PRIN

Pointwise Rotation-Invariant Network in PyTorch

News

An improved version of PRIN (SPRIN) is released here and described in PRIN/SPRIN: On Extracting Point-wise Rotation Invariant Features, which achieves much better results.

Overview

This repository is the Pytorch implementation of PRIN (Pointwise Rotation-Invariant Network).

Dependencies

• Install s2cnn (https://github.com/jonas-koehler/s2cnn) and its dependencies (pytorch, cupy, lie_learn, pynvrtc).
• Install pybind11 and compile the script under src (https://pybind11.readthedocs.io/)

Dataset and pretrained weights

• Download ShapeNet 17 Part Segmentation Dataset in h5py format from https://drive.google.com/drive/folders/1wC-DpeRtxuuEvffubWdhwoGXGeW052Vy?usp=sharing
• Download pretrained weights (trained on unrotated shapes) from https://drive.google.com/open?id=1QnFqQdWmx0cYtYeN9tJNlf-E5ZLawRBv

Usage

• For training, run "python train.py --log_dir log --model_path ./model.py --num_workers 4"
• For testing, run "python test.py --weight_path ./state.pkl --model_path ./model.py --num_workers 4"

License

MIT

References

Our paper is available on https://arxiv.org/abs/1811.09361.

Citation

@inproceedings{you2020pointwise,
  title={Pointwise rotation-invariant network with adaptive sampling and 3d spherical voxel convolution},
  author={You, Yang and Lou, Yujing and Liu, Qi and Tai, Yu-Wing and Ma, Lizhuang and Lu, Cewu and Wang, Weiming},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={34},
  number={07},
  pages={12717--12724},
  year={2020}
}