Uncracking the Bagel One Crumb at a Time

This repository contains the implementation of anomaly detection on MVTec 3D-AD dataset using point-voxel diffusion. We provide the report and poster under docs/ folder.

This project is a part of the course Computer Vision 2 at the University of Amsterdam (UvA) in the MSc Artificial Intelligence.

Figure of the anomaly detection pipeline

Requirements

Create the environment and install the requirements using the requirements.txt file

conda create -n pvd python=3.6
conda activate pvd
pip install -r requirements.txt

Dataset

To download the dataset locally, go to the MVTEC website and fill out your details. (Note: Only download the bagel dataset, since all data will take too much time and is not relevant for reproduction, since we only work on bagels).

To preprocess, run the preprocessing file in the 3D-ADS folder

python -u 3D-ADS/utils/preprocessing.py \
    --data_path /path/to/bagels_data \

This process can take up to ~4h or more depending on your system.

The preprocessing is done in-place, so the tiff files will remain tiff files. Here is an example of before and after of bagel/train/good/xyz/000.tiff.

Before	After

Running the experiments

Training the generator on normalized healthy samples for 2500 epochs:

python -u PVD/train_generation.py \
    --dataroot /path/to/data_folder \
    --normalize \

The testing can be run using the pvd_test.py file

python -u PVD/pvd_test.py \
    --dataroot path/to/data \
    --anomaly_time 20 \

Citation

The PVD repository

@inproceedings{Zhou_2021_ICCV,
    author    = {Zhou, Linqi and Du, Yilun and Wu, Jiajun},
    title     = {3D Shape Generation and Completion Through Point-Voxel Diffusion},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {5826-5835}
}

3D-ADS

@article{horwitz2022empirical,
  title={An Empirical Investigation of 3D Anomaly Detection and Segmentation},
  author={Horwitz, Eliahu and Hoshen, Yedid},
  journal={arXiv preprint arXiv:2203.05550},
  year={2022}
}