Large-Scale 3D Point Cloud Semantic Segmentation Benchmarks

Last updated: 2026-05-24

Why It Matters

An aggregated-LiDAR-map semantic segmentation pipeline is built, pre-trained, and validated against the public 3D point cloud segmentation benchmarks — no airside-domain benchmark exists (see ../overview/aggregated-map-semantic-segmentation.md §5.4). This page is the evaluation-mechanics reference for that benchmark landscape: splits, metrics, submission servers, label formats, licensing, and known issues. The companion §5 of the aggregated-map page is the selection guide (which dataset for what); this page is the usage reference (how to actually train and evaluate on them).

The practical point: these datasets shape what every 3D segmentation method reports. A "SOTA on SemanticKITTI" claim means little without knowing the split, the task variant, whether the number came from the hidden test server or a local validation set, and whether the method used multi-dataset pre-training.

Benchmark Landscape

Benchmark	Modality	Task variants	Eval classes	Metric	Leaderboard
SemanticKITTI	MLS (HDL-64E)	Single-scan; multi-scan (accumulated)	19 / 25	mIoU	Hidden test server (CodaLab)
SemanticTHAB	MLS (Ouster OS2-128)	Per-cloud semantic segmentation	20	mIoU / OA	Public Zenodo + repo; no hidden server
nuScenes-lidarseg	MLS (HDL-32E)	Single-frame lidarseg; panoptic	16	mIoU	Hidden test server (EvalAI, 3/yr cap)
Waymo 3D Seg	Multi-LiDAR vehicle	Per-frame semantic	22	mIoU	waymo.com/open
Semantic3D	TLS (static)	semantic-8; reduced-8	8	mIoU / OA	Online server (no annual cap)
Paris-Lille-3D (NPM3D)	MLS	Coarse segmentation	10	Av.IoU	Online test server
Toronto-3D	MLS	Segmentation	8	mIoU / OA	Public split (L002 test)
KITTI-360	MLS (accumulated)	3D segmentation; 2D-3D	~19	mIoU	Online server (cvlibs.net)
DALES	ALS (airborne)	Segmentation	8	mIoU / OA	Public split
GridNet-HD	UAV LiDAR + oblique RGB imagery	3D semantic; image segmentation; LiDAR-image fusion	11 evaluated groups	mIoU	Hugging Face hidden-label leaderboard
SensatUrban	UAV photogrammetry	Segmentation	13	mIoU / OA	Hidden test server (CodaLab)
STPLS3D	Aerial photogrammetry + synthetic	Segmentation; instance	up to 18	mIoU / AP	Public + challenge
CUS3D	UAV photogrammetry point cloud + mesh + imagery	3D/mesh/2D semantic segmentation	10	mIoU / OA / mAcc	Paper benchmark
SUM / SUM Parts	Textured urban mesh	Mesh semantic; part-level segmentation	6 / 21	mIoU	CVPR 2025 project/code/data
S3DIS	RGB-D (indoor)	Semantic segmentation	13	mIoU / OA	Area-5 or 6-fold (no server)
ScanNet v2 / ScanNet200	RGB-D mesh (indoor)	Semantic; instance	20 / 200	mIoU / mAP	Hidden test server
Point Cloud City / Open3D-ML PCC	Indoor public-safety point clouds	Semantic segmentation; label unification	collection-specific	IoU / mIoU	NIST program + Open3D-ML repo; access varies
FRACTAL	ALS (airborne)	Segmentation	7	mIoU / OA	Public split (HuggingFace)
GOOSE / GOOSE-Ex	LiDAR + RGB + NIR (off-road)	Segmentation (64 fine / 8 super)	64	mIoU	Public split; ICRA 2025 challenge
SemanticRail3D	MLS (railway corridor)	Semantic; instance	~11–12	mIoU	Public split
WHU-Railway3D	MLS (railway, multi-env)	Semantic	11	mIoU / OA	Public split
WHU-Urban3D	MLS + ALS	Semantic; instance; object detection	18 MLS / 7 ALS semantic classes	mIoU / OA / mAcc + instance metrics	Benchmark page; upload currently unavailable
Turin3D	ALS (airborne)	Semi-supervised segmentation	~8	mIoU	Public (val/test only)
CITYLID	ALS (airborne)	Fine street-feature segmentation	13	qualitative only	HuggingFace open dataset
ECLAIR	ALS (aerial)	Semantic segmentation	11	mIoU	CVPRW paper + GitHub/download form
YUTO Semantic	ALS (aerial campus)	Semantic segmentation	9	OA / mIoU	Hugging Face; one mission released
S.MID	Livox Mid-360 hybrid-solid LiDAR	Single-frame semantic segmentation	14 eval / 25 annotated	mIoU	Official dataset page + SFPNet repo
OpenTrench3D	Photogrammetric RGB point cloud	Semantic segmentation	5	mIoU	Public repo / Kaggle; not LiDAR
MLDAS	Multi-LiDAR vehicle platform	Domain-adaptation semantic segmentation	14	mIoU	Email-request access; license-limited
USCILab3D	32-beam LiDAR + 5 cameras	Long-term campus semantic point clouds	267	No stable public leaderboard	Raw bags/code available; processed data pending
Industrial3D	TLS industrial MEP	Semantic + cross-paradigm benchmark	12	mIoU	Watchlist; full dataset/code pending paper acceptance
City-Facade	MLS building-facade point clouds	Semantic + instance segmentation	9 facade classes	OA / mIoU	ISPRS JPRS 2026 + official repo; full access to verify
ZAHA	MLS facade point clouds (TUM-MLS-2016)	Facade semantic segmentation	5 LoFG2 / 15 LoFG3	OA / IoU / F1	WACV 2025 + TUM2TWIN benchmark + official repo
SIP (Site in Pieces)	TLS (single-station)	Construction-site segmentation	23	mIoU	Public split
Waymo-4DSeg / SAM4D	Camera + LiDAR (pseudo-labeled)	Class-agnostic masklets	class-agnostic	mask metrics	Derived (Waymo Open)

Dataset Selection Protocol for Aggregated-Map Segmentation

Do not rank datasets by leaderboard prestige alone. For an aggregated-map semantic layer, the first question is whether the public benchmark exercises the same map-production failure mode as the target product. A SemanticKITTI or Waymo score is useful for single-scan LiDAR semantics; it is not sufficient evidence that a model will preserve taxiway markings, terminal frontage details, utility poles, construction equipment, or stationary-but-transient clutter in a registered multi-session map.

