UnderOneFacade: Worldwide Facade Semantic Segmentation Benchmark Dataset

1Technical University of Munich, 2University of Cambridge, 3University of Nottingham, 4National University of Singapore

*Equal contribution

ECCV 2026

UnderOneFacade overview

Building on ZAHA, UnderOneFacade is — to date — the largest benchmark for facade semantic segmentation of point clouds.

UnderOneFacade in 2 Minutes

A quick walkthrough of the dataset and benchmark.

📖 Abstract

Globally consistent semantic digital twins require centimeter-accurate and geographically transferable 3D facade segmentation. However, progress in facade parsing is limited by the lack of large-scale, standardized benchmarks for evaluating cross-domain generalization. Existing datasets are geographically narrow, semantically inconsistent, or insufficiently precise. We introduce UnderOneFacade, the largest cross-country and cross-continent 3D facade benchmark to date, comprising centimeter-accurate point clouds with hierarchical, harmonized, and architecturally grounded semantic labels totaling 2.7 billion annotated points. Through a systematic evaluation of representative point-, graph- and transformer-based architectures, we show that current methods struggle to recognize fine-grained architectural elements and degrade significantly across geographic domains, with the best models achieving only up to 33 IoU on the fine-grained LoFG3 benchmark. By combining geometric precision with standardized semantics at unprecedented scale, UnderOneFacade establishes a rigorous benchmark for developing robust and transferable 3D segmentation models. The dataset, evaluation scripts, and pretrained models will be released upon publication.

2.7 B annotated points
3 continents (UK, Germany, Singapore)
15 / 5 classes at LoFG3 / LoFG2
8 benchmarked architectures

🌟 Highlights

  • Building on ZAHA, we extend the LoFG (Level of Facade Generalization) taxonomy from a single city to a harmonized scheme spanning continents
  • Combines Victorian, Haussmann, and Southeast Asian colonial/modern architectural styles across the UK, Germany, and Singapore subsets
  • Centimeter-accurate geometry from multi-sensor acquisition — Leica RTC360, Leica BLK360, and the MODISSA mobile mapping platform
  • Reveals strong cross-continental domain shift: several models degrade by more than 30 F1 points between European and Asian facades
  • ❗ The dataset link above provides the data at original point cloud density; for region-specific subsets (Singapore / Germany / UK), see here

📽️ Preview

A short overview of the UnderOneFacade dataset and benchmark.

Facade segmentation preview

🔎 Facade Semantic Segmentation Results

Facade Semantic Segmentation LoFG3
Facade Semantic Segmentation LoFG2

Figure left: Qualitative facade segmentation results on UnderOneFacade on the LoFG3. We visualize predictions of representative architectures across scenes from different countries. Rows show example facades, while columns correspond to different segmentation models. Despite correct segmentation of dominant structures such as walls and roofs, models struggle to consistently recognize fine-grained facade elements, including windows, doors, and decorative components. These examples illustrate the challenges posed by long-tailed semantics and cross-country architectural variability.

Figure right: Qualitative facade segmentation results on UnderOneFacade at the LoFG2 level. Compared to the LoFG3 level, the aggregated LoFG2 hierarchy allows models to more reliably capture dominant structural regions such as structural or floor. However, inconsistencies remain in regions corresponding to facade openings and decorative structures, which are often confused with surrounding structural elements. These examples illustrate that although semantic aggregation improves visual consistency, segmentation errors persist due to architectural variability and the long-tailed distribution of facade components.

📋 Per-Class Performance Comparison

Table: Per-class F1 performance comparison on the UnderOneFacade dataset (Results under an updated, method-specific training protocol. See more details in the main paper.). The best result in each row is shown in bold.

Metric/Class PointNet++ OctFormer Superpoint Transformer PTv1 PTv3 KPConv DGCNN GrowSP
LoFG3 (fine facade semantics)
OA60.957.960.262.558.343.743.790.6
μP47.245.148.048.745.341.539.736.4
μR41.638.946.642.537.832.846.931.3
μF142.539.845.344.137.831.438.131.3
μIoU29.727.332.231.527.321.025.725.4
wall70.571.969.973.665.237.849.170.3
window37.228.134.258.65.130.128.24.6
door13.910.317.717.316.40.524.213.9
balcony19.341.751.048.96.43.121.25.7
molding46.89.632.547.28.733.824.328.3
deco10.029.512.34.39.24.34.75.5
column45.729.549.132.743.349.039.85.4
arch42.631.543.530.552.831.164.10.9
stairs6.51.31.80.20.62.16.80.0
ground surface54.151.645.953.957.939.461.40.0
terrain70.467.772.873.175.072.071.789.8
roof68.870.371.563.569.350.655.876.9
blinds19.523.327.814.412.72.015.55.4
interior61.953.678.477.074.646.745.74.0
other70.763.470.466.169.568.159.296.1
LoFG2 (coarse facade semantics)
OA70.266.567.771.165.263.259.594.0
μP67.263.962.768.960.559.158.468.3
μR66.260.760.768.858.458.759.257.1
μF166.261.961.368.558.358.357.659.1
μIoU52.647.648.354.146.245.943.251.6
floor93.369.770.390.694.094.987.973.0
decoration45.144.323.450.530.425.438.511.6
structural73.160.745.873.566.163.864.519.6
opening44.387.890.452.425.335.235.393.9
other75.471.776.575.375.672.361.997.5

BibTeX

@InProceedings{UnderOneFacade2026,
  author    = {Wang, Yi and Wang, Fan and Gyawali, Prabin and Xu, Ziyang and Klimkowska, Anna
               and Jing, Yixiong and Yang, Wanru and Biljecki, Filip and Holst, Christoph
               and Busam, Benjamin and Sheil, Brian and Wysocki, Olaf},
  title     = {UnderOneFacade: Worldwide Facade Semantic Segmentation Benchmark Dataset},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}