HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting

Qiankun Gao^{1, 2}, Jiarui Meng¹, Chengxiang Wen¹, Jie Chen^{1, 2 ✉️}, Jian Zhang^{1, 3 ✉️}

¹School of Electronic and Computer Engineering, Peking University ²Peng Cheng Laboratory ³Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology,
Peking University Shenzhen Graduate School

[NeurIPS] [arXiv] [OpenReview]

🗞️ News

• [2024/12/19] Code released.
• [2024/10/29] Camera ready submitted.
• [2024/09/26] Accepted to NeurIPS 2024 as poster presentation!

⚙️ Experiment

The code is built on LibGS; please familiarize yourself with it before running the experiments.

1. Install dependencies

   pip install .

2. Run pipeline

   python main.py --config=config/dynerf.yaml --data.root=<PATH TO SCENE ROOT>
   

   The config directory contains pre-defined configurations for reproducing the results reported in the paper.

📖 Citation

@inproceedings{hicom2024,
  title = {HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting},
  author={Gao, Qiankun  and Meng, Jiarui and Wen, Chengxiang  and Chen, Jie and Zhang, Jian},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2024}
}

🙌 Acknowledgement

We sincerely thank the authors of 3DGStream for their inspiring work and valuable assistance. We also appreciate the contributions and accessible code provided by related research efforts, including 3DGS, 4DGaussians, and Dynamic3DGS, which have greatly supported our research.