README.md

MetroGS: Efficient and Stable Reconstruction of Geometrically Accurate High-Fidelity Large-Scale Scenes

Getting Started

• Installation
• Data Preparation
• Video Rendering

Run & Evaluate

The detailed setting of each step on GauU-Scene and MatrixCity Dataset can be found in ./scripts/run_metrogs_NAME.sh.

Train model and merge

# ${SCENE} is the name of training config file.
# ${NAME} is the name of model output file.
[Optional] export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export NCCL_SHM_DISABLE=1 # Shared memory (SHM) is currently not supported.
python main_bsz.py fit --config configs/metrogs/train/${SCENE}.yaml -n ${NAME}
python utils/merge_distributed_ckpts.py outputs/${NAME}

Evaluate Rendering Performance

python main.py test --config configs/metrogs/val/${SCENE}.yaml -n ${NAME}

Mesh extraction and evaluation

python utils/gs2d_mesh_extraction.py outputs/${NAME} \
        --post ${POST_FLAG} \
        --voxel_size ${VOXEL_SIZE} \
        --sdf_trunc ${SDF_TRUNC} \
        --depth_trunc ${DEPTH_TRUNC} \

python tools/eval_tnt/run.py \
        --scene your_gt_pcd \
        --dataset-dir data/geometry_gt/your_scene \
        --transform-path data/geometry_gt/your_scene/transform.txt \
        --ply-path "outputs/${NAME}/fuse.ply"

Checkpoints

Preprocessed pointmaps and checkpoints are available here.

TODO List

• [ ] Incorporate CLM-GS.
• [x] Release the checkpoints.

Citation

If you find this repository useful, please use the following BibTeX entry for citation.

@article{chen2025metrogs,
  title={MetroGS: Efficient and Stable Reconstruction of Geometrically Accurate High-Fidelity Large-Scale Scenes},
  author={Chen, Kehua and Mao, Tianlu and Ma, Xinzhu and Jiang, Hao and Li, Zehao and Liu, Zihan and Gao, Shuqin and Zhao, Honglong and Dai, Feng and Zhang, Yucheng and others},
  journal={arXiv preprint arXiv:2511.19172},
  year={2025}
}

Acknowledgements

This repo benefits from CityGSV2, CityGS-X, Grendel-GS, Gaussian Lightning, TriMip-RF. Thanks for their great work!