By Yiyang Chen, Lunhao Duan, Shanshan Zhao, Changxing Ding and Dacheng Tao
This is the official implementation of "Local-consistent Transformation Learning for Rotation-invariant Point Cloud Analysis" [arXiv]
- Python 3.7.0
- Pytorch 1.9.0
- CUDA 11.1
- Packages: numpy, pytorch3d, sklearn, h5py, tqdm
- 4 NVIDIA TITAN V GPUs
The ModelNet40 dataset will be automatically downloaded.
Accuracy on ModelNet40 dataset under rotation:
- Ours: 91.6% (z/z, z/SO(3)), 91.5% (SO(3)/SO(3))
- Ours*: 91.5% (z/z, z/SO(3)), 91.7% (SO(3)/SO(3))
Ours represents our original network. We further develop a lightweight
version by reducing the computational burden of our original network.
Ours* represents the lightweight version.
python main_cls.py --exp_name=modelnet40_cls --rot=ROTATION
python main_cls_l.py --exp_name=modelnet40_cls --rot=ROTATION
Here ROTATION should be chosen from aligned, z and so3. main_cls.py is the script for training
the original network, while main_cls_l.py is for training the lightweight version.
python main_cls.py --exp_name=modelnet40_cls --rot=ROTATION --eval=True --checkpoint=MODEL
Here MODEL should be chosen from model, model_vn, model_fuse, model_1 and model_2.
model_1 or model_2 achieves the best performance.
python main_cls_l.py --exp_name=modelnet40_cls --rot=ROTATION --eval=True --checkpoint=MODEL
Here MODEL should be chosen from model, model_vn, model_1.
model_1 achieves the best peformance.
You can also test our pretrained model directly:
python main_cls.py --exp_name=modelnet40_cls --rot=ROTATION --eval=True --model_path PATH
Here PATH can be set as pretrained/model_1_z.t7 or pretrained/model_1_so3.t7.
python main_cls_l.py --exp_name=modelnet40_cls --rot=ROTATION --eval=True --model_path PATH
Here PATH can be set as pretrained_l/model_1_z.t7 or pretrained_l/model_1_so3.t7.
If you find this repo useful, please cite:
@article{chen2024local,
title={Local-consistent Transformation Learning for Rotation-invariant Point Cloud Analysis},
author={Chen, Yiyang and Duan, Lunhao and Zhao, Shanshan and Ding, Changxing and Tao, Dacheng},
journal={arXiv preprint arXiv:2403.11113},
year={2024}
}
Our code borrows from: