LoG

LoG utilizes a single RTX 4090 for training highly realistic urban-scale models and for their real-time rendering. Visit our project page for more demos.

https://github.com/chingswy/LoGvideos/assets/22812405/abb200c1-5b9c-48fd-a9a1-f235499153d2

Our code is built upon PyTorch and leverages gaussian-splatting techniques.

Quick Start

For a smooth setup, follow the installation guide.

Dataset Preparation

We employ Colmap to prepare the dataset. Refer to the preprocessing documentation for detailed instructions. A minimal example dataset is provided here.

Training

Training the model is as simple as one command:

python3 apps/train.py --cfg config/example/test/train.yml split train

We automatically configure heuristic parameters based on the dataset size.

We provide a path for interpolation visualization

python3 apps/train.py --cfg config/example/test/train.yml split demo_interpolate ckptname output/example/test/level_of_gaussian/model_init.pth

The visualization video will be stored at output/example/test/level_of_gaussian/demo_interpolate/rgb.mp4

Immersive Visualization :rocket:

We will update a real-time rendering tool designed for immersive visualization.

Acknowledgements

We acknowledge the following inspirational prior work:

• gaussian-splatting
• LandMark
• UrbanBIS dataset
• UrbanScene3D dataset
• UAVD4L dataset

The rendering GUI is powered by our EasyVolcap tool.

Contributions are warmly welcomed! If you've made significant progress on any of these fronts, please consider submitting a pull request.

Citation

If you find this code useful for your research, please cite us using the following BibTeX entry.

@inproceedings{shuai2024LoG,
  title={Real-Time View Synthesis for Large Scenes with Millions of Square Meters},
  author={Shuai, Qing and Guo, Haoyu and Xu, Zhen and Lin, Haotong and Peng, Sida and Bao, Hujun and Zhou, Xiaowei},
  year={2024}
}