Signed Distance Function(SDF) means the method we estimate the point cloud to detect anomaly. We utilize the information of the RGB and the corresponding 3D point cloud to detect anomaly and complement each other to get the final score map.

Linux (Ubuntu 16.04)
Python 3.6+
PyTorch 1.7 or higher
CUDA 10.2 or higher

git clone https://github.com/jayliu0313/Shape-Guided.git
cd Shape-Guided
conda create --name myenv python=3.6
conda activate myenv
pip install -r requirement.txt

Here to download dataset

It will take few minutes to remove the backgoround of the point cloud.

python tools/preprocessing.py DATASET_PATH

Divided the point cloud into multiple local patches for each instance.

python cut_patches.py --datasets_path DATASET_PATH --save_grid_path data/

Make sure the order of execution of preprocessing.py is before cut_patches.py.

There is the best checkpoint of the 3D expert model in checkpoint/best_ckpt/ckpt_000601.pth, and you can skip this step. Alternatively, you can train the 3D expert model on your own. So, you need to execute the following commands to get the required training patches which are contained the noise points.
Recommend setting the save_grid_path in the same directory as above.

python cut_patches.py --datasets_path DATASET_PATH --save_grid_path data/ --pretrain

then,

python train_3Dmodel.py --grid_path data/ --ckpt_path "./checkpoint"

The result will be stored in the output directory.

python main.py --datasets_path DATASET_PATH --grid_path data/ --ckpt_path "checkpoint/best_ckpt/ckpt_000601.pth"

If our paper is useful for your research, please cite our paper. Thank you!

@InProceedings{pmlr-v202-chu23b,
  title = {Shape-Guided Dual-Memory Learning for 3D Anomaly Detection},
  author = {Chu, Yu-Min and Liu, Chieh and Hsieh, Ting-I and Chen, Hwann-Tzong and Liu, Tyng-Luh},
  booktitle = {Proceedings of the 40th International Conference on Machine Learning},
  pages = {6185--6194},
  year = {2023},
}

Our memory architecture is refer to https://github.com/eliahuhorwitz/3D-ADS
3D expert model is modified from https://github.com/mabaorui/PredictableContextPrior