Unsupervised Point Cloud Completion and Segmentation by Generative Adversarial Autoencoding Network

This repository contains PyTorch implementation for Unsupervised Point Cloud Completion and Segmentation by Generative Adversarial Autoencoding Network (NeurIPS 2022).

Start

Requirements

CUDA                            10.2    ~   11.1
python                          3.7
torch                           1.8.0   ~   1.9.0
numpy
lmdb
msgpack-numpy
ninja                              
termcolor
tqdm
open3d                          0.9.0 
h5py

We successfully build the pointnet2 operation lib with CUDA 10.2 + torch 1.9.0 and CUDA 11.1 + torch 1.8.0, separately. It should work with PyTorch 1.9.0+.

Install

cd util/pointnet2_ops_lib
python setup.py install

Pretrained Models

Download (NJU BOX code:ugaan, Baidu Yun code:d5ye) and extract our pretrained models to the log folder. The log folder should be

log
├── scannet
│   ├── bookshelf
│   │   ├── model-240.pkl
│   │   └── model-480.pkl
│   ├── chair
|   |   └── ...
│   └── ...
├── scannet_scanobj
│   └── ...
└── scanobj
    └── ...

Datasets

Download (NJU Box code:ugaan, Baidu Yun code:9wle) and extract our dataset, s3dis, and scanobjectnn to the data folder. The data folder should be

data
├── modelnet
|   └── ...
├── s3dis_coseg
├── scanobjectnn
├── shapenet
├── us_gt
└── ws

Evaluation

Evaluate segmentation results on our dataset.

python test_scannet.py --cate chair

Evaluate segmentation results on S3DIS using the weights trained on our dataset.

python test_s3dis.py --cate chair

Evaluate segmentation results on ScanObjectNN using the weights trained on our dataset.

python test_scannet_scanobj.py --cate chair

Evaluate segmentation results on ScanObjectNN using the weights trained on ScanObjectNN.

python test_scanobj.py --cate chair

Evaluate completion results on our dataset.

python test_scannet_com.py --cate chair

Train

Train on our dataset.

python train_scannet.py --cate chair

Train on our dataset for ScanObjectNN.

python train_scannet_scanobj.py --cate chair

Train on ScanObjectNN.

python train_scanobj.py --cate chair

License

MIT License

Acknowledgements

pointnet2 operation lib

Scan2CAD

ScanNet

S3DIS

ShapeNet

ModelNet

ScanObjectNN

Citation

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

@inproceedings{ma2022ugaan,
  title={Unsupervised Point Cloud Completion and Segmentation by Generative Adversarial Autoencoding Network},
  author={Ma, Changfeng and Yang, Yang and Guo, Jie and Pan, Fei and Wang, Chongjun and Guo, Yanwen},
  booktitle={NeurIPS},
  year={2022}
}