Rotation Invariant Convolutions for 3D Point Clouds Deep Learning

Zhiyuan Zhang, Binh-Son Hua, David W. Rosen, Sai-Kit Yeung

International Conference on 3D Vision (3DV) 2019

Introduction

This is the implementation of the rotation invariant convolution and neural networks for point clouds as shown in our paper. The key idea is to build rotation invariant features and use them to build a convolution to consume a point set. For details, please refer to our project.

@inproceedings{zhang-riconv-3dv19,
    title = {Rotation Invariant Convolutions for 3D Point Clouds Deep Learning},
    author = {Zhiyuan Zhang and Binh-Son Hua and David W. Rosen and Sai-Kit Yeung},
    booktitle = {International Conference on 3D Vision (3DV)},
    year = {2019}
}

Installation

The code is written in TensorFlow and based on PointNet, PointNet++, and PointCNN. Please follow the instruction in PointNet++ to compile the customized TF operators.

The code has been tested with Python 3.6, TensorFlow 1.13.2, CUDA 10.0 and cuDNN 7.3 on Ubuntu 14.04.

Usage

Classification

Please download the preprocessed ModelNet40 dataset here.

To train a network that takes XYZ coordinates as input to classify shapes in ModelNet40:

python3 train_val_cls.py

The evaluation is performed after every training epoch.

Part Segmentation

Download the preprocessed ShapeNetPart dataset here.

To train a network that takes XYZ coordinates as input to segments object parts:

python train_val_seg.py

The evaluation is performed after every training epoch.

License

This repository is released under MIT License (see LICENSE file for details).