L2G Auto-encoder: Understanding Point Clouds by Local-to-Global Reconstruction with Hierarchical Self-Attention
Created by Xinhai Liu, Zhizhong Han, Xin Wen, Yu-Shen Liu, Matthias Zwicker.
If you find our work useful in your research, please consider citing:
@inproceedings{liu2019l2gautoencoder,
title={ L2G Auto-encoder: Understanding Point Clouds by Local-to-Global Reconstruction with Hierarchical Self-Attention},
author={Liu, Xinhai and Han, Zhizhong and Wen, Xin and Liu, Yu-Shen and Zwicker, Matthias},
booktitle={Proceedings of the 27th ACM International Conference on Multimedia},
year={2019}
}
In L2G-AE, we focus on learning the local and global structures of point clouds in an auto-encoder architecture. Specifically, we propose hierarchical self-attentions to learn the correlation among point features in different semantic levels by highlight the importance of each element. In addition, we also introduce a RNN docoding layer to decode the features of different scale areas in the local region reconstruction.
In this repository we release code our L2G-AE classification as well as a few utility scripts for training, testing and data processing.
Install TensorFlow. The code is tested under TF1.4 GPU version and Python 2.7 on Ubuntu 16.04. There are also some dependencies for a few Python libraries for data processing like cv2, h5py etc. It's highly recommended that you have access to GPUs.
Before running the code, you need to compile customized TF operators as described in PointNet++.
To train a Point2Sequence model to classify ModelNet40 shapes (using point clouds with XYZ coordinates):
python train_hierarchical_attention.py
To see all optional arguments for training:
python train_hierarchical_attention.py -h
Follow the dataset in PointNet++, you can refer to here on how to prepare your own HDF5 files for either classification or segmentation. Or you can refer to modelnet_dataset.py on how to read raw data files and prepare mini-batches from them.
Our code is released under MIT License (see LICENSE file for details).
- PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space by Qi et al. (NIPS 2017)
- Attentional ShapeContextNet for Point Cloud Recognition by Xie et al. (CVPR 2018)
- PU-Net: Point Cloud Upsampling Network by Yu et al. (CVPR 2018)
- Point2Sequence: Learning the Shape Representation of 3D Point Clouds with an Attention-Based Sequence to Sequence Network by Liu et al. (AAAI 2019)