PCC-COT

Point Cloud Compression via Constrained Optimal Transport

(Official Pytorch Implementation, the code is modified from D-PCC)

Introduction

Point cloud compression (PCC) algorithms are typically designed to achieve the lowest possible distortion at a given low bit rate. However, the perceptual quality is often neglected. To tackle this, we innovatively regard PCC as a constrained optimal transport (COT) problem and propose a novel data-driven method to take the balance of distortion, perception, and the bit rate. Specifically, our method adopts a discriminator to measure the perceptual loss, and a generator to measure the optimal mapping from the original point cloud distribution to the reconstructed distribution.

Results

• Quantitative results

• Qualitative results

Installation

• Install the following packages

python==3.7
torch==1.7.1
torchvision==0.8.2
CUDA==11.0
numpy==1.20.3
open3d==0.9.0.0
einops==0.3.2
scikit-learn==1.0.1
compressai
ninja
pickle
argparse
tensorboard

Data Preparation

First download the ShapeNetCore v1 and SemanticKITTI datasets, and then divide them into non-overlapping blocks.

• ShapeNet

# install the `Manifold' program
cd ./dataset
git clone https://github.com/hjwdzh/Manifold
cd Manifold && mkdir build
cd build 
cmake .. -DCMAKE_BUILD_TYPE=Release
make 
cd ..

# divide into blocks
python prepare_shapenet.py --date_root path/to/shapenet

• SemanticKITTI

python prepare_semantickitti.py --data_root path/to/semantickitti

Train

# shapenet
python train.py --dataset shapenet
# semantickitti
python train.py --dataset semantickitti

Test

# shapenet
python test.py --dataset shapenet --model_path path/to/model
# semantickitti
python test.py --dataset semantickitti --model_path path/to/model

The decompressed patches and full point clouds will also be saved at ./output/experiment_id/pcd by default.

Acknowledgments

Our code is built upon the following repositories: D-PCC, DEPOCO, PAConv, Point Transformer and MCCNN, thanks for their great work.

Citation

If you find our project is useful, please consider citing:

@inproceedings{li2024point,
  title={Point Cloud Compression via Constrained Optimal Transport},
  author={Li, Zezeng and Wang, Weimin and Wang, Ziliang and Lei, Na},
  booktitle={2024 IEEE International Conference on Multimedia and Expo (ICME)},
  year={2024}
}