UnderOneFacade: Worldwide Facade Semantic Segmentation Benchmark Dataset¶

Conference: ECCV 2026
Paper: ECCV / Project Page
Area: 3D Vision / Segmentation
Keywords: facade semantic segmentation / point cloud benchmark dataset / cross-domain generalization / long-tail distribution / hierarchical taxonomy

TL;DR¶

UnderOneFacade collects centimeter-accurate facade point clouds of roughly 366 buildings in the UK (Nottingham), Germany (Munich), and Singapore using static terrestrial, backpack, and vehicle-mounted laser scanning, totaling 2.7B annotated points; it extends ZAHA's hierarchical Level-of-Facade-Granularity (LoFG) taxonomy (15 classes at LoFG3 / 5 at LoFG2) across countries with about 5,500 hours of manual annotation, and benchmarks six representative point cloud networks under a unified xyz-only, 100-epoch protocol, showing that the best model reaches only about 33 mIoU on fine-grained LoFG3 and that zero-shot cross-continental transfer roughly halves it again.

Background & Motivation¶

3D urban semantic segmentation has advanced quickly over the past decade on the back of LiDAR sensing and large-scale reconstruction, with Semantic3D, Paris-Lille-3D, Toronto-3D, and KITTI-360 turning urban scene understanding into a track with standard test sets. But their label sets stop at the level of buildings, roads, and vegetation; facade-level components (windows, doors, balconies, moldings) are either absent or collapsed into a coarse "building" class. Downstream tasks that actually need them — semantic digital twins, energy modeling, facade change detection — are precisely the ones that require per-element labels. Facades are also intrinsically hard: their class distribution is severely long-tailed, with dominant classes such as walls occupying most points while semantically critical elements (doors, windows, balconies) are rare, thin, and repetitive, so models drift toward the dominant classes and miss small structural elements even within a single architectural style.

Existing facade datasets are trapped in a triangle of small scale, single-city coverage, and dataset-specific label definitions: ArCH, TUM-FAÇADE, ZAHA, Tuscany, TrueCity, and City-Facade each define their own class list, range from under 0.1B to 0.6B facade-labeled points, and almost all come from one city and one sensing platform. Why has no large-scale facade benchmark appeared? First, centimeter-level geometry is only obtainable from laser scanning — facade parsing is accuracy-critical, and window frames or moldings only a few centimeters wide simply smear out in street-level image reconstruction. Second, facade annotation is extremely expensive: the elements are small, repetitive, and structurally intricate, and labeling a single facade takes many hours. ZAHA (WACV 2025) took an important step with the Level of Facade Generalization (LoFG), anchoring class definitions to architectural standards such as CityGML, IFC, and the Art and Architecture Thesaurus so that "window" means the same thing across datasets. But its realization covers a single German city and therefore validates the concept mainly on Central European typologies; cross-domain generalization itself has never actually been measured.

This paper's position is blunt: to ask whether facade models transfer, the dataset itself has to span domains, and the class definitions must be harmonized before that. It therefore inherits ZAHA's LoFG taxonomy and German subset wholesale, and acquires UK and Singapore data under the same class definitions, so that geometric precision, semantic consistency, and geographic diversity hold simultaneously, backed by a cross-country, cross-continental, multi-granularity evaluation protocol. Cross-country acquisition and unified semantics reinforce each other: the taxonomy only works across countries because it is grounded in architectural standards, and only cross-country data can reveal the classes on which it breaks down. Core idea: combine one architecture-standard-grounded hierarchical taxonomy with centimeter-accurate multi-sensor point clouds from three countries to turn facade segmentation from single-city fitting into a measurable cross-domain generalization benchmark.

Method¶

Overall Architecture¶

The input to UnderOneFacade is urban blocks in three countries — The Park residential district in Nottingham (UK), downtown Munich (Germany), and several major street corridors in Singapore — and the output is a facade point cloud benchmark with hierarchical semantic labels (2.7B points, ~366 facades, LoFG3 / LoFG2 / LoFG1 labels) plus a fixed training and evaluation protocol. Data production is one serial pipeline: multi-sensor laser scanning → global georeferencing and registration → per-building facade extraction using OSM building footprints (expanded by a 5 m buffer) → manual per-point annotation under the LoFG taxonomy → annotator-consensus cross-checking → per-building 70/20/10 spatially disjoint splits → benchmarking six point cloud segmentation networks under a unified protocol. The design goal is not to collect more points but to let three kinds of domain shift — acquisition setup, regional style, and class frequency — be discussed separately under one label set and one protocol.

On the evaluation side, the benchmark runs along three axes: cross-country / cross-continental domain shift, hierarchical semantic granularity (LoFG2 versus LoFG3), and realistic long-tailed class distributions. Aggregating countries produces a markedly heavier-tailed distribution than any single-country dataset (the European training set alone spends 58.3% of its points on wall and other, whereas the Singapore subset covers facade categories more evenly) — which is exactly what the benchmark intends to measure.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Cross-continental multi-sensor acquisition<br/>TLS / backpack / vehicle, cm-accurate"] --> B["LoFG hierarchical taxonomy<br/>15 / 5 / 1 classes"]
    B --> C["Manual annotation & consensus check<br/>~5,500 hours, ~10% corrected"]
    C --> D["Per-building splits<br/>70 / 20 / 10, spatially disjoint"]

Key Designs¶

1. Cross-continental multi-sensor acquisition: trade centimeter geometry for transferable facade detail

The UK subset was captured with a Leica RTC360 static terrestrial laser scanner (TLS), with 17 stations placed along the urban segment in The Park, Nottingham. Static multi-station acquisition substantially raises point density on vertical structures and reduces occlusion compared with mobile mapping, at the cost of a slow campaign; the scans were first registered with the scanner's visual–inertial system and then refined through global registration, while global georeferencing came from a ground control point (GCP) network measured with GNSS and RTK corrections, placing the clouds at centimeter-level accuracy in the national coordinate system. The German subset is inherited from ZAHA, whose upstream data is TUM-MLS-2016 in downtown Munich: the MODISSA vehicle platform carries two Velodyne HDL-64E LiDARs at the front and two VLP-16 sensors at the rear, georeferenced by an inertial navigation system with SAPOS RTK corrections. The Singapore subset used a Leica BLK360 mounted on a backpack and on street-level positions along major corridors, registered by geometric feature alignment and then referenced to geographic coordinates using OSM data. All three subsets then follow the same clipping procedure: buildings are extracted from the registered clouds using OSM footprints expanded by a 5 m buffer to absorb footprint inaccuracy, and deliberately no noise filtering is applied, preserving raw acquisition fidelity. The result is 997M points (~100 facades) for the UK, 601M (~66 facades) for Germany, and 1.117B (~200 facades) for Singapore.

Using three sensing platforms is a deliberate trade-off. Static TLS gives the best density and geometric quality but scales poorly; the vehicle platform covers ground quickly but samples vertical surfaces sparsely; the backpack system sits in between. Stacking them means cross-domain evaluation contains both architectural style differences and acquisition-mode differences, which brings the benchmark closer to real city-scale deployment — and also creates the disentanglement problem the authors acknowledge (see Limitations). The European pair is chosen deliberately: the UK and Germany share historical influences yet differ in facade materials, ornamentation, and structural composition (Victorian versus Haussmann), forming a within-continent style contrast, while Singapore represents Southeast Asian urban architecture — traditional Malay and Chinese elements mixed with modern high-density construction — creating the strongest stylistic gap against Europe.

2. LoFG hierarchical taxonomy: anchor class definitions to architectural standards rather than to one dataset

The annotation scheme inherits ZAHA's Level of Facade Generalization (LoFG) and adapts it across countries; the point is to hang label semantics on international urban modeling and architectural standards (CityGML, IFC, the Art and Architecture Thesaurus) instead of rewriting a class list for every newly acquired city. LoFG3 provides 15 fine-grained classes: wall, balcony, molding, deco, stairs, column, arch, window, blinds, door, roof, ground, terrain, interior, and other (mainly noise). LoFG2 folds them into five functional groups — structural (wall / column / arch / balcony / stairs), opening (blinds / door / window), decoration (molding / deco), floor (ground surface / terrain), and other elements (roof / other / interior) — while LoFG1 abstracts the entire facade into a single class.

The hierarchy is more than tidier labels: it doubles as a switchable evaluation setting. The gap a model shows between LoFG2 and LoFG3 directly exposes what aggregation hides — in the experiments the LoFG2 floor class reaches 93.3 F1 while the LoFG3 stairs class, once separated out again, drops to 6.5 (PointNet++), showing that merging fine classes makes numbers look better without changing the fact that the model never learned these components. This is also the main extension over ZAHA: the taxonomy is inherited as-is, but scale grows from 0.6B to 2.7B points (4.5×) and the same class list is validated in three countries for the first time.

3. Manual annotation and consensus checking: resolve class ambiguity at the annotation stage

All annotations were produced manually by trained annotators with expertise in architecture and geospatial modeling, working in the professional large-scale point cloud software Cyclone 3DR. A single facade took roughly 15 hours, giving about 5,500 hours in total for the full set of ~366 facades. Annotations were then harmonized and cross-checked through annotator consensus, with a correction rate of about 10% concentrated in genuinely ambiguous categories (typically the molding versus decoration boundary); the decision rules live in the annotation guidelines of the supplementary material (Suppl. Sec. 9).

This step deserves separate treatment because what it handles is a definition problem rather than a geometric one: moldings and decorative elements are often geometrically continuous, with no clean separating surface, so a model cannot learn from point clouds how the building standards draw the line — it can only rely on the annotation definition itself being stable. The ~10% correction rate is the price of pinning that definition down, and it is also a useful signal: it marks the least stable corner of the taxonomy, letting future work that reuses these labels anticipate where the label noise concentrates.

4. Per-building splits: let the test set evaluate unseen buildings, not unseen points

Each country subset is split 70/20/10 into training, validation, and test by total point count, with buildings rather than point blocks as the unit of splitting, so that a complete facade belongs to exactly one subset and the three subsets are spatially disjoint; dominant and underrepresented classes are deliberately preserved across all subsets. The reason for splitting per building is concrete: points of one building are highly correlated — same material, the same repeated windows, the same molding profile — so random point-level splitting would leak points of the same building into both training and test and inflate the scores. The paper also reports statistics that help attribute domain shift: mean facade heights of 8.2 / 10.6 / 12.1 m (Singapore / UK / Germany), and mean point densities of 11,483 / 28,713 / 33,199 pts/m² before voxelization, dropping to 575 / 427 / 2,446 pts/m² after.

Loss & Training¶

The baselines cover the three major working paradigms of point cloud networks: point-based (PointNet++, KPConv), graph-based (DGCNN), and transformer-based (OctFormer, Point Transformer v1, Point Transformer v3). They were chosen because they represent different feature aggregation mechanisms and serve as widely used reference points across point cloud benchmarks. The primary benchmark uses xyz coordinates only as input, with radiometric information analyzed separately, in order to separate geometric difficulty from appearance difficulty so that the main leaderboard reflects segmentation ability under pure geometry.

The training protocol is deliberately uniform: a fixed 100 epochs, no per-method hyper-parameter tuning, identical data preprocessing and uniform point sampling for all networks, and a unified 5 cm training voxelization that suppresses point density differences between country subsets. The loss is unweighted cross-entropy — the authors state this is intentional: without class-frequency weighting, the pressure of the long-tailed distribution is applied to every method as-is rather than being masked by reweighting.

Metrics are the standard segmentation set: Overall Accuracy (OA), Precision (P), Recall (R), F1, and IoU, reported at both LoFG2 and LoFG3 granularity; mIoU averages the per-class IoU over all classes at that granularity (15 classes at LoFG3, 5 at LoFG2). Because IoU is computed per class before averaging, high scores on dominant classes cannot offset zeros on rare ones — which is exactly why it is needed instead of OA: KPConv reaches 83.3 OA in the UK but only 23.4 mIoU.

Cross-continental generalization uses a zero-shot transfer protocol: a model trains on one continent (Europe = UK + German subset, Asia = Singapore subset) and is evaluated directly on the other without any target-domain adaptation, with results reported at both LoFG2 and LoFG3. The two directions are deliberately matched in training volume (about 1.5B points for Europe versus 1.2B for Asia) so that observed differences reflect distribution rather than dataset size.

Key Experimental Results¶

Main Results¶

Table 1 is the main leaderboard at LoFG3, the finest granularity. The Overall column reports models trained on the full UnderOneFacade dataset, while the per-country columns report models trained within a single country subset.

