Splataverse Streets
Textured photogrammetry meshes of real places converted to 3D Gaussian splats with the splataverse analytic mesh→3DGS converter (derived, hyperparameter-free density: one Gaussian lattice per face, spacing set by the source texture resolution). Collected for drone/car multi-agent visual localisation and SLAM. Each folder is a different source with its own licence, so check the table before use. The Sunshine Coast Maroochydore street mesh has its own repo: maxwildersmith/splataverse-sunshine-coast.
| folder | source | capture | Gaussians | texel | size | licence |
|---|---|---|---|---|---|---|
uavscenes/HKisland |
UAVScenes (DJI Terra mesh) | drone | 733M | 3.2 cm | 5.7 GB | CC BY-NC-SA 4.0 |
uavscenes/AMtown_BlockAB |
UAVScenes (DJI Terra mesh) | drone | 962M | 3.0 cm | 7.7 GB | CC BY-NC-SA 4.0 |
aachen/AC13 |
Aachen Day-Night v1.1, MeshLoc AC13-C textured MVS mesh | ground photos (street level) | 830M | n/a | 5.7 GB | CC BY-NC-SA 4.0 |
Gallery
UAVScenes from drone height (60 m above ground, pitched 35° down):
| HKisland | AMtown BlockAB |
|---|---|
 |
 |
 |
 |
Aachen: a real database photo (left) next to the splats rendered from that photo's COLMAP pose (right). Black areas are holes in the source MVS mesh, mostly ground and transient objects.
Layout
Each folder holds lod_index.json and tiles/<k>/<n>.spz (SPZ v3, recentred per tile). Every
node in the index lists its file, its world offset (add it to the tile's positions), its bounds,
its Gaussian count, its geometric error in metres and its children. All folders here are flat: every node is a leaf. The
frame string in each index gives the coordinate frame (Y-up, metres).
The Aachen splats share the frame of the Aachen Day-Night COLMAP database model once the index origin is added, so database and query poses can be used directly as render cameras.
Attribution
- UAVScenes: Wang et al., UAVScenes: A Multi-Modal Dataset for UAVs. CC BY-NC-SA 4.0. Derived; non-commercial, share-alike.
- Aachen Day-Night: Sattler et al. 2018 / Zhang et al. 2021; mesh from MeshLoc (Panek et al. 2022). CC BY-NC-SA 4.0. Derived; non-commercial, share-alike.
- Downloads last month
- 38