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.
This repository is the Pytorch implementation of PRIN (Pointwise Rotation-Invariant Network).
- 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/)
- 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
- 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"
MIT
Our paper is available on https://arxiv.org/abs/1811.09361.
@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}
}