PointWOLF: Point Cloud Augmentation with Weighted Local Transformations

This repository is the implementation of PointWOLF.

Sihyeon Kim^1*, Sanghyeok Lee^1*, Dasol Hwang¹, Jaewon Lee¹, Seong Jae Hwang², Hyunwoo J. Kim^1†, Point Cloud Augmentation with Weighted Local Transformations (ICCV 2021).
¹Korea University ²University of Pittsburgh

Installation

Dependencies

• CUDA 10.2
• Python 3.7.1
• torch 1.7.0
• packages : sklearn, numpy, h5py, glob

Download

Clone repository

$ git clone https://github.com/mlvlab/PointWOLF.git

Download ModelNet40

Notes : When you run the main.py, ModelNet40 is automatically downloaded at .../PointWOLF/data/.
If you want to download dataset on your ${PATH}, see below.

$ cd ${PATH}
$ wget https://shapenet.cs.stanford.edu/media/modelnet40_ply_hdf5_2048.zip --no-check-certificate
$ unzip modelnet40_ply_hdf5_2048.zip
$ rm modelnet40_ply_hdf5_2048.zip

Runnig the code

train

• Run the training without PointWOLF & AugTune:

$ python main.py --exp_name=origin --model=dgcnn --num_points=1024 --k=20 --use_sgd=True

• Run the training with PointWOLF:

$ python main.py --exp_name=PointWOLF --model=dgcnn --num_points=1024 --k=20 --use_sgd=True --PointWOLF

• Run the training with PointWOLF & AugTune:

$ python main.py --exp_name=PointWOLF_AugTune --model=dgcnn --num_points=1024 --k=20 --use_sgd=True --PointWOLF --AugTune

eval

• Run the evaluation with trained model located at ${PATH}:

$ python main.py --exp_name=eval --model=dgcnn --num_points=1024 --k=20 --use_sgd=True --eval=True --model_path=${PATH}

Citation

@InProceedings{Kim_2021_ICCV,
    author    = {Kim, Sihyeon and Lee, Sanghyeok and Hwang, Dasol and Lee, Jaewon and Hwang, Seong Jae and Kim, Hyunwoo J.},
    title     = {Point Cloud Augmentation With Weighted Local Transformations},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {548-557}
}

License

MIT License

Acknowledgement

The structure of this codebase is borrowed from DGCNN.