📄 Paper: Industrial3D: A Terrestrial LiDAR Point Cloud Dataset and Cross-Paradigm Benchmark for Industrial Infrastructure (Under Review)
Industrial3D is a large-scale, high-resolution point cloud dataset for industrial Mechanical, Electrical, and Plumbing (MEP) scene understanding. Furthermore, we benchmark 9 representative methods on the Industrial3D across 4 DL paradigms (fully supervised, weakly supervised, unsupervised, and foundation model).
- Scale: 612.7 million labeled points from 7 water treatment facilities
- Resolution: 6mm terrestrial laser scanning (TLS)
- Diversity: 12 semantic classes (MEP + structural elements)
- Authenticity: Real industrial environments with realistic occlusion, noise, and complexity
Figure: Industrial3D graphical abstract showing dataset overview and benchmark framework.
Figure: Industrial water treatment facility with UAV photography and annotated point cloud examples.
Figure: All 12 semantic classes in Industrial3D: Duct, Elbow, Flange, I-beam, Pipe, Pump, Reducer, Rectangular beam, Strainer, Tank, Tee, Valve.
Figure: Scene gallery of 4 representative scenes in Industrial3D.
Figure: Additional scenes in Industrial3D.
Figure: Statistical distribution of 612.7M labeled points across 3 tiers. Industrial3D has a 215:1 class imbalance (head:tail), 3.5× more severe than S3DIS.
20 unique rooms across 13 areas. 4 representative scenes:
| # | Area | Room | Split | Video |
|---|---|---|---|---|
| 1 | Area 2 | Service Gallery | Train | Watch ▶ |
| 2 | Area 12 | SPH Pump Room | Test | Watch ▶ |
| 3 | Area 6-1 | 93m Psu | Test | Watch ▶ |
| 4 | Area 3 | 93m Tank | Train | Watch ▶ |
Area 2: Service Gallery (Train) - Largest at 79.6M points with highest MEP density
Area 12: SPH Pump Room (Test) - Test set representative with compact equipment
Area 6-1: 93m Psu (Test) - Test set with moderate complexity
Area 3: 93m Tank (Train) - Large tank structure showing geometric diversity
Why these 4:
- Service Gallery: Largest at 79.6M points with highest MEP density
- SPH Pump Room: Test set representative with compact equipment
- 93m Psu: Test set with moderate complexity
- 93m Tank: Large tank structure showing geometric diversity
📁 Full dataset coming soon! All 20 rooms (RGB + ground truth videos and renders) will be uploaded to Google Drive upon paper acceptance. Current preview materials available (videos, renders). Stay tuned for the complete collection!
| Metric | Value |
|---|---|
| Total Scanned Area | 20,000+ m² (estimated) |
| Labeled Points | 612.7 million |
| Raw Scan Data | 2.3+ billion points |
| Point Density | 6mm TLS resolution |
| Annotation Effort | 754 person-hours |
| Facilities | 13 type of water treatment facilities |
| Areas | 13 areas |
| Rooms | 20 unique rooms/scenes |
Fully-Supervised (6 methods):
- KPConv
- PosPool
- RandLA-Net
- ResPointNet++
- PTv3 (Point Transformer V3)
- Boundary-CB
Weakly-Supervised (1 method):
- SQN (0.1% labels, 0.01% labels)
Unsupervised (1 method):
- GrowSP
Foundation Models (1 method):
- Point-SAM (Oracle, One-vs-Rest)
| Paradigm | Method | mIoU |
|---|---|---|
| Fully-Supervised | Boundary-CB | 55.74% |
| Fully-Supervised | KPConv | 53.65% |
| Fully-Supervised | PosPool | 53.18% |
| Fully-Supervised | ResPointNet++ | 52.48% |
| Fully-Supervised | PTv3 | 41.90% |
| Fully-Supervised | RandLA-Net | 39.83% |
| Weakly-Supervised | SQN (0.1% labels) | 44.29% |
| Weakly-Supervised | SQN (0.01% labels) | 33.16% |
| Unsupervised | GrowSP | 11.73% |
| Foundation Model | Point-SAM (Oracle) | 21.08% |
| Foundation Model | Point-SAM (One-vs-Rest) | 15.79% |
@article{yin2026industrial3d,
title={Industrial3D: A Terrestrial LiDAR Point Cloud Dataset and Cross-Paradigm Benchmark for Industrial Infrastructure},
author={Yin, Chao and Yue, Hongzhe and Han, Qing and Hu, Difeng and Liang, Zhenyu and Lin, Fangzhou and Sun, Bing and Wang, Boyu and Li, Mingkai and Yao, Wei and Cheng, Jack C.P.},
journal={arXiv preprint arXiv:2603.28660},
year={2026}
}@article{yin2021,
title={Automated semantic segmentation of industrial point clouds using ResPointNet++},
author={Yin, Chao and Wang, Boyu and Gan, Vincent JL and Wang, Mi and Cheng, Jack CP},
journal={Automation in Construction},
volume={130},
pages={103874},
year={2021},
publisher={Elsevier},
doi={10.1016/j.autcon.2021.103874}
}
@article{yin2023,
title={Label-efficient semantic segmentation of large-scale industrial point clouds using weakly supervised learning},
author={Yin, Chao and Yang, Bo and Cheng, Jack CP and Gan, Vincent JL and Wang, Boyu and Yang, Ji},
journal={Automation in Construction},
volume={148},
pages={104757},
year={2023},
issn = {0926-5805},
doi={10.1016/j.autcon.2023.104757},
}
@article{Yin2026arXiv,
title={Resolving Primitive-Sharing Ambiguity in Long-Tailed Industrial Point Cloud Segmentation via Spatial Context Constraints},
author={Yin, Chao and Han, Qing and Hou, Zhiwei and Liu, Yue and Dai, Anjin and Hu, Hongda and Yang, Ji and Yao, Wei},
journal={arXiv preprint arXiv:2601.19128},
eprint={2601.19128},
year={2026}
}