https://github.com/pxiangwu/RCNet

Code for "Point cloud processing via recurrent set encoding"
https://github.com/pxiangwu/RCNet

Last synced: 14 days ago
JSON representation

Code for "Point cloud processing via recurrent set encoding"

• Host: GitHub
• URL: https://github.com/pxiangwu/RCNet
• Owner: pxiangwu
• Created: 2020-03-09T19:21:55.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-03-28T05:03:02.000Z (over 4 years ago)
• Last Synced: 2024-08-01T03:45:53.521Z (3 months ago)
• Language: Python
• Size: 30.3 KB
• Stars: 6
• Watchers: 2
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: readme.md

Awesome Lists containing this project

README

        
## Code for ["Point cloud processing via recurrent set encoding"](https://arxiv.org/pdf/1911.10729.pdf).

## Requirements:

- PyTorch 0.3/0.4

- Python 3.6+

- CUDA 8.0 (Not sure if CUDA > 8.0 will work. This depends on PyTorch.)

## Usage and file structures:

- For classification, run `train_rcnet_cls.py`. For shape segmentation, run `run_seg.py`

- The dataset can be downloaded from https://github.com/charlesq34/pointnet:

```

# classification data

wget https://shapenet.cs.stanford.edu/media/modelnet40_normal_resampled.zip

# segmentation data

wget https://shapenet.cs.stanford.edu/media/shapenetcore_partanno_segmentation_benchmark_v0_normal.zip

```

- The dataloader for classification task is in file `utils/datasets.py`. The dataloader for segmentation task is in file `utils/part_dataset.py`.

- `utils/pointnet.py1` provides some modules, which are about spatial transformer (STN).  

- `utils/orderpoints.py` provides functions for partitioning the ambient space into structured beam. Currently the code is not the most efficient. For more efficient version, please refer to [here](https://github.com/pxiangwu/MotionNet/blob/master/data/data_utils.py#L105).

- `utils/provider` provides functions for some basic data augmentation, such at random jitter, random scale, etc.

- Directory `utils/gen_point_cloud` includes some codes (you may not need) for converting mesh into point clouds. It relies on pyntcloud library.

## Reference:

```

@inproceedings{wu2019point,

  title={Point cloud processing via recurrent set encoding},

  author={Wu, Pengxiang and Chen, Chao and Yi, Jingru and Metaxas, Dimitris},

  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},

  volume={33},

  pages={5441--5449},

  year={2019}

}

```