AnomalyDINO: Boosting Patch-based Few-shot Anomaly Detection with DINOv2

Simon Damm, Mike Laszkiewicz, Johannes Lederer, Asja Fischer

This is the official code to reproduce the experiments in the paper AnomalyDINO: Boosting Patch-based Few-shot Anomaly Detection with DINOv2.

Prerequisits

Create a virtual environment (e.g., python -m venv .venvAnomalyDINO), activate it (e.g., source .venvAnomalyDINO/bin/activate) and install the required dependencies for AnomalyDINO:
```
pip install -r requirements.txt
```
Info: If you want to use faiss with GPU-acceleration we recommend setting up a conda environment with the required packages instead (only conda installation is supported, see, e.g., here). To perform similarity search on CPU set the additional flag --faiss_on_cpu.

Download and prepare the datasets MVTec-AD and VisA from their official sources. For VisA, follow the instruction in the official repo to organize the data in the official 1-class splits. The default data roots are data/mvtec_anomaly_detection for MVTec-AD, and data/VisA_pytorch/1cls/ for VisA. Please adapt the function calls below if necessary. Alternatively, prepare your own dataset accordingly:

your_data_root
├── object1
│   ├── ground_truth        # anomaly annotations per anomaly type
│   │   ├── anomaly_type1
│   │   ├── ...
│   ├── test                # test images per anomaly type & 'good'
│   │   ├── anomaly_type1    
│   │   ├── ...
│   │   └── good
│   └── train               # train/reference images (without anomalies)
│       └── good
├── object2
│   ├── ...

Usage

Short Demo

Get started with the minimal demo to perform few-shot anomaly detection (demo_AD_DINO.ipynb).

Few-shot anomaly detection

For the full evaluation, run the script run_anomalydino.py on the selected dataset for a given number of shots and repetitions (seeds). The preprocessing to your dataset can be specified in src/utils.py in get_dataset_info, default is "agnostic" (apply masking whenever PCA-based masking works well & augment reference samples by rotations, see the paper).

The results for the default setting, i.e., all considered shots, three repetitions, and agnostic preprocessing, can be reproduced by calling:

python run_anomalydino.py --dataset MVTec --shots 1 2 4 8 16 --num_seeds 3 --preprocess agnostic --data_root data/mvtec_anomaly_detection

python run_anomalydino.py --dataset VisA --shots 1 2 4 8 16 --num_seeds 3 --preprocess agnostic --data_root data/VisA_pytorch/1cls/

For a faster inspection use, e.g.,

python run_anomalydino.py --dataset MVTec --shots 1 --num_seeds 1 --preprocess informed --data_root data/mvtec_anomaly_detection

The script automatically creates some example plots, plots some anomaly maps for each object, and automatically evaluates each run (activate evaluation of segementation with --eval_segm if applicable).

Evaluation results are saved in the respective results directory as metrics_seed={seed}.json for each seed.

Batched-Zero-Shot Anomay Detection

To reproduce the results in the batched zero-shot scenario, run run_anomalydino_batched.py with appropriate arguments:

python run_anomalydino_batched.py --dataset MVTec --data_root data/mvtec_anomaly_detection

python run_anomalydino_batched.py --dataset VisA --data_root data/VisA_pytorch/1cls/

This work uses the following ressources and datasets:

DINOv2, code and model available under Apache 2.0 license.
The MVTec-AD dataset, available under the CC BY-NC-SA 4.0 license.
The VisA dataset, available under the CC BY 4.0 license.

If you find this repository useful in your research/project, please consider citing the paper: