Awesome

A-CNN: Annularly Convolutional Neural Networks on Point Clouds

Created by <a href="https://github.com/artemkomarichev" target="_blank">Artem Komarichev</a>, <a href="http://www.cs.wayne.edu/zzhong/" target="_blank">Zichun Zhong</a>, <a href="http://www.cs.wayne.edu/~jinghua/" target="_blank">Jing Hua</a> from Department of Computer Science, Wayne State University.

teaser image

Introduction

Our paper (<a href="https://arxiv.org/abs/1904.08017" target="_blank">arXiV</a>) proposes a new approach to define and compute convolution directly on 3D point clouds by the proposed annular convolution.

To appear, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

A-CNN usage

We provide the code of A-CNN model that was tested with Tensorflow 1.3.0, CUDA 8.0, and python 3.6 on Ubuntu 16.04. We run all our experiments on a single NVIDIA Titan Xp GPU with 12GB GDDR5X.

Classification Task

<a href="https://1drv.ms/u/s!ApbTjxa06z9CgQfKl99yUDHL_wHs">Download</a> ModelNet-40 dataset first. Point clouds are sampled from meshes with 10K points (XYZ + normals) per shape and provided by PointNet++.

To train a classification A-CNN model on ModelNet-40 dataset type the following command:
```
  python train.py
```
To evaluate a trained model run the following script:
```
  python evaluate.py
```
Part Segmentation Task

<a href="https://1drv.ms/u/s!ApbTjxa06z9CgQnl-Qm6KI3Ywbe1">Download</a> ShapeNet-part dataset first. Each point cloud represented by 2K points (XYZ + normals) and provided by PointNet++.

To train a part segmentation A-CNN model on ShapeNet-part dataset type the following commands:
```
  cd part_segm
  python train.py
```
To evaluate a trained segmentation model run the following script:
```
  ./evaluate_job.sh
```
Semantic Segmentation Task

Download <a href="http://buildingparser.stanford.edu/dataset.html">S3DIS</a> and <a href="https://shapenet.cs.stanford.edu/media/scannet_data_pointnet2.zip">ScanNet</a> datasets provided by PointNet/PointNet++. S3DIS contains XYZ + RGB information. ScanNet only has geometry information (XYZ only), no color.

To estimate normals we used PCL library. The script to estimate normals for ScanNet data could be found here:
```
  cd scannet/normal_estimation
  ./run.sh
```

Citation

If you find our work useful in your research, please cite our work:

@InProceedings{komarichev2019acnn,
    title={A-CNN: Annularly Convolutional Neural Networks on Point Clouds},
    author={Komarichev, Artem and Zhong, Zichun and Hua, Jing},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    year={2019}
}