Family	Method	UK OA / mIoU	Singapore OA / mIoU	Germany OA / mIoU	Overall OA / mIoU
Point	PointNet++	68.5 / 22.1	59.6 / 29.8	66.4 / 25.6	60.9 / 29.7
Point	KPConv	83.3 / 23.4	41.5 / 11.2	55.8 / 11.2	49.0 / 15.0
Graph	DGCNN	77.8 / 28.3	62.2 / 31.3	71.1 / 33.4	62.7 / 33.4
Transformer	Point Transformer (PTv1)	79.0 / 31.0	44.7 / 19.8	75.0 / 41.6	48.8 / 15.1
Transformer	PTv3	79.0 / 26.5	58.5 / 29.6	56.8 / 15.7	58.3 / 27.3
Transformer	OctFormer	75.5 / 27.7	61.5 / 34.6	44.9 / 13.5	57.9 / 27.3

Table 2 reports zero-shot cross-continental transfer (xyz input only); "Europe → Asia" trains on UK + the German subset and tests on Singapore, and vice versa.

Method	Europe → Asia LoFG2 OA / mIoU	Europe → Asia LoFG3 OA / mIoU	Asia → Europe LoFG2 OA / mIoU	Asia → Europe LoFG3 OA / mIoU
PointNet++	56.4 / 33.9	40.5 / 14.0	65.8 / 38.8	51.9 / 23.6
KPConv	48.6 / 25.7	32.1 / 6.6	48.9 / 28.8	33.9 / 9.4
DGCNN	56.4 / 38.3	42.6 / 14.8	63.9 / 40.5	44.8 / 17.4
OctFormer	59.2 / 37.0	39.0 / 11.5	62.3 / 40.0	50.5 / 16.2
PTv1	55.1 / 34.2	33.8 / 14.0	58.1 / 38.1	41.9 / 20.4
PTv3	60.9 / 40.8	41.5 / 11.0	53.4 / 34.2	49.0 / 15.7

Ablation Study¶

Table 3 ablates radiometric features on the fixed UK LoFG3 test set, replacing pure xyz input with xyz + projected RGB + laser intensity (RGBI); ∆ is the change relative to xyz.

Method	Setting	OA	µIoU	window F1	door F1	molding F1	interior F1
PointNet++	xyz	68.5	22.1	13.7	5.7	23.6	23.2
PointNet++	+rgbi	73.9	29.4	48.8	30.2	41.5	64.4
PointNet++	∆	+5.4	+7.3	+35.1	+24.5	+17.9	+41.2
PTv1	xyz	79.0	31.0	23.1	21.7	46.3	55.4
PTv1	+rgbi	83.9	39.1	54.5	44.8	60.9	74.8
PTv1	∆	+4.9	+8.1	+31.4	+23.1	+14.6	+19.4
DGCNN	xyz	77.8	28.3	27.9	10.3	40.7	49.7
DGCNN	+rgbi	76.5	30.2	38.2	43.0	47.1	39.8
DGCNN	∆	−1.3	+1.9	+10.3	+32.7	+6.4	−9.9

Table 4 breaks the cross-country drift down to the class level (F1 difference ∆ = SG − UK on UK / SG, showing four representative methods; the full six-method result is Table 4 of the paper).

Class	PointNet++	DGCNN	PTv1	OctFormer
wall (LoFG3)	−18.1	−21.5	−39.9	−16.7
terrain (LoFG3)	−33.0	−34.2	−39.7	−25.9
window (LoFG3)	−10.6	−25.3	−17.9	−10.5
molding (LoFG3)	+36.6	+22.6	−23.4	+27.3
column (LoFG3)	+56.8	+48.8	+11.8	+65.4
arch (LoFG3)	+62.0	+67.3	+57.6	+80.0
opening (LoFG2)	−71.1	−77.4	−90.9	−60.3
structural (LoFG2)	+40.2	+25.7	+10.3	+31.5

