You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains 3D Gaussian splats derived from the Helsinki 3D reality mesh 2017 by the City of Helsinki (CC BY 4.0). By requesting access you agree to credit © City of Helsinki as the source and to follow the CC BY 4.0 terms.
Log in or Sign Up to review the conditions and access this dataset content.
Splataverse Helsinki
Complete source tileset, verified. Helsinki, Finland: a hierarchical level-of-detail (HLOD) tree of 30,813 splat tiles (13,375 finest leaves), 5,988,621,632 Gaussians at the finest level (8,927,322,933 across all levels), in 24 archives (approximately 75.90 GB). Public dataset with manual access approval.
The leaf footprint is approximately 104.8 km²: the union of the finest-level tile bounding boxes on the ground plane, rasterised at 1/1500 of the root tile extent. This is a tile-coverage area, not occupied mesh surface or municipal area. The source is real aerial photogrammetry: the City of Helsinki 3D reality mesh 2017, built from more than 42,000 aerial photographs (summer 2017, ~7.5 cm ground sampling distance).
Archive and tile manifest · Release metadata · LOD index · Source and license · All cities
Data
| Tiles (HLOD nodes) | 30,813 (13,375 leaves; every textured source content node converted, 0 missing) |
| Gaussians, finest level | 5,988,621,632 |
| Gaussians, all levels | 8,927,322,933 |
| Tile payload | 75.88 GB (30,800 SPZ + 13 NPZ tiles) |
| Archives | 24 tar files, 1.66 to 3.76 GB each |
| Leaf footprint | 104.8 km² (union of leaf tile boxes) |
| Recipe | analytical_source_error_v1_aniso2 (Splataverse balanced preset) |
Layout:
lod_index.json: the tree.nodes[i]={file, offset, lo, hi, n, error, children};rootslists the root node ids.fileis the tile path after extraction,offsetthe tile origin in the city frame,lo/hiits axis-aligned bounds (city frame, metres),nits Gaussian count,errorthe source geometric error (metres) andchildrenthe node ids that refine it (REPLACE refinement).shards/tiles-<first dir>[-<last dir>].tar: uncompressed tars of consecutive tile directoriestiles/<id // 1000>/. Extracting every archive in the dataset root restorestiles/<id // 1000>/<id>.spzexactly as named inlod_index.json. Each tile is SPZ v4 (zstd, degree-0 SH) with positions relative to itsoffset; the 13 top-level tiles that exceed SPZ's ±2 km position range are.npz(float32xyz, rgb, log_scales, quats, opacity).shards/<name>.json: per-archive manifest: archive SHA-256 and size, and every tile with its node id, Gaussian count, byte size, SHA-256 and tar header offset (the tile bytes start 512 bytes later, so single tiles can be read with an HTTP range request without downloading the archive).manifest.json: all archives with SHA-256 and size, plus one entry per tile (id, file, archive, bytes, sha256, n).nodes_metadata.jsonl: one line per node with its conversion report (recipe, reference extent, geometry spacing, per-material texture size, mean mip level and maximum sampled band RMS, time, bytes).release_metadata.json: recipe, configuration, scene adaptation, per-level counts, footprint and validation summary.sources.json: source, license and attribution.
Only derived Gaussians are distributed; no source meshes or textures are included. The tiles hold no additional per-Gaussian channels beyond position, scale, rotation, opacity and colour.
Download
After requesting access and logging in with hf auth login, fetch the index and one archive, or the whole city:
hf download Splataverse/helsinki --repo-type dataset --include 'lod_index.json' 'manifest.json' 'shards/tiles-000-005.*' --local-dir helsinki
cd helsinki && tar -xf shards/tiles-000-005.tar # restores tiles/<dir>/<id>.spz
sha256sum shards/tiles-000-005.tar # compare with shards/tiles-000-005.json
manifest.json maps any tile id to its archive (tile_entries), so a region can be fetched by selecting nodes from
lod_index.json by their lo/hi bounds first.
Rendering by level of detail
import json, numpy as np
from splataverse.formats.spz import load_spz # or any SPZ v4 reader (nianticlabs/spz)
index = json.load(open("lod_index.json"))
node = index["nodes"][index["roots"][0]]
g = np.load(node["file"]) if node["file"].endswith(".npz") else load_spz(node["file"])
xyz_city = g["xyz"] + np.asarray(node["offset"]) # tile-local -> city frame
To render a view, start at the roots and replace a node by its children while its projected geometric error
error * focal_px / distance exceeds the pixel threshold (e.g. 1.5 px), then draw all selected tiles together. Parents
and children are alternatives (REPLACE), not additive: never draw a node together with its descendants. This is the
same screen-space-error refinement the source 3D Tiles tileset was authored for; splataverse.eval.lod_tree.load_index / select / render implements it with gsplat.
