.1	97.1
Screw	98.9	94.6
Toothbrush	98.7	93.1
Transistor	96.0	88.2
Zipper	96.9	90.3
Mean textures	97.8	92.8
Mean	97.9	93.6

Qualitative Results

The first row is input image, the second row is ground truth anomaly segmentation and the third is the anomaly score map produced by our model. The last row is the segmentation result when FPR of 0 for texture classes and FPR of 0.001 for object classes on corresponding training set are given.

Reference

https://github.com/YoungGod/DFR

https://github.com/lucidrains/vit-pytorch

https://github.com/asyml/vision-transformer-pytorch