3DAC: Learning Attribute Compression for Point Clouds (CVPR 2022)
This is the official implementation of 3DAC, a learning-based data compression framework for point cloud attributes. For technical details, please refer to:
3DAC: Learning Attribute Compression for Point Clouds
Guangchi Fang, Qingyong Hu, Hanyun Wang, Yiling Xu, Yulan Guo.
[Paper]
(1) Setup
This code has been tested with Python 3.7, torch 1.7.0, CUDA 10.2.
- Clone the repository
git clone git@github.com:fatPeter/ThreeDAC.git && cd ThreeDAC
- Setup python environment
conda create -n ThreeDAC python=3.7
conda activate ThreeDAC
pip install -r helper_requirements.txt
- Install torchsparse
sudo apt-get install libsparsehash-dev
pip install git+https://github.com/mit-han-lab/torchsparse.git@v1.1.0
(2) ScanNet
- Download the ScanNet dataset from the official website and move it to
./dataset. - Voxelize the point cloud geometry with a 9-depth octree and deduplicate redundant points. We merge this process to the dataloader to simplify our code.
- Train/eval and compress/decompress scripts are in
./ScanNet:
cd ScanNet
- Train on ScanNet (Qstep denotes the quantization step):
python train.py --Qstep=10
- Quick eval:
python eval.py --Qstep=10
- Compress scans of ScanNet:
python compress.py --Qstep=10
- Decompress:
python decompress.py --Qstep=10
(3) SemanticKITTI
- SemanticKITTI dataset can be found here. Download the files and extract everything into the same folder. Uncompress the folder and move it to
./dataset. - Voxelize the point cloud geometry with a 12-depth octree and deduplicate redundant points. We merge this process to the dataloader to simplify our code.
- Train/eval and compress/decompress scripts are in
./SemanticKITTI:
cd SemanticKITTI
- Train on SemanticKITTI (step: use part of data for quick testing):
python train.py --Qstep=10 --step=20
- Compress scans of ScanNet:
python compress.py --Qstep=10 --step=200
- Decompress:
python decompress.py --Qstep=10 --step=200
Citation
If you find our work useful in your research, please consider citing:
@article{fang20223dac,
title={3DAC: Learning Attribute Compression for Point Clouds},
author={Fang, Guangchi and Hu, Qingyong and Wang, Hanyun and Xu, Yiling and Guo, Yulan},
journal={arXiv preprint arXiv:2203.09931},
year={2022}
}