Coordinate frame
Local ENU at the tileset centre on the WGS84 ellipsoid, Y-up: x=east, y=up, z=south. Each tile stores positions relative to its own offset (lod_index.json nodes[i].offset). Frame string: ENU at lat 60.221450 lon 25.024192 (x=east, y=up, z=south). Units are metres. The source glTF content (Y-up) is rotated to
Z-up and offset by RTC_CENTER per the 3D Tiles specification, converted from ECEF to the local ENU frame and written Y-up.
Levels
| depth | geometric error (m) | tiles | Gaussians | GB | reference extent (m) | geometry floor spacing (m) |
|---|---|---|---|---|---|---|
| 0 | 128 | 1 | 23,490 | 0.00 | 25597.4 | 199.979 |
| 1 | 64 | 2 | 170,932 | 0.01 | 19367.6 | 151.310 |
| 2 | 32 | 4 | 569,655 | 0.03 | 10271.0 | 80.242 |
| 3 | 16 | 6 | 1,012,843 | 0.06 | 5489.9 | 42.890 |
| 4 | 8 | 13 | 2,390,921 | 0.02 | 2749.0 | 21.477 |
| 6 | 4 | 25 | 5,771,740 | 0.06 | 1376.9 | 10.757 |
| 7 | 2 | 76 | 17,334,452 | 0.18 | 692.5 | 5.410 |
| 8 | 1 | 260 | 56,334,166 | 0.57 | 348.4 | 2.722 |
| 9 | 0.5 | 895 | 184,536,748 | 1.76 | 176.1 | 1.376 |
| 10 | 0.25 | 3,312 | 614,183,632 | 5.52 | 89.8 | 0.701 |
| 12 | 0.125 | 12,927 | 2,070,720,613 | 17.49 | 46.0 | 0.360 |
| 13 | 0.0625 | 13,292 | 5,974,273,741 | 50.18 | 45.1 | 0.352 |
Conversion recipe
analytical_source_error_v1_aniso2: One source-error analysis and one CPU emission per glTF material (scripts/classic_sampling.convert_classic, the analytical balanced path of splat-tool mesh-to-splat --preset balanced): per-face spacing = min(reference_extent/128, texel spacing x 2^lod) where lod is the coarsest mip level whose cumulative band RMS stays <= 0.03; Gaussians are shaped to their lattice cell with major/minor sigma <= 2. No rendering, no count search, no Gaussian cap.
Scene adaptation. The geometry floor uses a per-LOD-level reference extent instead of each tile's own vertex extent: for every geometric-error level, sqrt(2) x the median bounding-sphere radius of all content nodes on that level (the side of a flat square tile with that circumradius). Neighbouring tiles of one level therefore get the same floor, and small clipped fragment tiles are not over-sampled; the texture term is per-face and unchanged.
Each Gaussian has opacity 0.99 and a colour sampled from the source texture mip level that matches its spacing. Nothing is trained or optimised.
Verification and validation scope
Every textured content node of the source tileset (30,813) was converted with the single recipe
above, and every tile file matches its recorded byte size. After upload, every archive's remote byte size and SHA-256
on the Hub was checked against manifest.json, and every tile's SHA-256 inside the archives was recorded at packing time.
Image quality is a source-error proxy (per-material maximum sampled mip-band RMS ≤ 0.03), plus a small render
comparison against the previous native-density conversion recorded in release_metadata.json; no whole-scene LPIPS
against the source mesh is claimed.
License and attribution
Source: Helsinki 3D reality mesh 2017 (Helsingin kolmioverkkomalli 2017), © City of Helsinki (Helsingin kaupunki, Kaupunginkanslia), available at https://kartta.hel.fi/3d/mesh/Helsinki_2017/ and listed on Helsinki Region Infoshare (helsingin-3d-kaupunkimalli), under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. This derived dataset is released under CC BY 4.0 as well.
Attribution: Source: © City of Helsinki, Helsinki 3D reality mesh 2017, CC BY 4.0 (https://kartta.hel.fi/3d/mesh/Helsinki_2017/). Converted to 3D Gaussian splats by Splataverse.
Changes: the textured source meshes were converted to 3D Gaussian splats and repackaged into tar archives; no source meshes or textures are redistributed.
Limitations
- Aerial photogrammetry: facades, overhangs and street-level detail are only as good as the 2017 mesh; moving objects, vegetation and water carry reconstruction artefacts.
- Colours are the source photographs' baked daylight shading (degree-0 SH only), not relightable albedo.
- Tiles overlap slightly at their borders, as in the source.
- The balanced density is chosen from source texture and geometry error, not from rendered image quality.
- Downloads last month
- 24