Use the following protocol before choosing a pre-training or validation pool:

1. Match acquisition geometry first. Vehicle MLS and accumulated mobile-mapping datasets are the highest-value sources for ground-robot maps; ALS/UAV datasets are site-survey complements; TLS/indoor datasets are managed-building or terminal-interior proxies; mesh/photogrammetry datasets are digital-twin transfer aids rather than LiDAR sensor-noise evidence.
2. Match the class pressure second. Choose datasets that contain the failure-critical classes: road markings, kerbs, poles, wires, signs, facade openings, utilities, workers, temporary equipment, rail/corridor furniture, or industrial MEP. A generic "outdoor" class set is not enough.
3. Separate training evidence from release evidence. Photogrammetric and mesh corpora are acceptable for pre-training features or taxonomy stress, but release claims for a LiDAR map need LiDAR validation data with comparable density, intensity behavior, registration quality, and sensor pose.
4. Require split discipline. Favor spatially disjoint geographic splits over random point splits. Aggregated maps leak context easily: adjacent tiles can share the same facade, road marking, or pole-line pattern even when point IDs differ.
5. Audit license and access before planning. Several useful 2025-2026 additions are email-gated, challenge-gated, non-commercial, or still pending full release. Treat them as research proxies until download, redistribution, and commercial-use terms are verified.
6. Score map hygiene separately. Datasets built for semantic segmentation rarely label dynamic residuals, static-but-transient objects, or false deletions. Pair this page with moving/static and map-cleaning benchmarks whenever the target product is a publishable map, not a per-frame perception output.

Release question	Primary dataset evidence	Secondary evidence	What still needs in-house validation
Can a model label a ground-level registered LiDAR map?	KITTI-360, Paris-Lille-3D, Toronto-3D, WHU-Urban3D, SemanticRail3D, WHU-Railway3D	SemanticKITTI multi-scan, SemanticTHAB, MLDAS	Site-specific registration drift, map tiling, intensity calibration, and static/transient quarantine
Can it preserve non-road managed-site classes?	Point Cloud City / Open3D-ML PCC, SIP, S.MID, Industrial3D, USCILab3D	S3DIS, ScanNet200, CUS3D	Outdoor transfer, vehicle-mounted density, safety-critical minority classes, operational clutter
Can it use LiDAR plus image evidence without requiring camera at release time?	GridNet-HD, KITTI-360, CUS3D, H3D	2DPASS/ScaLR-style distillation papers, SAM4D pseudo-labels	Calibration residuals, projection provenance, image-unavailable fallback, LiDAR-only artifact compatibility
Can it handle facade and terminal-frontage semantics?	ZAHA, City-Facade, SUM Parts, CUS3D, WHU-Urban3D	H3D, Toronto-3D	Full-site ground classes, facade-to-ground boundary policy, BIM/digital-twin handoff rules
Can it handle overhead/long-thin infrastructure?	GridNet-HD, ECLAIR, WHU-Railway3D, SemanticRail3D, DALES	OpenTrench3D for utility taxonomy pressure	Wire/pole recall at target sensor density, false deletion during map cleaning, camera projection visibility
Can it support airside, port, depot, or industrial-yard rollout?	GOOSE-Ex, SIP, S.MID, Industrial3D, WHU-Urban3D, KITTI-360	STPLS3D, CUS3D, SemanticTHAB	Airport/yard-specific objects, GSE/vehicle vocabulary, FOD exclusion, stationary people, staged equipment
Can it produce a releaseable semantic-map artifact?	No public dataset is sufficient by itself	MapBench, SceneEdited, HKCD, moving/static datasets, map-hygiene protocols	A held-out, manually reviewed site map with semantic labels, removal labels, provenance, and publication gates

Non-Road Urban-District Benchmark Bundles

For managed districts, use public datasets as a bundle, not as a single leaderboard proxy. The bundle should cover acquisition geometry, class pressure, long-tail objects, and release-state risk. No public bundle replaces a held-out local map, but the combinations below expose failures that a road-only benchmark will miss.

Target district slice	Minimum public proxy bundle	Classes to stress	Release-state risks to add locally
Airport apron, depot, or service-road network	KITTI-360, Toronto-3D, Paris-Lille-3D, WHU-Urban3D, SemanticRail3D or WHU-Railway3D, GridNet-HD for thin assets	pavement/markings, kerbs, poles, signs, fences, gantries, long linear boundaries	parked aircraft/GSE, stationary people, FOD candidates, wet-apron artifacts, work-zone equipment
Campus or pedestrian district	YUTO Semantic, USCILab3D, SemanticTHAB, MLDAS, Point Cloud City / Open3D-ML PCC	walkways, plazas, buildings, vegetation, furniture, poles, bicycles/carts, delivery robots	event furniture, parked bikes/carts, stationary people, temporary barriers, open-vocabulary campus objects
Port, logistics yard, or industrial depot	SIP, S.MID, Industrial3D, GOOSE-Ex, OpenTrench3D, WHU-Urban3D	containers, trailers, cranes/gantries, pipes/ducts, cabinets, trenches, gravel/asphalt transitions	moved containers, staged pallets, tools/cables, construction clutter, dust/rain artifacts
Utility corridor or overhead-infrastructure site	GridNet-HD, ECLAIR, DALES, SemanticRail3D, WHU-Railway3D, Toronto-3D	pylons, poles, overhead lines, structural cables, insulators, vegetation, water/soil/road	false deletion of wires/poles, vegetation encroachment freshness, camera-projection gaps, artifact-vs-wire confusion
Terminal frontage, facade, or managed-building exterior	City-Facade, ZAHA, SUM Parts, CUS3D, Point Cloud City / Open3D-ML PCC	wall/window/door/sign/awning/balcony, facade-to-ground boundary, roof/overhang parts	advertisements, scaffolds, temporary signs, construction wraps, BIM-vs-current-map differences
Terminal interior or public-safety facility	Point Cloud City / Open3D-ML PCC, S3DIS, ScanNet200, USCILab3D	doors, corridors, structural columns, furniture, equipment, clutter, tail classes	movable furniture, people, temporary queues, maintenance tools, access-control changes

The local benchmark should add a release-state overlay for every bundle: permanent_static, dynamic_residual, movable_static, static_transient, fod_candidate, artifact, and unknown_review. Public mIoU answers whether the semantic class is plausible; the overlay answers whether the point is safe to publish or use as training data.

