3d bridge segmentation

The objective of the present project is the automated creation of digital twins for existing bridge structures. To this end, the initial step involves semantic segmentation of the provided point clouds prior to subsequent processing. This repository contains:

1. A set of 20 real-world bridges scanned and manually annotated
2. A set of 23 synthetic bridges that were scanned and annotated automatically
3. Simple basic starting code to train 3 different 3D models for large 3D point cloud scenes

For more details we refer to the following two paper:

• 3D bridge segmentation using semi-supervised domain adaptation
• SemanticBridge - A Dataset for 3D Semantic Segmentation of Bridges and Domain Gap Analysis

Code

The code is containerized. For development, VS Code DevContainers can be used. Easily use the Remote Development extension.

Setup

Build the container with the tag <3d_bridge_segmentation>, rename the .devcontainer/devcontainer.example to .json. Update and mount your data-folder into the docker:

	},
	// Uncomment to use the Docker CLI from inside the container. See https://aka.ms/vscode-remote/samples/docker-from-docker.
	"mounts": [
		"source=<path-to-data>,target=/data,type=bind"
	]

Afterwards, you have to compile the cuda/c++ dependencies:

cd lib/
./setup.sh

Configure the required config.yml file (examples can be found in config directory) and simply run:

python3 train.py <config.yml>

Data

We scanned 20 real world bridges. 10 located in Germany and 10 in UK. In all cases a Terrestrial Laser Scanning (TLS) scanner was used. Additionally we scanned 7 bridges with a Mobile Laser Scanning (MLS) scanner.

Data split

• Total real bridge point clouds: 20
• Training set: 15
• Testing set: 5

Real Data

For lightweight GitHub visualization, training data examples are shown below (15 samples, 5 per row). In each cell: top is point cloud preview, bottom is semantic annotation (_gt).

bridge_2_cb_faro	bridge_3_cb_faro	bridge_4_cb_faro	bridge_5_cb_faro	bridge_6_cb_faro
bridge_8_cb_faro	bridge_9_cb_faro	bridge_10_cb_faro	bridge_11_fr_rtc	bridge_12_fr_rtc
bridge_14_fr_rtc	bridge_15_fr_rtc	bridge_16_fr_rtc	bridge_18_fr_rtc	bridge_20_fr_rtc

The following real-world bridges are for testing purposes:

• bridge_1_cb_faro
• bridge_7_cb_faro
• bridge_13_fr_rtc/blk
• bridge_17_fr_rtc/blk
• bridge_19_fr_rtc/blk

Test data examples are shown below (point cloud on top, semantic annotation _gt on bottom).

bridge_1_cb_faro

bridge_7_cb_faro

bridge_13_fr_rtc

bridge_17_fr_rtc

bridge_19_fr_rtc

In total, we offer 20 different scenes of bridges. We focus on the following classes:

ID	Class name	TLS points	MLS points
0	unlabeled	4.848.483	716.059
1	ground	106.884.341	29.217.777
2	high vegetation	45.193.593	19.097.682
3	abutment	8.994.167	2.575.750
4	superstructure	37.847.082	7.657.857
5	top surface	19.846.553	8.828.246
6	railing	11.715.179	3.905.237
7	traffic sign	223.234	155.333
8	pillar	9.544.994	397.365

Synthetic

In addition, we offer 23 different scenes of bridges. For simplicity we focused on the following classes:

ID	Class name	nr points
0	unlabeled	628.371
1	ground	66.630.880
2	high vegetation	10.532.641
3	abutment	10.838.768
4	superstructure	42.497.680
5	top surface	48.302.362
6	railing	2.169.738
7	traffic sign	341.185
8	pillar	3.462.674

Download

The synthetic data can be downloaded via the mobilithek or directly via zenodo

The real data can be downloaded via the mobilithek or directly via zenodo

Citation

This synthetic data was created for the work presented in 3D bridge segmentation using semi-supervised domain adaptation. The real world data are created in cooperation with the Department of Engineering, Cambridge. If you find our work useful in your research, please consider citing:

@article{SemanticBridge,
title = {SemanticBridge—A dataset for 3D semantic segmentation of bridges and domain gap analysis},
journal = {Developments in the Built Environment},
volume = {26},
pages = {100912},
year = {2026},
issn = {2666-1659},
doi = {https://doi.org/10.1016/j.dibe.2026.100912},
url = {https://www.sciencedirect.com/science/article/pii/S2666165926000700},
author = {Maximilian Kellner and Mariana {Ferrandon Cervantes} and Yuandong Pan and Ruodan Lu and Ioannis Brilakis and Alexander Reiterer},
}

@article{SynthBridge,
	title = {3D bridge segmentation using semi-supervised domain adaptation},
	journal = {Automation in Construction},
	volume = {172},
	pages = {106021},
	year = {2025},
	issn = {0926-5805},
	doi = {https://doi.org/10.1016/j.autcon.2025.106021},
	url = {https://www.sciencedirect.com/science/article/pii/S0926580525000615},
	author = {Maximilian Kellner and Timothy König and Jan-Iwo Jäkel and Katharina Klemt-Albert and Alexander Reiterer},
	keywords = {3D deep learning, Bridge segmentation, Point cloud generation, Point cloud processing, Domain adaptation},
}

Models

Sparse ResUNet

3D UNet with residual connections using sparse convolutions.

@misc{spconv2022,
    title={Spconv: Spatially Sparse Convolution Library},
    author={Spconv Contributors},
    howpublished = {\url{https://github.com/traveller59/spconv}},
    year={2022}
}

KPConv

Kernel Point Convolution, which is a special designed point convolution operator for processing point clouds.

@article{thomas2019KPConv,
    Author = {Thomas, Hugues and Qi, Charles R. and Deschaud, Jean-Emmanuel and Marcotegui, Beatriz and Goulette, Fran{\c{c}}ois and Guibas, Leonidas J.},
    Title = {KPConv: Flexible and Deformable Convolution for Point Clouds},
    Journal = {Proceedings of the IEEE International Conference on Computer Vision},
    Year = {2019}
}

Point Transformer V2

Transformer architecture designed specifically for point cloud data, updated with grouped attention and grid-based pooling.

@inproceedings{wu2022point,
  title     = {Point transformer V2: Grouped Vector Attention and Partition-based Pooling},
  author    = {Wu, Xiaoyang and Lao, Yixing and Jiang, Li and Liu, Xihui and Zhao, Hengshuang},
  booktitle = {NeurIPS},
  year      = {2022}
}