3DSES: Indoor TLS Point Cloud Segmentation Dataset

This repository provides a mirror of the 3DSES dataset to facilitate access for research and benchmarking purposes, with per-tier folders for selective downloads.

Disclaimer This dataset is not owned nor created by the maintainer of this Hugging Face repository. The official and authoritative source is the original Zenodo record: https://zenodo.org/records/13323342

3DSES (3D Segmentation of ESGT point clouds) is a dataset of dense indoor Terrestrial Laser Scanning (TLS) colorized point clouds covering 427 m² of an engineering school (ESGT, Le Mans, France). It features dual annotation: semantic labels at the point level alongside a complete 3D CAD building model.

Dataset Summary

	Stations	Points	Intensity	Manual labels	Pseudo-labels	Columns	Size
Gold	7	44.5M	Yes	Yes	Yes	9	3.0 GB
Silver	27	195.5M	Yes	Yes	Yes	9	14 GB
Bronze	39	384.9M	Yes	No	Yes	8	23 GB
test_area	3	20.7M	No	No	Yes	7	1.1 GB

Total: ~645M points across 39 unique scan stations.

File Format

Each .npy file contains one scan station as a NumPy array (float64).

Gold / Silver (9 columns):

Column	Content
0-2	X, Y, Z (local survey coordinates)
3-5	R, G, B (0-255)
6	Intensity (normalized 0-1)
7	Manual semantic label
8	Pseudo-label (from CAD model alignment)

Bronze (8 columns): same as above without column 7 (no manual labels).

test_area (7 columns): X, Y, Z, R, G, B, pseudo-label only.

Directory Structure

Gold/           # 7 stations with full annotations
  S163.npy ... S179.npy
Silver/         # 27 stations with full annotations
  S150.npy ... S179.npy, S295.npy, S296.npy
Bronze/         # 39 stations with pseudo-labels only
  S140.npy ... S179.npy, S295.npy, S296.npy
test_area/      # 3 held-out stations for evaluation
  S170.npy, S171.npy, S180.npy
3DSES_cad_model.obj   # 3D CAD building model (320 MB)
Illustration.png      # Dataset illustration

Usage

import numpy as np
from pathlib import Path

# Load a single station
data = np.load("Gold/S163.npy")
xyz = data[:, :3]          # coordinates
rgb = data[:, 3:6]         # color (0-255)
intensity = data[:, 6]     # normalized intensity
labels = data[:, 7]        # manual semantic labels
pseudo = data[:, 8]        # pseudo-labels from CAD

# Load with projax3d
from projax3d.opendata import get_provider
provider = get_provider("3dses")
scene = provider.load_scene("gold", output_dir="./data/3dses")

Tiers

• Gold: Stations with the highest annotation quality (manually verified labels
  • pseudo-labels + intensity). Best for training and evaluation.
• Silver: Larger set with manual labels. Superset of Gold stations.
• Bronze: All stations with pseudo-labels only (automatically generated from CAD model alignment, >95% accuracy). Largest coverage.
• test_area: Held-out stations for benchmarking (pseudo-labels only).

Citation

@inproceedings{merizette2025_3dses,
  title     = {{3DSES}: an indoor Lidar point cloud segmentation dataset
               with real and pseudo-labels from a {3D} model},
  author    = {M{\'e}rizette, Maxime and Audebert, Nicolas and
               Kervella, Pierre and Verdun, J{\'e}r{\^o}me},
  booktitle = {Proceedings of the 20th International Joint Conference on
               Computer Vision, Imaging and Computer Graphics Theory and
               Applications (VISAPP)},
  year      = {2025},
  address   = {Porto, Portugal},
  note      = {arXiv:2501.17534}
}

License

This dataset is distributed under the Creative Commons Attribution Share Alike 4.0 International (CC-BY-SA 4.0) license.

Links

• Paper: arXiv:2501.17534
• Original dataset: Zenodo (DOI: 10.5281/zenodo.13323342)
• Code: github.com/merizetm/3dses
• Benchmark: Codabench competition

Authors

• Maxime Merizette (ESGT / QUARTA)
• Nicolas Audebert (LaSTIG, IGN-ENSG)
• Pierre Kervella (QUARTA / ESGT)
• Jerome Verdun (ESGT)

Contributors: Lea Corduri, Judicaelle Djeudji Tchaptchet, Damien Richard, Lilian Ribet, Elisabeth Simonetto.