license: cc-by-sa-4.0
pretty_name: LaMAR splats
tags:
- gaussian-splatting
- visual-localization
- lamar
- indoor
LaMAR as Gaussian splats
Gaussian-splat conversions of the NavVis scan meshes of the LaMAR benchmark (Sarlin et al., LaMAR: Benchmarking Localization and Mapping for Augmented Reality, ECCV 2022), one top-level folder per location: hge/ (ETH main building), cab/ (ETH CAB building) and lin/ (Lindenhof, an old-town area of Zurich).
Original data: https://cvg-data.inf.ethz.ch/lamar/, (c) ETH Zurich Computer Vision and Geometry Group and Microsoft, licensed CC BY-SA 4.0.
Changes were made. The dataset's vertex-colored NavVis meshes were converted to Gaussians, spatially tiled and stored as SPZ v4 (details below). These derived splats are released under the same license, CC BY-SA 4.0, with attribution to the original authors.
No source meshes, point clouds, images or session files are included. sources.json lists where the originals are (URLs, member paths, sizes, sha256).
Contents
| Session | Source mesh | Source triangles | Gaussians | SPZ GB |
|---|---|---|---|---|
navvis_2022-02-06_12.55.11 |
mesh_simplified |
89,028,324 | 166,800,456 | 1.71 |
navvis_2022-02-26_16.21.10 |
mesh_simplified |
88,930,914 | 168,329,046 | 1.73 |
navvis_2021-02-18_10.02.08 |
mesh |
337,935,451 | 546,820,597 | 5.38 |
navvis_2021-02-18_11.33.08 |
mesh |
270,605,718 | 460,777,974 | 4.50 |
navvis_2021-05-10_19.34.30 |
mesh |
404,160,039 | 647,286,891 | 6.29 |
cab/ and lin/
| Location | Session | Source mesh | Source triangles | Gaussians | SPZ GB |
|---|---|---|---|---|---|
| CAB | navvis_2022-06-21_09.28.22 |
mesh_simplified |
80,791,650 | 153,009,837 | 1.54 |
| CAB | navvis_2022-06-25_11.14.36 |
mesh_simplified |
137,074,217 | 260,550,686 | 2.67 |
| LIN | navvis_2022-07-03_08.30.21 |
mesh_simplified |
191,194,343 | 364,232,909 | 3.73 |
Only sessions that ship a source mesh are converted (the HoloLens and iPhone sessions of these locations have none).
Layout: <location>/<session>/index.json (tile list, tile-local SPZ positions plus offset, Gaussian counts, recipe) and <location>/<session>/tiles/tNNN.spz.
Frame: positions are in the LaMAR global frame of the location. Each session mesh was transformed by the pose_graph_optimized, __absolute__ pose of its proc/alignment_global.txt (stored as to_global, qw qx qy qz tx ty tz, in every index.json). Add the tile offset to the SPZ positions.
How the splats were made
CPU-only, deterministic, no rendering or learned estimation. Recipe analytical_source_error_v1_edge2t1.5_aniso2 of the Splataverse mesh-to-splat converter: one Gaussian per source triangle, refined on elongated triangles (edge lattice 2, threshold 1.5), Gaussians shaped to their cell (anisotropy cap 2), color from barycentrically interpolated vertex colors, opacity 0.99. Tiles are an adaptive XY partition of at most 3M source triangles. The recipe and its options are recorded in each tile receipt and index.json.
hge/ 2022 sessions and all cab/ and lin/ sessions use the dataset's mesh_simplified.ply; the two 2021-02 and the 2021-05 hge/ sessions use the full mesh.ply. We did not decimate any mesh.
Known limitations
- Appearance is per-vertex color from the dataset's meshes (no textures or materials), so surfaces are softer than the original imagery and carry no view-dependent effects.
- Glass is reconstructed with spurious geometry in the source meshes; windows show odd shapes and bright blobs. This comes from the source scan, not from the conversion.
- Sessions overlap in space and time; they are not merged or deduplicated.