USSPA: Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion

This repository contains PyTorch implementation for Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion (CVPR2023).

YouTube: https://youtu.be/1iWvKcR9DzA

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                           
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 code/util/pointnet2_ops_lib
python setup.py install

Pretrained Models

Download (NJU BOX code:usspa, Baidu Yun code:boqx) and extract our pretrained models as the weights folder in code/network. The weights folder should be

weights
├── usspa
│   ├── all
│   │   └── model-120.pkl
│   ├── chair
|   |   └── ...
│   └── ...
├── scannet_scanobj
│   └── ...
└── scanobj
    └── ...

Datasets

Download (NJU Box code:usspa, Baidu Yun code:sbo2) and extract our dataset and ShapeNet dataset to the data folder. And download PCN dataset following PoinTr. The data folder should be

data
├── PCN
|   └── ...
├── RealComData
└── RealComShapeNetData

Evaluation

cd code/network

Evaluate completion results of USSPA on our dataset for single-category and multi-category.

python test.py --class_name [all, chair, ...]

Evaluate completion results of USSPA(classifier) on our dataset for multi-category.

python test_classifier.py

Evaluate completion results of USSPA on PCN dataset.

python test_pcn.py --class_name [chair, table, ...]

Train

cd code/network

Train USSPA on our dataset for single-category and multi-category.

python train.py --class_name [all, chair, ...]

Train USSPA(classifier) on our dataset for multi-category.

python train_classifier.py

Train USSPA on PCN dataset.

python train_pcn.py --class_name [chair, table, ...]

License

MIT License

Acknowledgements

pointnet2 operation lib

Scan2CAD

ScanNet

ShapeNet

Citation

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

@inproceedings{ma2023usspa,
  title={Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion},
  author={Ma, Changfeng and Chen, Yinuo and Guo, Pengxiao and Guo, Jie and Wang, Chongjun and Guo, Yanwen},
  booktitle={CVPR},
  year={2023}
}