Skip to content

maggie0106/Graph-CNN-in-3D-Point-Cloud-Classification

Repository files navigation

Graph-CNN-in-3D-Point-Cloud-Classification (PointGCN)

This is a TensorFlow implementation of using graph convolutional neural network to solve 3D point cloud classification problem. Details are decribed in the short paper A GRAPH-CNN FOR 3D POINT CLOUD CLASSIFICATION and master project report in the folder Documents.

If you find this code usefule please cite the following paper:

Yingxue Zhang and Michael Rabbat, "A GRAPH-CNN FOR 3D POINT CLOUD CLASSIFICATION", International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018

Link: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8462291

Bibtex:

@inproceedings{ZhangR_18_gcnn_point_cloud,

author = {Yingxue Zhang and Michael Rabbat},

title = {A Graph-CNN for 3D Point Cloud Classification},

booktitle = {International Conference on Acoustics, Speech and Signal Processing (ICASSP)},

address = {Calgary, Canada},

year = {2018}

}

Getting Started

Prerequisites

Python 2.7
tensorflow (>0.12)

Installing instructions

  1. Clone this repository.
git clone git@github.com:maggie0106/Graph-CNN-in-3D-Point-Cloud-Classification.git
  1. Install the dependencies.
pip install -r requirements.txt
  1. Download data
    We are using the data from 3D benchmark data set ModelNet http://modelnet.cs.princeton.edu/.
    The mesh polygon data format from ModelNet is preprocessed into Point Cloud format by Charles R. Qi et al. (PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation, CVPR 2017 https://arxiv.org/abs/1612.00593)
    Download the data and put it in the data folder from the following link:
https://shapenet.cs.stanford.edu/media/modelnet40_ply_hdf5_2048.zip 

Model

You can choose between two models using different pooling scheme including global pooling and multi-resolution pooling. And two training schemes have been provided to alleviate the unbalanced data, please change the batchWeight line in the model.py accordingly.

  • global pooling: no subsampling process, only aims at picking the global features.
  • multi-resolution pooling: doing subsampling after each convolutional layer to shrink the graph dimension by farthest subsampling a subset of centroid points and preform max-pooling on each cluter formed by the nearest neighbor around each point in the subset.

Run the demo

To run global pooling model

cd global_pooling_model
python main.py

To run multi-resolution pooling model

cd multi_res_pooling_model	
python main_multi_res.py	

Useful github repo for graph convolutional neural network and deep learning on point set related research

  1. ChebyNet https://github.com/mdeff/cnn_graph
    Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst, Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering, Neural Information Processing Systems (NIPS 2016)
  2. GCN https://github.com/tkipf/gcn
    Thomas N. Kipf, Max Welling, Semi-Supervised Classification with Graph Convolutional Networks (ICLR 2017)
  3. PointNet https://github.com/charlesq34/pointnet
    Charles R. Qi, Hao Su, Kaichun Mo, Leonidas J. Guibas, PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation (CVPR 2017)

Using scope

This implementation can be used to achieve 3D point cloud classification and can be easily applied to point cloud part segmentation by simply removing the global features aggregation process to achieve pointwise classification. This model also has the potential to extend into any problem relate to the interaction between graph structure and graph signal or purely graph classification problem.

License

This project is licensed under the MIT License - see the [LICENSE.md] file for details

Acknowledgments