Anomaly Multi-classification in Industrial Scenarios: Transferring Few-shot Learning to a New Task

This repository is based on PatchCore's official implementation. This repository contains the implementation proposed in our paper.

This repository also doesn't provide a evaluation process, all results can be trained and tested very soon so we only save the model being pretrained.

---

Requirements

Our results were computed using Python 3.8, with packages and respective version noted in requirements.txt. In general, the majority of experiments should not exceed 24GB of GPU memory;

Quick Guide

We recommand you to use Visual Studio Code, as we provide a "launch.json" file to launch our project.

All settings can be found in "launch.json" and run_patchcore.py

Setting up DTD

Download the DTD from here:https://www.robots.ox.ac.uk/~vgg/data/dtd/ Make sure that it follows the following data tree:

generate_anomaly_pkg
|-- dtd
|-----|----- images
|-----|----- imdb
|-----|----- labels
|-- generate_anomaly.py
|-- data_loader_for_draem.py
|-- ...

Setting up MVTec AD and generate synthesized anomaly

To set up the main MVTec AD benchmark, download it from here: https://www.mvtec.com/company/research/datasets/mvtec-ad. Make sure that it follows the following data tree:

mvtec_anomaly_detection
|-- bottle
|-----|----- ground_truth
|-----|----- test
|-----|--------|------ good
|-----|--------|------ broken_large
|-----|--------|------ ...
|-----|----- train
|-----|--------|------ good
|-- cable
|-- generate_foreground.py
|-- generate_foreground copy.py
|-- ...

containing in total 15 subdatasets: bottle, cable, capsule, carpet, grid, hazelnut, leather, metal_nut, pill, screw, tile, toothbrush, transistor, wood, zipper.

Then you can run "generate_foreground.py" and "generate_foreground copy.py" respectively with "run Currrent file" in "launch.json", the generated image will be placed in folder "generate_anomaly_pkg"

Training

run ./bin/run_patchcore.py with "run_patchcore"

License

This project is licensed under the Apache2.0 License.