3D-Rotation-Equivariant Quaternion Neural Networks (PyTorch)

Introduction

This repository is an official implementation of 3D-Rotation-Equivariant Quaternion Neural Networks (arXiv,SpringerLink) which has been published at ECCV 2020.

Note that, we have found a bug and solved it. Therefore, the experimental results obtained based on this repository are slightly different from that in the paper. However, this did not essentially change our conclusions. New experimental results are as follows.

Method		ModelNet40			3D MNIST
	Baseline w/o rotation	Baseline w/ rotation	REQNN(ours)	Baseline w/o rotation	Baseline w/ rotation	REQNN(ours)
PointNet++	25.87	29.25	62.03	44.19	51.48	72.01
DGCNN	32.08	33.78	84.57	45.90	50.00	84.57
PointConv	25.01	26.46	81.93	45.51	48.08	85.71

Besides, we have modified a few writing errors. In the second and third paragraph of the Experiment section, and the caption of Table 4, we have modified all “z-axis rotations” to “y-axis rotations”.

Installation

To run the program successfully, you need to include all packages in requirements.txt on your server.

pip install -r requirements.txt

Usage

To train a model to classify point clouds.

Run the training script:

python main.py 

Log files and network parameters will be saved to checkpoint folder in default.

You can specify models, datasets and other train configurations in training script. For example:

python main.py --exp_name=dgcnn_reqnn_1024_train --model=dgcnn_reqnn --dataset=modelnet --use_sgd=True --lr=0.001

See HELP for the training script:

python main.py -h

Run the evaluation script with trained models:

python main.py --exp_name=dgcnn_reqnn_1024_eval --eval=True --eval_model_path=your_trained_model_path

Citation

If you use this project in your research, please cite it.

@inproceedings{shen20193d,
	title={3d-rotation-equivariant quaternion neural networks},
	author={Shen, Wen and Zhang, Binbin and Huang, Shikun and Wei, Zhihua and Zhang, Quanshi},
	booktitle={ECCV},
	year={2020}
}