Per-Benchmark Evaluation Detail

SemanticKITTI — the AV-domain reference

• Sensor: Velodyne HDL-64E, 64-beam rotating LiDAR, 360° FOV, ~0.09° angular resolution. Vehicle-mounted MLS survey drives around Karlsruhe, Germany.
• Scale: 23 sequences, 43,000+ scans, ~4,549 million annotated points.
• Split: Train sequences 00–07 + 09–10 (23,201 scans); validation sequence 08 (~4,071 scans); test sequences 11–21 (~20,351 scans, labels withheld). Sequence 08 is the de-facto development reference; report it during architecture search.
• Single-scan task — 19 evaluated classes (raw ontology has 28; "other-structure," "other-object," and "outlier" excluded due to high intra-class variance): road, sidewalk, parking, other-ground, building, car, truck, bicycle, motorcycle, other-vehicle, vegetation, trunk, terrain, person, bicyclist, motorcyclist, fence, pole, traffic-sign.
• Multi-scan / moving-object task — 25 classes: adds moving variants (moving-car, moving-truck, moving-other-vehicle, moving-person, moving-bicyclist, moving-motorcyclist) — stationary and in-motion instances are distinct classes. This is the variant closest to aggregated-map segmentation.
• Scene completion task: same 19-class space over voxelised occupancy grids.
• Label file format: Binary .label files, one per scan, N uint32_t values. Bit layout: lower 16 bits = semantic class ID; upper 16 bits = temporally-consistent instance ID. Load with np.fromfile(path, dtype=np.uint32). Class remapping defined in semantic-kitti-api/config/semantic-kitti.yaml.
• Metric: mIoU, TP/FP/FN aggregated globally across all test frames before dividing.
• Submission server: CodaLab; requires institutional or corporate email (free-provider addresses rejected). No published hard annual cap, but arXiv preprint link required — prevents anonymous score fishing.
• Licence: CC BY-NC-SA. Must cite both SemanticKITTI (ICCV 2019) and the KITTI Vision Benchmark.
• Known pitfalls: HDL-64E is discontinued; modern dense-pillar encoders gain easy points partly from its simple beam pattern. Moving-object labels were generated by tracking — short-lived motion can be mislabelled. Class imbalance is severe: "person" ~0.2%, "motorcyclist" ~0.08% of all points. Sequence 08 shares road geography with nearby training sequences, so val mIoU can overfit the geography; val numbers typically run 2–4 points above test mIoU.

SemanticTHAB — modern high-resolution MLS sensor proxy

• Sensor: Ouster OS2-128 Rev7, a modern high-resolution 128-channel spinning LiDAR rather than the older HDL-64/HDL-32 units common in the canonical driving benchmarks.
• Scale: 4,750 labelled 3D LiDAR point clouds from urban autonomous-driving environments.
• 20 semantic classes: The Zenodo record lists road, car, pedestrian, and building among the class set; use the repository taxonomy before building a loader.
• Access: Zenodo v3, published 2025-02-21; associated repository kav-institute/SemanticLiDAR.
• Metric: Intended for SemanticKITTI-style semantic-segmentation evaluation, but without a hidden public leaderboard.
• Known pitfalls: This is a high-resolution single-cloud sensor proxy, not an aggregated-map benchmark. Its value for an airside map pipeline is density and modern-beam-pattern coverage, not map-scale tiling or multi-session change handling.

nuScenes-lidarseg — the multi-sensor urban reference

• Sensor: 1× Velodyne HDL-32E (32-beam, 360°), keyframe-only at 2 Hz. Boston (USA) and Singapore; 1,000 scenes × 20 s each.
• Scale: ~1.4 billion annotated LiDAR points across ~40,000 keyframes.
• Split: Train 700 / Val 150 / Test 150 scenes; test labels withheld.
• 16 evaluated classes: barrier, bicycle, bus, car, construction_vehicle, motorcycle, pedestrian, traffic_cone, trailer, truck, driveable_surface, other_flat, sidewalk, terrain, manmade, vegetation. Mapping from raw 32-class ontology defined in lidarseg_idx2name_mapping.json in the nuScenes devkit.
• Label file format: Per-sample binary .bin files, one uint8 per point (class index 0–15; 0 = void/noise). Named {lidar_sample_data_token}_lidarseg.bin.
• Metric: mIoU over 16 classes.
• Submission server: EvalAI. Hard cap: 3 submissions per team per year. This is the tightest leaderboard restriction in the major AV benchmarks — plan submissions carefully.
• Extension — Panoptic nuScenes: Adds per-point instance IDs for 7 thing classes; introduces PAT metric and LSTQ (LiDAR Segmentation and Tracking Quality).
• Licence: CC BY-NC-SA; download requires registration.
• Known pitfalls: No explicit "road" class — "driveable_surface" covers road plus nearby dirt, which confuses transfer to SemanticKITTI. HDL-32E has lower vertical resolution than HDL-64E; SemanticKITTI-trained methods may underperform without density augmentation. Keyframe-only (2 Hz) cannot test per-scan online inference rates. Singapore scenes use left-hand traffic convention — affects flip augmentations.

Waymo Open Dataset — 3D Semantic Segmentation

• Sensor: 5× LiDARs — 1× mid-range top (64-channel, truncated to 75 m) + 4× short-range (front/sides/rear, truncated to 25 m). Two returns per beam. Diverse US cities, mixed urban/suburban/highway.
• Scale: ~1,150 segments; approximate split ~24,000 training / 7,000 validation / 3,000 test frames. Per-point labels added March 2022.
• 23 classes: car, truck, bus, other_vehicle, motorcyclist, bicyclist, pedestrian, sign, traffic_light, pole, construction_cone, bicycle, motorcycle, building, vegetation, tree_trunk, curb, road, lane_marker, other_ground, walkable, sidewalk (22 named + implicit "undefined/void").
• Metric: Per-class IoU + mIoU over 22 named classes.
• Submission: waymo.com/open platform; challenge closes ~May each year; year-round submission permitted post-challenge. 2024 SOTA: PTv3-Extreme 72.76% test mIoU (1st); MixSeg3D 69.83% (2nd); vFusedSeg3D (3rd, camera-LiDAR fusion).
• Licence: Waymo Open Dataset Licence (non-commercial, no redistribution).
• Known pitfalls: Multi-return (two returns per ray) is unusual — most frameworks use strongest or last return only. Short-range LiDARs have a different beam pattern from the top LiDAR; naive concatenation causes density artefacts. "Lane marker" is typically the lowest per-class IoU class.

Semantic3D — the dense static-scan reference

• Sensor: Static terrestrial laser scanners (TLS), multiple units; Central Europe (churches, streets, railroad tracks, squares, villages, castles).
• Scale: >4 billion manually labelled points. Training and test sets approximately equal in size.
• 8 classes: man-made terrain, natural terrain, high vegetation, low vegetation, buildings, hard scape, scanning artefacts, cars. Each point carries (x, y, z, intensity, R, G, B, label).
• Label file format: Plain-text .labels files, one integer per line, index-aligned with .txt point files.
• Two test splits: semantic-8 (full) and reduced-8 (subsampled, lower submission cost). Always state which split is used — numbers are not interchangeable.
• Train/test split: No official validation split; community convention is to hold out a random ~20% of training tiles.
• Metric: mIoU (Jaccard) + per-class OA, computed server-side.
• Submission server: http://semantic3d.net/ — upload prediction files. No annual submission cap.
• Licence: Custom academic-use licence (free for non-commercial research).
• Known pitfalls: Static TLS viewpoint — radial density gradient from scanner origin; range-image projection methods adapted for spinning LiDAR will degrade. "Scanning artefacts" class (ghosting, multi-path returns) is highly variable. No moving objects; geometry density ranges 100–1,000 pts/m² near scanner but sparse at range — aggressive subsampling is standard.

