heig-vd-geo/3DSES

---
license: cc-by-sa-4.0
task_categories:
- token-classification
tags:
- point-cloud
- lidar
- 3d
- semantic-segmentation
- indoor
- terrestrial-laser-scanning
- BIM
pretty_name: "3DSES: Indoor TLS Point Cloud Segmentation"
size_categories:
- 100M<n<1B
---

# 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

```python
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

```bibtex
@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)](https://creativecommons.org/licenses/by-sa/4.0/legalcode)
license.

## Links

- **Paper**: [arXiv:2501.17534](https://arxiv.org/abs/2501.17534)
- **Original dataset**: [Zenodo (DOI: 10.5281/zenodo.13323342)](https://zenodo.org/records/13323342)
- **Code**: [github.com/merizetm/3dses](https://github.com/merizetm/3dses)
- **Benchmark**: [Codabench competition](https://www.codabench.org/competitions/6927/)

## 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.