TrueCity: Real and Simulated Urban Data for Cross-Domain 3D Scene Understanding

Accepted to 3DCV 2025

| PDF | | arXiv | | Code | Dataset (Coming Soon)

TrueCity introduces real-world annotated point clouds, a semantic 3D city model, and 3D-model simulated point clouds of the same location, enabling coherent evaluation of the sim-to-real domain gap in 3D scene understanding.

Abstract

We introduce TrueCity, the first urban semantic segmentation benchmark with cm-accurate annotated real-world point clouds, semantic 3D city models, and annotated simulated point clouds representing the same city. TrueCity proposes segmentation classes aligned with international 3D city modeling standards, enabling consistent evaluation of synthetic-to-real gap. Our extensive experiments on common baselines quantify domain shift and highlight strategies for exploiting synthetic data to enhance real-world 3D scene understanding.

TrueCity Benchmark Dataset

Real-world point cloud (2nd row), which was manually labeled according to the class list, used for manual modeling of semantic 3D models (3rd row), which in turn were used to simulate and auto-label synthetic point clouds (4th row).

TrueCity proposes a class list of 12 classes harmonized with the standards CityGML 2.0 and OpenDRIVE 1.4:

Experiments

Top-down schematic of S--R mixtures along a continuous streetscape. Solid lines mark train/validation/test splits; dashed lines mark boundaries between contiguous synthetic and real segments for each mixture ratio.

Results

Model	100S--0R		75S--25R		50S--50R		25S--75R		0S--100R
	mIoU	OA	mIoU	OA	mIoU	OA	mIoU	OA	mIoU	OA
Point-based
PointNet	6.03	30.36	10.74	48.10	10.89	49.29	13.10	47.99	14.51	49.82
PointNet++	9.72	34.39	20.95	62.80	23.18	65.36	25.38	63.27	23.39	63.15
RandLA-Net	8.98	35.40	13.25	50.32	15.73	59.37	16.89	57.09	17.71	54.57
Kernel-based
KPConv	15.84	50.07	21.55	62.08	28.50	61.62	22.33	61.92	29.90	62.80
Transformer-based
Point Transformer v1	16.30	57.54	19.79	60.29	23.43	67.54	24.66	68.70	28.89	67.98
Point Transformer v3	14.13	53.15	19.29	60.22	25.30	65.94	24.64	65.72	25.24	60.75
Superpoint Transformer	14.31	54.17	17.01	58.62	14.22	54.63	19.61	56.98	15.96	54.64
OctFormer	13.07	53.30	14.17	55.34	14.22	49.71	13.91	50.97	17.65	56.28

Resources

• Paper: PDF, arXiv
• Code: https://github.com/tum-gis/TrueCity
• Dataset: Coming Soon

Citation

If you use TrueCity in your research, please cite our work:

@inproceedings{truecity2025,
  title={TrueCity: Real and Simulated Urban Data for Cross-Domain 3D Scene Understanding},
  author={Duc Nguyen and Yan-Ling Lai and Qilin Zhang and Prabin Gyawali and Benedikt Schwab and Olaf Wysocki and Thomas H. Kolbe},
  booktitle={International Conference on 3D Vision (3DCV)},
  year={2025},
  url={https://arxiv.org/abs/2511.07007}
}

Contact

For questions or issues, please open an issue on the GitHub repository.

---