Paris-Lille-3D (NPM3D) — the primary MLS pre-training source

• Sensor: Stereopolis II MLS system (Velodyne HDL-32E). On-vehicle survey drives through Paris and Lille, France.
• Scale: 1,940 m total distance, 143.1 million points, two cities.
• Class counts: Raw ontology = 50 fine-grained classes. Benchmark uses 10 coarse classes: Ground, Building, Pole, Bollard, Trash can, Barrier, Pedestrian, Car, Natural, Other.
• Label file format: .txt with (x, y, z, intensity, label_id). No RGB. Coarse-class integer IDs in the benchmark split.
• Metric: Average IoU (Av.IoU) over 10 coarse classes. Submissions via NPM3D benchmark page at npm3d.fr.
• Licence: CC BY-NC-ND 3.0 — non-commercial, no derivatives.
• Known pitfalls: 50-class fine taxonomy is too granular for most DL pipelines; the 10-class coarse reduction loses structure that could be useful for airside transfer. Only ~143 M points — smaller than SemanticKITTI for deep network training. Urban Western-European road geometry only.

Toronto-3D — the road-marking reference

• Sensor: Teledyne Optech Maverick mobile laser scanner (vehicle-mounted, 32-channel equivalent) + RGB camera. ~1,000 pts/m² on road surfaces.
• Scale: ~78.3 million points, 4 segments (L001–L004) over ~1 km of Spadina Avenue, Toronto.
• Split: Standard — 3 segments train, 1 segment test; the community commonly uses L002 as test.
• 8 classes: Road, Road marking, Natural, Building, Utility line, Pole, Car, Fence.
• Label file format: LAS/LAZ with (x, y, z, R, G, B, intensity, GPS time, scan angle rank, label) as integer attribute.
• Metric: mIoU + Overall Accuracy.
• Licence: Open; research use.
• Known pitfalls: Only 1 km of a single road in one city — minimal geographic diversity. "Utility line" and "Road marking" are very sparse (<1% of points each); mIoU masks poor rare-class performance. Single MLS run; no multi-season or weather variation.

KITTI-360 — the accumulated MLS map reference

• Sensor: 2× Velodyne HDL-64E (one spinning, one pushbroom/FARO-style) + fisheye + perspective cameras. Vehicle-mounted, ~73.7 km of suburban drives in Karlsruhe.
• Scale: >150,000 images; ~1 billion annotated 3D points in the accumulated, registered multi-scan map.
• Classes: 37 raw semantic/instance labels; 3D segmentation benchmark evaluates over ~19 Cityscapes-aligned classes. Full 37-class annotation in .xml bounding-box and dense point-cloud layers.
• Aggregated-map nature: KITTI-360 is an explicitly accumulated, registered point cloud (all scans fused into a single metric map with per-point timestamps) — more directly relevant for MLS survey-map segmentation than per-scan benchmarks.
• Label file format: LAS-style binary with semantic + instance labels; loaded via the KITTI-360 devkit.
• Metric: mIoU; leaderboard at cvlibs.net/datasets/kitti-360/.
• Licence: CC BY-NC-SA.
• Known pitfalls: Pushbroom LiDAR produces a different scan pattern than rotating scanners. Large area means ground/building dominates class distribution. Annotations were semi-automated (2D-to-3D lifted from camera labels) with known artefacts at annotation boundaries.

DALES — the airborne LiDAR density bracket

• Sensor: Airborne LiDAR system (ALS), fixed-wing platform, ~500 m altitude. Dayton, Ohio, ~10 km² of mixed urban/suburban/commercial.
• Scale: >500 million hand-labelled points, 40 scenes (~0.25 km² each).
• 8 classes: Ground, Vegetation, Cars, Trucks, Power lines, Fences, Poles, Buildings.
• Split: Typically 30 scenes train / 10 scenes test. No official validation split — community convention varies.
• Label file format: LAS files with class attribute; intensity and return-number available.
• Metric: Mean IoU, per-class IoU, Class Consistency Index (CCI), Overall Accuracy. Extension: DALES Objects (2021) adds instance-level annotations.
• Licence: Open for research.
• Known pitfalls: Aerial nadir viewpoint — density drops severely for vertical objects (poles, fences). "Power lines" are extremely sparse (one return per line segment); all existing methods achieve very low IoU on this class. US Midwest geography only.

GridNet-HD — multimodal utility-infrastructure LiDAR/image benchmark

• Sensor / acquisition: UAV LiDAR plus high-resolution oblique RGB imagery with camera poses and calibration files. The official dataset card describes 36 geographic zones, each with images, masks, LiDAR, and pose folders.
• Scale: 7,694 images and 2,448,762,950 LiDAR points across the official train/test split.
• 11 evaluated semantic groups: pylon, conductor cable, structural cable, insulator, high vegetation, low vegetation, herbaceous vegetation, rock/gravel/soil, impervious soil/road, water, building; the unassigned/unlabeled group is ignored.
• Tasks and baselines: LiDAR-only Superpoint Transformer, image-vote projection, and late-fusion MLP baselines are released. The dataset card reports 66.90 mIoU for SPT, 69.10 for ImageVote, and 74.22 for late fusion.
• Access and license: Hugging Face dataset and leaderboard; CC-BY-4.0. The card recommends huggingface_hub.snapshot_download rather than the auto-converted Parquet dataset view.
• Benchmark role: Best current public proxy for long-thin utility infrastructure and calibrated LiDAR-image fusion in a large dense point cloud. It is relevant to poles, masts, cables, gantries, and overhead/edge infrastructure in non-road sites.
• Known pitfalls: UAV utility-corridor acquisition is not vehicle-mounted MLS and is not airside. Treat it as a thin-class and fusion stress test, not as validation for pavement semantics, FOD retention, GSE, or dynamic map cleaning.

SensatUrban — the scale / tiling benchmark

• Sensor: Photogrammetric dense point clouds (SfM + MVS from UAV oblique imagery — NOT direct LiDAR pulses). The dataset paper reports Birmingham, Cambridge, and York, covering 7.6 km²; benchmark descriptions may refer to the labelled two-city release subset.
• Scale: Nearly 3 billion annotated points.
• 13 classes: Ground, Vegetation, Building, Wall, Bridge, Parking, Rail, Traffic Road, Street Furniture, Car, Footpath, Bike, Water.
• Split: Train/test by tile as provided in the repository; no formal validation split.
• Label file format: PLY files with per-point integer class label.
• Metric: mIoU; CodaLab submission (competitions.codalab.org/competitions/31519). Urban3D Challenge hosted at ECCV 2022.
• Licence: Academic use; requires registration.
• Known pitfalls: Photogrammetric point clouds — different noise and density distribution from spinning LiDAR; intensity not available. "Rail," "Bridge," and "Bike" classes are rare; mIoU can mask near-zero performance on them.

