G3AR: Graph-Guided Neural Visual Geometry for Scalable Multi-Sequence Aerial Registration

SIGGRAPH Asia 2026 Technical Communications

Jeng Wen Joshua Lean¹, Ting-Yu Yen¹, Wei-Fang Sun², Simon See², Hung-Kuo Chu¹, Shih-Hsuan Hung¹

¹National Tsing Hua University    ²NVIDIA AI Technology Center (NVAITC)

1. Setup

Linux, an NVIDIA GPU, CUDA 13-compatible drivers, Python 3.12, and uv are required.

uv sync
uv run g3ar setup

The setup command downloads the pinned retrieval and DA3 checkpoints into the ignored checkpoints/ directory. DA3 weights are provided separately under CC BY-NC 4.0 and are intended for non-commercial use.

2. Run

Input is a COLMAP-style directory containing images/. Nested folders below images/ are treated as separate temporal sequences.

uv run g3ar /path/to/dataset
uv run g3ar /path/to/dataset --gpus 0,1
uv run g3ar /path/to/dataset --config configs/default.yaml

Results are written to outputs/<dataset>-<timestamp>/. Every disconnected result is exported independently under components/NNN/ with a point cloud, camera JSON, and COLMAP text model. Failed runs remain as .partial directories. Successful runs remove inference caches unless --keep-intermediates is set.

The checked-in default config documents all public options. Unknown or invalid config values fail before reconstruction.

Citation

@article{lean2026g3ar,
  title={G3AR: Graph-Guided Neural Visual Geometry for Scalable Multi-Sequence Aerial Registration},
  author={Lean, Jeng Wen Joshua and Yen, Ting-Yu and Sun, Wei-Fang and See, Simon and Chu, Hung-Kuo and Hung, Shih-Hsuan},
  journal={arXiv preprint arXiv:2609.16603},
  year={2026}
}

License

Code is released under Apache-2.0. See NOTICE for bundled third-party code and model-license information.

Acknowledgements

We thank VGGT-Long and MERG3R for their influential ideas and open-source contributions to scalable neural visual geometry.

This implementation is built with the following open-source projects:

• Depth Anything 3
• LightGlue and ALIKED
• UltraVPR

We sincerely thank all their authors and maintainers.

We are also deeply grateful to Taiwan's National Science and Technology Council (NSTC) and the NVIDIA AI Technology Center (NVAITC) for their extraordinary support.