CVPR2022 - Anomaly Detection via Reverse Distillation from One-Class Embedding

1. Environment

   python=3.8
   torch1.10.0+cu113
   torchvision0.11.0+cu113
   scikit-learn
   opencv-python
   scikit-image
   pandas
   

2. Dataset

   MVTec_AD
   DATASET_HUAWEI
   

3. Download pretain weights

   autodl: https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth => /root/.cache/torch/hub/checkpoints/
   

4. Train and Test the Model

   python main.py

5. results:

   1. make first conv1 (which achieves downsample from 224 to 112) of pretrained feature extracter wide_Resnet50 from Conv2d(k7x7,s2,p3) to row_Conv【Conv2d(k1x7,s(1,2),p(0,3))】 & col_Conv【Conv2d(k7x1,s(2,1),p(3,0))】 => fiber auroc:0.67 (epoch=250) => bad