STPLS3D — synthetic-to-real aerial

• Sensor (real): DJI Phantom 4 Pro, 75–85% image overlap, 25–70 m altitude, SfM+MVS reconstruction. 4 real sites.
• Sensor (synthetic): Unreal Engine 4 / AirSim rendering, Bentley ContextCapture reconstruction — 3 synthetic versions (SyntheticV1–V3) totalling ~16 km².
• Scale: 1.27 km² real + 16 km² synthetic.
• 18 classes (SyntheticV3): Building, Low vegetation, Medium vegetation, High vegetation, Passenger car, Truck, Aircraft, Military vehicle, Bike, Motorcycle, Light pole, Street sign, Clutter, Fence, Road, Grass, Window, Dirt. Real data uses a subset.
• Split: Real test = WMSC site; generalisation tested on Fort Drum cantonment.
• Label file format: Point-wise semantic + instance annotation, compatible with ScanNet/S3DIS loading conventions.
• Metric: Semantic mIoU + OA; instance mAP, mAP50, mAP25.
• Licence: Academic research use.
• Known pitfalls: Synthetic-to-real gap persists despite pipeline-level simulation; domain adaptation needed. "Aircraft" class appears only in synthetic data (USC airport proximity) — not well-represented in real tiles. Real tiles are small (1.27 km²) relative to synthetic.

CUS3D and SUM Parts — urban mesh / photogrammetric map references

• CUS3D acquisition: UAV photogrammetry from 10,840 aerial images, reconstructed into a 2.85 km² urban/rural scene with point cloud, triangular mesh, and high-resolution 2D imagery.
• CUS3D scale and labels: 152,298,756 labelled 3D points, 289,404,088 labelled triangles, and 10 semantic categories: building, road, grass, car, high vegetation, playground, water, building site, farmland, and ground.
• CUS3D benchmark role: Not a LiDAR dataset. Its value is a geometry+RGB+mesh reference for pipelines that colorize a LiDAR map, mesh it for inspection/simulation, or compare point-cloud and mesh segmentation heads. The paper reports KPConv as the best of six tested 3D methods with 59.72% mIoU, 89.42% OA, and 97.88% mAcc.
• SUM Parts acquisition: Textured urban meshes, derived from the SUM mesh lineage, with part-level semantic annotations over about 2.5 km².
• SUM Parts classes: 21 part-level classes, including terrain, high/low vegetation, water, car, boat, wall, roof surface, facade surface, chimney, dormer, balcony, roof installation, window, door, impervious surface, road, road marking, cycle lane, and sidewalk.
• Benchmark role: SUM Parts is the right proxy when the release artifact is a textured mesh or a digital twin layer rather than a raw point cloud. It is especially useful for class-taxonomy design around markings, sidewalks, facades, roofs, and other map-product surfaces that do not map cleanly onto a simple road/ground/building ontology.
• Known pitfalls: Photogrammetric and mesh datasets do not reproduce LiDAR pulse noise, multi-return behaviour, intensity calibration, or vehicle-height occlusion. Treat them as RGB/mesh pipeline references and annotation-schema pressure tests, not as LiDAR-statistics substitutes.

S3DIS — the indoor structural reference

Architecture papers routinely report on S3DIS for benchmarking backbone generality. Cross-reference with the method pages before assuming numbers are comparable to outdoor benchmarks.

• Sensor: Matterport RGB-D cameras + structured-light scanners. 6 large indoor areas across 3 Stanford buildings; 271 rooms total.
• Scale: 695 million reconstructed points, per-point (x, y, z, R, G, B) + label.
• 13 classes: ceiling, floor, wall, beam, column, window, door, table, chair, sofa, bookcase, board, clutter.
• Evaluation protocol: Two conventions appear in the literature — always specify which:
  • Standard: Area 5 held out for test, Areas 1,2,3,4,6 for train. Report mIoU.
  • Alternate: 6-fold cross-validation (each area held out once), report mean mIoU and mean OA.
• Label file format: Per-point .txt annotation files.
• Metric: mIoU (primary), OA. No hidden test server — evaluation is reproducible locally.
• Licence: Non-commercial academic.
• Known pitfalls: RGB-D fused reconstruction differs from LiDAR pulse data — colour features matter here but are absent in most outdoor benchmarks. Area 5 is disproportionately large, making train/test size asymmetric. Near-perfect Area 5 performance does not generalise to real-world interior MLS scans.

ScanNet v2 / ScanNet200 — the long-tail indoor reference

Architecture papers consistently report on ScanNet v2 and ScanNet200 for head/tail class performance. Required for cross-reference with the method pages.

• Sensor: BundleFusion structured-light depth sensor + RGB camera, handheld scanning. 1,513 scans of 707 distinct indoor spaces.
• Scale (v2): 1,613 training / 312 validation / 100 test scans; evaluated on 20 object classes.
• ScanNet v2 — 20 benchmark classes (selected from 40 annotated categories): common indoor objects.
• ScanNet200: Extends to 200 classes grouped as Head (66) / Common (68) / Tail (66). mIoU is evaluated per group to expose tail-class collapse — report all three group scores alongside the aggregate.
• Label file format: Volumetric mesh + per-vertex labels; also provided as fused voxel grids at 2 cm resolution.
• Metric: mIoU (semantic); mAP at IoU 25%, 50%, 50–95% (instance). Hidden test-server submission at ScanNet portal.
• Licence: Non-commercial academic.
• Known pitfalls: Handheld scan — non-uniform density, noisy depth — very different from vehicle-mounted or UAV LiDAR. ScanNet200 Tail classes may have only 1–3 training scans; few-shot generalisation is the real challenge. v1/v2 distinction matters: v2 has cleaner annotations; some papers still cite v1 numbers without flagging this.

FRACTAL — the large ALS corpus

• Sensor: Leica, Riegl, and Teledyne/Optech ALS platforms (French Lidar HD national programme, ~10 pulses/m², ~37 pts/m²). Colourised from ORTHO HR 0.20 m NRGB imagery.
• Scale: 250 km² total, 9,261 million points, 100,000 tiles of 50×50 m from 5 spatial domains in southern France.
• 7 classes: Ground, Vegetation, Building, Water, Bridge, Permanent structure, Other.
• Split: 80/10/10 (train 225 km² / val 25 km² / test 25 km²). Test from geographically separate areas.
• Label file format: LAZ 1.4 with (x, y, z, intensity, return info, scan angle, colour channels NRGB).
• Metric: mIoU (primary), OA (secondary). Baseline (RandLa-Net): mIoU 77.5%, OA 86.7%.
• Access: HuggingFace IGNF/FRACTAL. Licence: permissive open (French government open-data compatible).
• Known pitfalls: Southern France only — forest and scrubland dominate the vegetation class. Imagery and LiDAR are not co-temporal (up to 3-year lag) — RGB/NIR features may be misaligned. "Other" class is semantically ill-defined. Very coarse taxonomy — no vehicles, poles, or infrastructure detail.