Key Findings¶

Fine-grained facade segmentation is far from solved. On the hardest LoFG3 setting the strongest model trained on all data (DGCNN) reaches only 33.4 mIoU; most methods show acceptable OA (57–63) but mIoU stuck between 15 and 33, meaning the score comes mostly from dominant classes such as wall, terrain, and roof.
The graph network overtakes the transformers. DGCNN leads the Overall column at 33.4 mIoU, clearly ahead of PTv1 (15.1) and PTv3 (27.3) — one of the paper's most counter-intuitive results. The authors argue that current transformer designs for point clouds do not yet capture the thin, repetitive, locally geometric structure that dominates facades.
Single-country training is deceptive. PTv1 trained within Germany reaches 41.6 mIoU, but trained on the three-country mixture it drops to 15.1 Overall: single-country numbers reflect overfitting to one local style rather than transferable segmentation ability.
Hierarchical aggregation does not solve the long tail. Under LoFG2 the floor class reaches 93.3 F1 (PointNet++), while the same methods score only 0.7–9.1 on LoFG3 stairs, 9.2–10.0 on deco, and at most 26.7 on door. Merging classes hides the problem rather than solving it; opening also shows huge variance across architectures (DGCNN 92.1 versus PTv3 16.4), showing that coarse granularity retains strong architectural sensitivity.
Cross-continental transfer is clearly asymmetric. At LoFG3, Asia → Europe (51.9 OA / 23.6 mIoU) beats Europe → Asia (42.6 / 14.8) across the board. The paper's data-grounded explanation is source-domain class balance: wall and other alone account for 58.3% of European training points, whereas the Singapore subset covers facade categories more evenly, so Asian data provides stronger supervision for rare classes. Once volumes are matched, what drives transfer is source-domain semantic balance, not scale — the most transferable finding in the paper.
Cross-country drift is highly non-uniform across classes. wall and terrain lose 17–40 F1 in Singapore and window loses 10–25, yet molding, column, and arch actually gain there (arch is 0.0 F1 for nearly every method in the UK but 57.6–80.0 in Singapore). Cross-domain degradation therefore mixes two things: shapes the model genuinely fails to learn, and classes that simply have no samples in the source domain. LoFG2's opening degrades the most (PTv1 97.7 → 6.8), which the authors attribute to the far more intricate openings of Southeast Asian buildings compared with the repetitive windows of UK Victorian architecture.
Radiometric features are a double-edged sword. Adding RGBI lifts window and door F1 by more than 30 (PointNet++ window +35.1, door +24.5), showing that color is a strong cue for geometrically weak components; yet DGCNN loses 1.3 OA and 9.9 interior F1, which the authors read as models learning dataset-specific appearance shortcuts rather than robust representations. Large planar structures are barely affected (terrain changes by −6.3 to +0.2).

Highlights & Insights¶

Treating the taxonomy itself as a methodological contribution. LoFG is not an annotation convention but an evaluation switch: the gap a model shows between LoFG2 and LoFG3 localizes exactly what aggregation hides. The recipe transfers to any long-tailed segmentation benchmark — draw the hierarchy first, then decide which level to report.
Cross-country acquisition turns "domain shift" from rhetoric into measurement. By matching training volume (1.5B European versus 1.2B Asian points) and running both transfer directions, the paper can conclude that source-domain class balance bounds transfer, instead of vaguely asserting that cross-domain is hard.
Deliberately making the setting harder. Unweighted cross-entropy plus xyz-only input refuses every compensation that would make numbers look good, exposing long-tail and geometric difficulty directly to the leaderboard. A self-imposed-hardening protocol worth borrowing elsewhere.
Uniform 5 cm voxelization as cheap cross-sensor density de-identification. Point densities differ threefold across countries (11,483 versus 33,199 pts/m²); unified training voxelization removes this term so mIoU differences are more likely to come from architecture and acquisition geometry than from sampling rate, at minimal loss of usable signal.
"Class balance before data volume" is directly actionable for acquisition planning. Since the Asian subset transfers better with fewer points, adding coverage of rare classes across cities beats acquiring more square meters in the same city — a conclusion relevant to any long-tailed 3D dataset.
Per-class signed gains expose a dataset's stylistic fingerprint. arch scores 0 for every method in the UK and jumps to tens of F1 in Singapore; this extreme asymmetry is by itself a characterization of how the two countries use building components, and can serve as ready-made material for studying class-existence shift in domain adaptation.

