[AAAI 2023] PaRot: Patch-Wise Rotation-Invariant Network via Feature Disentanglement and Pose Restoration

Official implementation of "PaRot: Patch-Wise Rotation-Invariant Network via Feature Disentanglement and Pose Restoration", AAAI 2023. [Paper] [Supp.] [Video]

We've optimized the code and released more experimental data, including class mIoU and instance mIoU, to facilitate comparisons with other methods and enable further analysis.

Requirements

Python 3.7
Pytorch 1.10
CUDA 10.2
Packages: pytorch3d, tqdm, sklearn, visualdl, opencv-python

Data

The ModelNet40 and ShapeNetPart dataset will be automatically downloaded. For ScanObjectNN, you need to fill out an agreement to get the download link.

Performance

Accuracy on ModelNet40 under rotation: 91.0% (z/SO(3)), 90.8% (SO(3)/SO(3)).
Accuracy on ScanObjectNN OBJ_BG classification under rotation:

z/z	z/SO(3)	SO(3)/SO(3)
82.4%	82.1%	82.6%

Averaged mIoU on ShapeNetPart segmentation under rotation:

	z/z	z/SO(3)	SO(3)/SO(3)
Class mIOU	79.1%	79.2%	79.5%
Insta. mIOU	82.8%	82.9%	82.8%

Citation

If you find this repo useful in your work or research, please cite:

@article{Zhang_Yu_Zhang_Cai_2023,
title={PaRot: Patch-Wise Rotation-Invariant Network via Feature Disentanglement and Pose Restoration},
author={Zhang, Dingxin and Yu, Jianhui and Zhang, Chaoyi and Cai, Weidong},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={37}, 
number={3},
year={2023},
month={Jun.},
pages={3418-3426}
}

Training Command

For ModelNet40 model train (1024 points)

python main_cls.py --exp_name=modelnet40_cls --train_rot=z --test_rot=so3

For ShapeNetPart segmentation model train (2048 points)

python main_seg.py --exp_name=shapenet_seg --train_rot=z --test_rot=so3

Acknowledgement

Our code borrows a lot from:

dingxin-zhang/PaRot