GOOSE / GOOSE-Ex — the off-road / unstructured-environment pair

• Sensor (GOOSE, MuCAR-3 vehicle): 1× Ouster OS1-128 (128-channel spinning LiDAR), 2× Ouster OS1-32, front-facing RGB+NIR camera. Mixed outdoor environments — forest, gravel roads, fields, campus, paved areas. Germany-based.
• Sensor (GOOSE-Ex): Liebherr R924 excavator + Boston Dynamics Spot quadruped. Excavator: 3× Ouster OS1-64 + 1× Ouster OS1-128 at boom + prism cameras. 4 environmental settings: construction, quarry, landfill, open terrain.
• Scale: GOOSE 10,000 annotated frame pairs; GOOSE-Ex 5,000 annotated multimodal frames (2,800 excavator + 2,200 Spot), 100+ sequences. Combined ~13,076+ labeled clouds.
• Ontology: 64 classes hierarchically organised (8 superclasses); ~90% of annotated area is vegetation/terrain/sky. Classes with <20 occurrences excluded from evaluation. GOOSE-Ex subset uses 36 classes present in the distribution.
• Label format: SemanticKITTI-compatible point-wise label files. Per-point labels produced from multiple frames merged by platform odometry — i.e., aggregated clouds, same accumulation regime as an aggregated map.
• Split (GOOSE-Ex): Train 3,989 / Val 407 / Test 604 (test labels withheld).
• Metric: mIoU. GOOSE 3D Semantic Segmentation Challenge hosted at ICRA 2025 (arXiv:2506.06995).
• Licence: Open at goose-dataset.de.
• Known pitfalls (GOOSE-Ex): Excavator top-mounted LiDAR observes mostly ground-plane geometry; vertical objects appear at extreme angles — architectures trained on standard vehicle-height sensor configurations will encounter inverted density profiles.

SemanticRail3D — the MLS corridor reference

• Sensor: Optech LYNX Mobile Mapper (two LiDARs on MLS), ~980 pts/m², 5 mm precision.
• Scale: 87.6 km total, 438 × 200 m segments, ~2.8 billion annotated points.
• Classes: 11 or 12 semantic classes plus instance labels. The v1/v2 distinction introduces a possible class count discrepancy — verify in the Scientific Data publication before citing exact numbers.
• Split: 5 training shards + validation + held-out test (labels reviewed by railway domain experts).
• Label file format: LAZ with (x, y, z, intensity, class, instance_id).
• Metric: mIoU (implied; not explicitly stated in public documentation — verify in full paper).
• Licence: Not confirmed in open documentation — check before use.

WHU-Railway3D — the multi-environment railway reference

• Sensor: 3 Mobile Mapping Systems: Optech Lynx (urban), HiScan-Z dual-sensor MLS (rural), 32-line rMMS (plateau, China).
• Scale: 4.6 billion annotated points, 120 tiles across 3 environments (urban/rural/plateau).
• 11 classes: Rails, Track bed, Masts, Support devices, Overhead lines, Fences, Poles, Vegetation, Buildings, Ground, Others.
• Split: 60/20/20 per environment (tile-based).
• Label file format: .npy files, uint8, values 0–10.
• Metric: mIoU, per-class IoU, Overall Accuracy.
• Licence: Not specified (IEEE TITS paper) — verify before use.

WHU-Urban3D — MLS + ALS urban semantic and instance benchmark

• Sensor / acquisition: Mixed MLS and ALS point clouds over large road and urban scenes.
• Scale: Official site reports 3.6 × 10^6 m² and more than 300 million points, with point-wise semantic and instance labels.
• Tasks: Semantic segmentation, instance segmentation, and object detection; countable classes also carry 3D bounding boxes.
• Semantic benchmark classes: Current benchmark page reports 18 MLS semantic classes and 7 ALS semantic classes. The MLS class list includes tree, non-driveway, building, box, light, electrical/municipal poles, low vegetation, board, driveway, mark, vehicle, pedestrian, traffic light, detector, fence, wire, and pole; the ALS list is coarser.
• Metrics / submission: Semantic segmentation reports oAcc, mAcc, and mIoU; instance segmentation reports coverage, precision/recall, F1, and semantic metrics. The benchmark page currently states result upload is not available.
• Airside transfer: High for dense urban district mapping, service-road corridors, and urban non-road campus environments because it combines ground-level MLS and aerial ALS views. It is also a good stress test for the "same place, different acquisition geometry" problem that airside survey and site-survey products will face.
• Known pitfalls: Licence and redistribution terms are not explicit in the public landing page. Use it for research benchmarking unless terms are confirmed.

