Uni10K from A LoD of Gaussians: Unified Training and Rendering for Ultra-Large-Scale Reconstruction with External Memory

Felix Windisch¹ Thomas Köhler¹ Lukas Radl¹ Mattia D'Urso¹ Michael Steiner¹ Dieter Schmalstieg^1,2 Markus Steinberger^1,3

¹Graz University of Technology, ²University of Stuttgart, ³Huawei Technologies

SIGGRAPH 2026

 

COLMAP sparse reconstruction with cameras from Uni10K.

General Information

Uni10K is a large-scale outdoor reconstruction dataset of a university campus covering approximately 100,000 m². The scene is captured from both ground-level and aerial viewpoints and contains more than 10,000 4K-resolution images.

A standard COLMAP reconstruction at this scale would typically require weeks of computation because of the complexity of image matching and mapping. To make reconstruction feasible, we use spatial and temporal priors together with a coarse-to-fine reconstruction scheme.

First, GPS information is used to reduce the image-matching complexity from quadratic to near-linear by limiting image comparisons to a predefined spatial radius. Second, because the frames are sampled from video sequences, we initially reconstruct a baseline model from frames sampled uniformly at 1 frame per second. Camera coverage is then densified by incrementally registering, triangulating, and refining additional images using local bundle adjustment. The process concludes with several rounds of global bundle adjustment.

The final reconstruction contains more than 10,000 images, 6.2 million sparse points, and an overall mean reprojection error of approximately 0.63 pixels. Training is performed on the full-resolution images. Every 8th frame, sorted alphabetically, is held out as part of the test set.

Dataset Format

The dataset follows the structure below. Images are undistorted and ready for 3D Gaussian Splatting.

undistorted/
├── sparse/                  # Structure-from-Motion outputs in COLMAP format
│   ├── cameras              # Intrinsic camera parameters
│   ├── images               # Extrinsic parameters / camera poses
│   └── points3D             # Sparse 3D point cloud
|
└──images/                   # Extracted video frames
    ├── aerial/
    │   ├── frame_000000.jpg
    │   └── ...
    └── ...

Citation

If you find this dataset useful in your research, please consider citing:

@article{windisch2025lod,
  title={A LoD of Gaussians: Unified Training and Rendering for Ultra-Large-Scale Reconstruction with External Memory},
  author={Windisch, Felix and K{\"o}hler, Thomas and Radl, Lukas and D'Urso, Mattia and Steiner, Michael and Schmalstieg, Dieter and Steinberger, Markus},
  journal={arXiv preprint arXiv:2507.01110},
  year={2025}
}