Limitations & Future Work¶

The main limitation the authors admit is the entanglement of sensor and style: the dataset mixes vehicle-mounted mobile scanning, a backpack system, and static TLS, so cross-country and cross-continental shift cannot be attributed to a single factor — sensor type, point density, acquisition geometry, and architectural style all vary together. They provide partial disentanglement (uniform 5 cm training voxelization to reduce density differences, plus the radiometric analysis of Table 5 to isolate appearance shift) but acknowledge that controlled acquisitions of the same environment would be needed to fully separate sensor effects.
Preserving the naturally long-tailed distribution means some rare classes have very few training samples (stairs and arch score 0.0 F1 for nearly every method in the UK). This reflects realistic urban scenes but implies high variance for those classes and calls for caution when interpreting their results.
Coverage is limited to three cities (Nottingham, Munich, Singapore); despite their stylistic spread they cannot represent global facade variability. The authors stress that the LoFG taxonomy and the annotation protocol are designed to be extensible, so additional cities and sensing setups can be integrated later.
My own observations: only six baselines and one unified training budget (100 epochs, no tuning) are reported, with no recent stronger point cloud backbones (e.g., PointNeXt or superpoint-based methods) and no domain-adaptation or fine-tuning upper bound, so "33 mIoU" reads more like a lower bound for off-the-shelf methods than a ceiling; the ~10% correction rate indicates that disputed points were changed but gives no harder reliability measure such as inter-annotator agreement (e.g., Cohen's kappa); per-country test split sizes and class distributions are also not broken down.
Possible improvements: release the test set with a public leaderboard to fix the evaluation protocol; add domain-adaptation / fine-tuning baselines to quantify how much of the cross-domain gap is removable; extend the radiometric analysis from the UK only to a cross-country comparison; and run rare-class experiments under few-shot or class-weighted losses to separate insufficient data from insufficient model capacity.

vs ZAHA (WACV 2025): This paper inherits its LoFG hierarchical taxonomy and the German subset, but scales from 0.6B to 2.7B points (4.5×), adds multi-sensor acquisition, and introduces a cross-country, cross-continental evaluation protocol. ZAHA validates the concept on a single German city and Central European typologies; this work validates the same class definitions across three continents. The taxonomy design credit belongs to ZAHA — the increment here is scale, geographic diversity, and the evaluation protocol.
vs TUM-FAÇADE / ArCH (0.1B-scale, one city or one heritage site, 10–17 classes): they demonstrated that facade-level fine-class segmentation works in 3D, but at a scale too small to train deep models or to run systematic cross-domain evaluation; this work raises facade annotation by an order of magnitude and supplies a unified protocol.
vs City-Facade (2026, ~0.2B, 9 classes, single country) and TrueCity (2025, ~0.1B, mixed real and synthetic): both are recent facade datasets, but both stay within one country and lack hierarchical classes; the difference here is cross-country label consistency plus a zero-shot cross-continental protocol.
vs general urban point cloud datasets (Semantic3D, Paris-Lille-3D, Toronto-3D, KITTI-360, Paris-CARLA-3D): they are the main arena for urban semantic segmentation but label at the granularity of buildings, roads, and vegetation. This work pushes granularity down to facade components, trading much higher acquisition and annotation cost for labels that can directly support semantic digital twins and LoD3 reconstruction.
Takeaway: for 3D domain adaptation research, this paper offers a ready-made testbed combining class-existence shift, long tails, and multiple sensors; for work on segmentation losses, the per-class F1 distribution under unweighted cross-entropy is a ready-made control for measuring how much reweighting or resampling actually buys.

Rating¶

Novelty: ⭐⭐⭐⭐ The hierarchical taxonomy is inherited from ZAHA; the real contribution is cross-continental scale, multi-sensor acquisition, and the first systematic cross-country/cross-continental generalization study — methodologically incremental but the engineering and protocol design are solid.
Experimental Thoroughness: ⭐⭐⭐⭐ Six baselines across three paradigms, two semantic granularities, two transfer directions, and a radiometric ablation, with consistent conclusions and data-grounded explanations; however it lacks a domain-adaptation/fine-tuning ceiling and reliability measures such as inter-annotator agreement.
Writing Quality: ⭐⭐⭐⭐ Motivation and design trade-offs are clearly laid out, numbers in the text agree with the tables, and the authors are candid about the sensor entanglement and long-tail limitations.
Value: ⭐⭐⭐⭐⭐ 2.7B annotated points, released under CC BY 4.0 with evaluation scripts and pretrained models — a rare directly usable cross-domain benchmark for facade parsing and long-tailed 3D segmentation that will be heavily cited soon.