mbendjilali/DALES-2

DALES 2

Dataset summary

DALES 2 is a renovated large-scale aerial LiDAR benchmark for 3D scene understanding, built on the geographic coverage of the original DALES release with re-annotated semantic and instance labels, refined taxonomy, and companion scene-graph structure (objects and relations) described in the USM3D workshop paper at CVPR 2026.

This Hub repository distributes the annotated point clouds (typically LAZ/LAS per tile). Graph JSON and the interactive graph editing tool live in the companion code repository (see Scene graph tool).

What is new vs. DALES Objects

• Human-in-the-loop revision of instances and semantics (e.g. powerline granularity per span, tree instances promoted from vegetation, fence as non-instantiable stuff).
• 15-class semantic taxonomy with per-point classification and cross-class unique instance IDs within each tile (see Class IDs).
• Optional use with relational graphs (edges such as adjacent, near, support, extension) when generated with the official pipeline.

Supported tasks

• Semantic and instance segmentation on aerial LiDAR
• 3D detection / panoptic-style parsing in urban / infrastructure scenes
• Scene graph generation and reasoning (with external graph assets and tools)

Data modality

• 3D point clouds (LAZ or LAS), aerial / large-footprint tiles

Class IDs

ID	Class	Notes (high level)
0	Ground	Stuff
1	Vegetation	Stuff
2	Car	Vehicle instance
3	Powerline	Conductor instance
4	Fence	Stuff
5	Tree	Tree instance
6	Pick-up	Vehicle instance
7	Van & Truck	Vehicle instance
8	Heavy-duty	Vehicle instance
9	Utility pole	Pole instance
10	Light pole	Pole instance
11	Traffic pole	Pole instance
12	Habitat	Building instance
13	Complex	Building instance
14	Annex	Building instance

Instance field: Each point carries an instance attribute. Non-stuff objects use distinct positive instance IDs; stuff classes use instance 0 in the released convention. Instance IDs are unique across semantic classes within a tile (one ID does not appear on two different classes in the same file).

Dataset structure (repository layout)

Layout may evolve between releases; a typical layout is:

• One file per spatial tile, e.g. *.laz / *.las at the repository root or under a tiles/ (or similar) prefix.
• Optional sidecar assets (metadata, splits, checksums) as you add them on the Hub.

After uploading, configure Dataset > Files and versions so the Hub (and datasets) can see supported patterns if you add a load_dataset script later. Until then, use huggingface_hub to download files (below).

Download

huggingface_hub (recommended for raw files)

pip install huggingface_hub

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mbendjilali/DALES-2",
    repo_type="dataset",
    local_dir="dales2",
)

Git LFS

git lfs install
git clone https://huggingface.co/datasets/mbendjilali/DALES-2

Scene graph tool

Scene graphs (graph_*.json, graph_*_edges.json, geom_*.json) and the web editor are maintained in the Github codebase. Documentation, schema, and build instructions are in the project README shipped with that repository.

Citation

If you use this dataset, please cite:

@inproceedings{bendjilali2026dales2,
  title     = {{DALES} 2: A Renovated Aerial {LiDAR} Benchmark for 3D Scene Understanding},
  author    = {Bendjilali, Moussa and Peyran, Claire and Velumani, Kaaviya and Mauri, Antoine and Luminari, Nicola and Alliez, Pierre},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
  year      = {2026},
  note      = {USM3D workshop},
}

(Update pages or volume fields when official workshop proceedings are available.)

License

This dataset is released under the MIT License, consistent with the Hub dataset metadata.

Acknowledgments

DALES 2 builds on the original DALES aerial LiDAR data and the DALES Objects instance benchmark. Please cite the original DALES and DALES Objects papers when comparing to prior work on that geography.

Contact

For questions about the release or benchmark, open a discussion on this dataset page or contact the authors via the affiliations listed in the paper.