Recent additions: non-road, site-survey, utility, campus, and industrial proxies

• Turin3D — ALS, ~1.43 km², ~70 M points, central Turin. Training set is unlabeled by design; only validation and test sets are annotated. Purpose-built for semi-supervised and domain-adaptation evaluation, not a fully-supervised leaderboard. Classes not enumerated in public abstract — verify in arXiv:2504.05882.
• CITYLID — Citywide ALS of Berlin (1,060 tiles, ~15 billion points). 13 classes: 3 standard (ground, buildings, trees); 5 fine street features (medians, driveways, bikepaths, walkways, on-street parking); 5 shadow bins (0–3, 3–5, 5–7, 7–10, 10–12 h). Licence: DL-DE (allows commercial + research use, redistribution). No quantitative DL benchmark — visual/qualitative validation only; the researcher must define their own splits and metrics. Access: HuggingFace Deepank/CITYLID.
• SIP (Site in Pieces) — TLS single-station, 40 scenes (27 indoor, 13 outdoor construction sites), ~140–200 M total points, 23 classes covering structural elements, temporary objects, and site context. Single-station viewpoint preserves radial density decay and self-occlusion — the explicit counterpoint to aggregated maps. Licence: publicly available (terms in arXiv:2512.09062).
• Waymo-4DSeg / SAM4D — Pseudo-labeled, cross-modal (camera + LiDAR) masklet dataset (~30 M LiDAR masks). Labels are class-agnostic (instance/masklet identity, not a fixed taxonomy). Pre-training and promptable-segmentation resource, not a closed-taxonomy leaderboard. Released as part of SAM4D (arXiv:2506.21547).
• ECLAIR — Aerial LiDAR over urban/utility corridors, covering about 10 km² with close to 600 M points, 11 semantic classes, colorized points, and 1,246 tiles (ground-truth plus pseudo-label tiles). Strong aerial/site-survey and utility-infrastructure proxy; not ground-level MLS and not direct airside acceptance evidence.
• YUTO Semantic — ALS campus dataset over York University, with approximately 738 M points over 9.46 km² and 9 semantic classes. Use it as a campus-scale aerial/site-survey proxy; do not treat it as ground-vehicle MLS.
• S.MID — Industrial substation LiDAR dataset captured by an industrial robot with Livox Mid-360 hybrid-solid LiDAR: 38,904 frames, 25 annotated categories merged to 14 single-frame evaluation classes, and SemanticKITTI-like .bin / .label files. Use it as an industrial hybrid-solid LiDAR proxy, not as map-scale aggregated segmentation evidence.
• OpenTrench3D — 310 photogrammetric RGB point clouds of open utility trenches, totaling about 528 M points and 5 utility/trench classes. It is useful for underground-utility, works-zone, and RGB point-cloud taxonomy pressure, but it is not LiDAR and should not be used as LiDAR sensor-noise evidence.
• MLDAS — Multi-LiDAR domain-adaptation dataset with 31,875 synchronized 128/64/32-beam LiDAR scans across campus and urban-street scenarios, annotated into 14 classes. Use it for sensor-transfer and density-robustness experiments; access is email-gated and non-commercial.
• USCILab3D — Long-term USC campus robot dataset with 5 cameras, 32-beam 360° LiDAR, pose-stamped data, foundation-model-assisted 2D-to-3D semantic labels, and reported 267 semantic categories. Treat it as a high-interest campus/label-backprojection watch item because the project site says raw bags and processing code are available while processed data is still coming soon.
• Industrial3D — 2026 TLS industrial-infrastructure benchmark with 612.7 M labeled points at 6 mm resolution, 12 MEP/structure classes, and cross-paradigm baselines. The public repository still marks full dataset/code release as tied to paper acceptance, so keep it as a watchlist/proxy note until the full release is available.
• Point Cloud City / Open3D-ML PCC — NIST public-safety indoor point-cloud collections from Enfield, Memphis Map901, and Hancock County, routed through an Open3D-ML integration with dataset processing, SemanticKITTI-style conversion, KPConv/RandLA-Net configuration examples, and TensorFlow/PyTorch semantic-segmentation pipelines. The 2026 NIST WF-PST publication analyzes Enfield/Memphis label unification and KPConv-compatible IoU evaluation, and explicitly warns that class imbalance and small safety-critical features remain hard. Use it as a managed-building, terminal-interior, and label-harmonization proxy; do not treat it as outdoor AV, airside, or leaderboard-grade evidence. The GitHub repository is no longer maintained by NIST, and collection access varies by awardee.
• City-Facade — 2026 ISPRS JPRS MLS building-facade dataset from Xiamen, China, with approximately 200 M labeled points over more than 60 km of roads for semantic and instance segmentation of facade elements. The official repository describes labeled facade clouds with 9 classes plus unlabeled street-landscape clouds, and the paper lists wall, window, door, roof, advertisement, air conditioner, rain shed, balcony, and other/unclassified labels. Use it as a vertical-structure, terminal-frontage, BIM/digital-twin, and facade-sublabel proxy; do not treat it as a full map benchmark for roads, markings, poles, vegetation, vehicles, or dynamic-object removal. Verify full train/test access and license terms before benchmark-grade or production claims.
• ZAHA — WACV 2025 / TUM2TWIN facade benchmark derived from real mobile laser scans over Munich/TUM campus facades, with 601 M annotated points across 66 facades. Its Level of Facade Generalization (LoFG) hierarchy evaluates both five coarse LoFG2 groups (floor, decoration, structural, opening, other elements) and 15 fine LoFG3 classes (wall, window, door, balcony, molding, deco, column, arch, stairs, ground surface, terrain, roof, blinds, interior, other). Use it as the strongest public facade-hierarchy stress test for terminal frontage, BIM/digital-twin, and vertical-structure sublabels; do not treat it as a road/yard/apron benchmark, a LiDAR-image benchmark, or a dynamic-object-removal source.

Cross-Dataset Taxonomy Mismatch

Naively combining datasets from different sources introduces label conflicts that can degrade performance — measured as negative transfer. The PTv3 Point Prompt Training paper documents −3.3% to −2.1% mIoU degradation from naive multi-dataset joint training vs. single-dataset baselines.

Core incompatibilities

Issue	Example
Coarser vs. finer granularity	nuScenes "manmade" vs. SemanticKITTI's separate building, fence, pole, traffic-sign classes
Different ground definitions	nuScenes has no "road" class; "driveable_surface" is broader than SemanticKITTI "road"
Moving-object labelling	SemanticKITTI has explicit moving-variant classes; nuScenes and Waymo use instance IDs instead
Vegetation hierarchy	Waymo separates "vegetation" and "tree_trunk"; SemanticKITTI merges them
Evaluation class count	SemanticKITTI evaluates 19/25; nuScenes 16; Waymo 22 — no universal set

Safe cross-dataset class overlap across SemanticKITTI, nuScenes, and Waymo is approximately 10 classes: car, bicycle, motorcycle, truck, other_vehicle, pedestrian, driveable_surface/road, sidewalk, terrain, vegetation. Only SemanticKITTI, SemanticUSL, and SemanticSTF share label definitions directly.

Label-mapping approaches

Coarse common-class mapping (standard practice): Reduce all datasets to the lowest-common-denominator 7–10 class subset for unified training. Simplest to implement; loses fine-grained structure.

Union label space (M3Net / UniSeg approach): Train with a union of all class labels, keeping dataset-specific heads but sharing a backbone. Requires negative-sample masking — only the source dataset's class space is valid for each sample. M3Net (CVPR 2024) reports 72.0% mIoU on SemanticKITTI and 80.9% on nuScenes using this strategy. Source: arXiv:2405.01538.

Language-guided alignment (Point Prompt Training / PPT): Use CLIP text embeddings of class names to find semantic correlations across datasets, enabling soft label transfer between semantically similar classes. Demonstrated for indoor datasets (ScanNet + S3DIS + Structured3D) at CVPR 2024. The same principle can be applied to outdoor driving datasets.

OmniLiDAR harmonisation: Aggregates 12 public driving datasets (~22 million raw scans, 15 distinct sensor models) with standardised label spaces and raw-scan normalisation. Enables cross-domain ground segmentation (TerraSeg). Source: arXiv:2603.27344.

Negative-transfer risk: Any multi-dataset training that omits explicit alignment (label-mapping layer, negative-sample masking, or language-guided alignment) risks measurable mIoU degradation on every target domain. Always ablate single-dataset vs. multi-dataset baselines before deploying a multi-source training pipeline.

Synthetic Datasets and Sim-to-Real

Synthetic data addresses two problems: (1) pre-training on large corpora without annotation cost, and (2) rare-class rebalancing — real datasets have insufficient examples of uncommon objects (motorcyclists, cyclists, pedestrians) to train reliable detectors.

SynLiDAR — large-scale pre-training corpus

• Generator: Custom purpose-built LiDAR simulator (not CARLA), virtual urban/suburban worlds.
• Scale: 13 sequences, ~19 billion points, ~20,000 scans. 32 classes (superset of SemanticKITTI + additional categories). Class definitions in annotations.yaml.
• Label format: .label files, SemanticKITTI-compatible (uint32).
• Licence: MIT.
• Sim-to-real results: Adding SynLiDAR to SemanticKITTI training improves mIoU; Point Cloud Translator (PCT) reduces the domain gap further. Effective as a pre-training corpus before fine-tuning on real data. Source: arXiv:2107.05399 (AAAI 2022).

