MALiBU3D

A Multitask Aerial LiDAR Benchmark for Large-Scale 3D Scene Understanding.

Paper (arXiv, TODO): https://arxiv.org/abs/XXXX.XXXXX

Training code (GitHub, TODO): https://github.com/louisgeist/MALiBU3D

MALiBU3D is a large-scale multitask ALS benchmark: 59 billion LiDAR points over 2,221 km² of metropolitan France (overlap of the national LiDAR HD programme and IGNF/FLAIR-HUB). It spans urban, agricultural, forested, mountainous, and coastal landscapes, with point-wise intensity and aerial RGB.

Supervision covers five complementary tasks:

• land-cover segmentation (15 classes)
• forest-cover segmentation
• natural-habitat distribution (four ecological axes)
• road-network prediction
• elevation regression

Zones are ~1 km² and split into 100 × 100 m tiles. Train / val / test follow the FLAIR-HUB departmental split.

This Hub repo ships the training-ready arrays (NumPy .npy inside one zip per ROI), not the raw GeoTIFFs / PLYs. Unzip onto local scratch for training (np.load / mmap). The zip is the distribution format.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ORG/MALiBU3D",
    repo_type="dataset",
    allow_patterns=["data/train/D075-2021_LIDARHD/*", "labels.json", "tiles.csv"],
)

Do not load_dataset() on the point clouds (variable N ≈ 1e5–3e5). The Dataset Viewer reads tiles.parquet (one row per tile, no xyz). tiles.csv is the same table when parquet was not built. Tiles with missing LiDAR coordinates are omitted from the catalog.

A FLAIR-HUB patch that is not in tiles.csv has no LiDAR HD coverage (or, rarely, its LiDAR coordinates were missing): MALiBU3D only ships tiles with LiDAR.

A tiny extracted ROI lives under toy/ for inspection.

Layout

labels.json
palettes.json
tiles.parquet                # tile table, read by the Hub viewer
tiles.csv                    # same table (split, flags, zip_path, n_points, forest georef, …)
val_subset_2000tiles.csv     # fixed 2,000-tile validation subset (split, tile_id)
SHA256SUMS
toy/                          # one extracted ROI
data/{train,val,test}/{dept}_LIDARHD/{roi}.zip

Each {roi}.zip (flat, no wrapping {roi}/ folder):

{tile_id}/coord.npy                 float32 (N, 3)   XYZ relative
{tile_id}/coord_translation.npy     float64 (3,)     Lambert-93 offset
{tile_id}/color.npy                 uint8   (N, 3)
{tile_id}/segment.npy               uint8   (N,)     land cover, Void=15
{tile_id}/strength.npy              float32 (N,)     LiDAR intensity ~[0, 1]
{tile_id}/elevation.npy             float32 (N,)     z − DTM, if present
{tile_id}/natural_habitat.npy       uint8   (N, 4)   ecological axes, if present
{tile_id}/forest_2d.npy             uint8   (1, H, W)
{deptcode}_{roi}_ROADS_graph.gpkg    EPSG:2154, if the ROI has roads

Absolute coordinates: xyz_abs = coord + coord_translation (EPSG:2154).

Not every tile has a DTM or CarHab coverage, and not every ROI has roads. elevation.npy is omitted where has_elevation is false, natural_habitat.npy where has_natural_habitat is false, and the GeoPackage where has_roads_graph is false (columns of tiles.csv).

Not included: per-tile meta.json, network.npy, per-point forest.npy, land_use.npy, rail / transmission-line graphs.

Tile table (tiles.csv)

tiles.csv / tiles.parquet is the single tile table: one row per shipped 100 m tile, every row with LiDAR.

column	content
split, dept_year, roi, scene_i_j	FLAIR-HUB split and location
tile_id	{dept_year}_{roi}_{scene_i_j} (= FLAIR-HUB patch id)
zip_path	relative ROI zip (data/train/D075-2021_LIDARHD/UU-S1-4.zip)
has_natural_habitat, has_elevation	natural_habitat.npy / elevation.npy present
has_roads_graph, roads_gpkg	ROI roads GeoPackage present / its file name
date_aerial_rgb, date_lidarhd, date_gap_days	acquisition dates
n_points, n_voxels	points per tile / occupied 0.1 m voxels
forest_origin_x, forest_origin_y, forest_width, forest_height	forest_2d grid georeferencing

Reconstruct a tile directory after unzipping:

{data_root}/{split}/{dept_year}_LIDARHD/{roi}/{tile_id}

forest_2d cell size (pixel_m = 1.0), CRS, south-up axis, and class values are global (labels.json).

val_subset_2000tiles.csv lists a fixed subset of 2,000 validation tiles, representative of the label distribution of the whole validation split. The reference training code validates on it to save time. Test results always use the full test split.

Labels

See labels.json. Land cover (segment.npy): 15 train classes + Void=15.

Natural habitat (natural_habitat.npy): (N, 4) uint8, already remapped from CarHab to the 4 ecological axes. Column order is labels.json → natural_habitat.columns:

col	key	classes	Void
0	nathab_habitat_type	Open, Forest, Mineral, Aquatic	4
1	nathab_moisture_regime	Humid, Mesic, Dry	3
2	nathab_soil_chemistry	Acidic, Alkaline	2
3	nathab_bioclimatic_zone	Temperate, Mediterranean, Alpine	3

import json, numpy as np
labels = json.load(open("labels.json"))
nh = np.load("natural_habitat.npy")          # (N, 4)
moisture = nh[:, 1]                          # Void = 3
# Pointcept: no LUT. Assign column i to task labels["natural_habitat"]["columns"][i].

forest_2d → points

South-up Lambert-93 grid. Origin and size come from the catalog; cell size is global (labels.json → forest_2d.pixel_m = 1.0). Width/height vary slightly per tile.

import json
import numpy as np
import pandas as pd

pixel_m = json.load(open("labels.json"))["forest_2d"]["pixel_m"]
row = pd.read_csv("tiles.csv").set_index("tile_id").loc[tile_id]
coord = np.load("coord.npy")
t = np.load("coord_translation.npy")
raster = np.load("forest_2d.npy")  # (1, H, W), 0 / 1 / 2=void
x = coord[:, 0] + t[0]
y = coord[:, 1] + t[1]
ix = np.floor((x - row.forest_origin_x) / pixel_m).astype(int)
iy = np.floor((y - row.forest_origin_y) / pixel_m).astype(int)
h, w = int(row.forest_height), int(row.forest_width)
inside = (ix >= 0) & (ix < w) & (iy >= 0) & (iy < h)
out = np.full(len(coord), 2, dtype=np.uint8)
out[inside] = raster[0, iy[inside], ix[inside]]

Road graphs: GeoPackage layers nodes, edges, metadata, coordinates absolute EPSG:2154.

Licence and attribution

TODO: licence to be defined (see LICENSE).