RareBoost3D — CARLA-based rare-class rebalancing

• Generator: CARLA v0.9.15 (Unreal Engine). Simulated Velodyne HDL-64E: 64 channels, ~125k–138k pts/scan.
• Scale: 8 sequences × 60,000 scans = 480,000 scans total. Environments: rural (Map 7), small towns (Maps 1,2,4), large cities (Maps 3,5,6,10).
• Classes: 29 raw labels mapped to 16 unified evaluation classes aligned to SemanticKITTI.
• Sim-to-real technique: Cross-domain Semantic Consistency (CSC) loss — contrastive alignment of feature representations across domains.
• Gains: +2.5% mIoU on SemanticKITTI val (MinkUNet baseline); rare classes see +3–10% per-class IoU improvement.
• Source: arXiv:2510.10876.

SynthmanticLiDAR — SemanticKITTI-aligned CARLA dataset

• Generator: Modified CARLA simulator with additional classes and consistent labelling with SemanticKITTI.
• Nature: Designed specifically for sim-to-real transfer to SemanticKITTI; sequential point cloud dataset.
• Access: Dataset + simulator code on GitHub. Source: arXiv:2501.19035.

KITTI-CARLA — domain adaptation benchmark

• Generator: CARLA v0.9.10, KITTI-like output format.
• Purpose: Transfer-learning benchmark from synthetic to real (SemanticKITTI). Purpose-built for evaluating domain adaptation methods. Source: ResearchGate 2021.

STPLS3D synthetic split — aerial sim-to-real

The STPLS3D synthetic split (SyntheticV1–V3, ~16 km², up to 18 classes) is the primary aerial photogrammetry sim-to-real resource. See the STPLS3D section above for sensor and format details. The "Aircraft" class in the synthetic split is particularly relevant for airside pre-training, though it is not represented in the real tiles.

Sim-to-real considerations for airside use

No purpose-built airside synthetic LiDAR dataset exists. The closest available resources are STPLS3D synthetic (aerial, includes "Aircraft" class), RareBoost3D (ground-vehicle, rare-class rebalancing methodology transferable to airside GSE), and SynLiDAR (large-scale urban pre-training). Any sim-to-real airside pipeline will require: (1) domain-specific world building (apron geometry, aircraft stands, GSE classes), (2) an explicit sim-to-real alignment step (CSC loss, style transfer, or domain randomisation), and (3) validation on real airside scans before deployment.

Metrics and Reporting Conventions

Primary metric: mIoU

Mean Intersection-over-Union is the standard primary metric across all benchmarks. Always report per-class IoU alongside the aggregate mIoU — a benchmark with one dominant class (ground, building) hides rare-class collapse behind a healthy headline number. On SemanticKITTI, "person" and "motorcyclist" IoU are the most informative single-class indicators of rare-object performance.

For ScanNet200, report mIoU separately for the Head, Common, and Tail groups — aggregate mIoU alone is meaningless for tail-class analysis.

Secondary metrics

• Overall accuracy (OA): reported for comparability on Semantic3D, DALES, Toronto-3D; dominated by ground/building and should not be used as a primary ranking metric.
• Instance segmentation metrics: mAP at IoU thresholds 25%, 50%, 50–95% used by ScanNet, STPLS3D (instance task), and DALES Objects.
• Panoptic metrics: PAT, LSTQ (nuScenes Panoptic); PQ/SQ/RQ (ScanNet panoptic). Report all three components when available.

Submission-limit and leaderboard-hygiene rules

Benchmark	Annual submission cap	Notes
nuScenes-lidarseg (EvalAI)	3 per team per year	Tightest cap; plan submissions carefully
SemanticKITTI (CodaLab)	No published hard cap	Institutional email required; arXiv link required — prevents anonymous score fishing
Waymo 3D Seg (waymo.com)	Challenge closes ~May each year	Year-round submission permitted post-challenge
SensatUrban (CodaLab)	No confirmed cap	Standard CodaLab convention
Semantic3D	No annual cap	Online prediction upload

Validation-vs-test discipline: On SemanticKITTI, val mIoU typically runs 2–4 points above test mIoU because sequence 08 is geographically close to training sequences and is often used for hyperparameter selection. Always develop against a fixed validation split, submit to the server sparingly, and never select hyperparameters on the test set.

Multi-dataset pre-training disclosure: Increasingly, top-performing entries (e.g., PTv3-Extreme in the 2024 Waymo Challenge) use pre-training on multiple datasets before fine-tuning on the target benchmark. The label-mapping layer used for pre-training is often not disclosed. When citing leaderboard results, note whether the method uses: (a) single-dataset training only, (b) multi-dataset joint training, or (c) a large-scale 2D/3D pre-trained backbone (DINOv2, PointMAE, etc.).

Overfitting risk: CodaLab and EvalAI leaderboards that display individual score updates enable iterative tuning against the held-out test set. Scores are sometimes rounded (e.g., to 2 decimal places) to reduce exploitation. The SemanticKITTI leaderboard's arXiv-link requirement and nuScenes' 3-submission annual cap exist precisely to prevent this.

A SOTA claim is only meaningful with the task variant, split, evaluation source, and pre-training regime all stated. See ../overview/aggregated-map-semantic-segmentation.md §13.3 for the full benchmarking-rigor protocol.

Data Formats and Label Conventions

Benchmark	Format	Point fields	Label encoding
SemanticKITTI	.bin + .label	float32 (x,y,z,intensity)	uint32: lower 16 bits = semantic, upper 16 bits = instance
SemanticTHAB	Zenodo .zip + repo tools	OS2-128 point clouds	20-class semantic labels; verify exact field layout in repo
nuScenes-lidarseg	.bin	float32	uint8 class index (0 = void)
Waymo 3D Seg	tfrecord via devkit	(x,y,z,intensity, return)	per-point class index
Semantic3D	.txt + .labels	(x,y,z,intensity,R,G,B)	one integer per line in .labels
Paris-Lille-3D	.txt	(x,y,z,intensity,label_id)	integer class index; no RGB
Toronto-3D	LAS/LAZ	(x,y,z,R,G,B,intensity,GPS time,scan angle,label)	integer attribute
KITTI-360	LAS-style binary	(x,y,z,intensity,timestamp)	semantic + instance via devkit
DALES / FRACTAL	LAS / LAZ 1.4	(x,y,z,intensity,return info,colour)	class attribute
GridNet-HD	LAS + images + masks + poses	LiDAR point cloud with ground_truth field plus RGB projection assets	grouped semantic class IDs; test labels hidden for leaderboard
ECLAIR	LAZ point clouds + labels.json	aerial LiDAR points with colorized attributes	11-class semantic labels; ground-truth and pseudo-label tile split
YUTO Semantic	LAS/LAZ-style release via Hugging Face	aerial LiDAR points with intensity, returns, GPS time, scan angle, and class	9 semantic classes; one mission currently released
S.MID	SemanticKITTI-like .bin + .label	Livox Mid-360 hybrid-solid LiDAR points	25 raw categories merged to 14 evaluation labels
OpenTrench3D	PLY-style photogrammetric point clouds	XYZRGB point clouds from smartphone-video photogrammetry	5 utility/trench classes; Misc ignored in training/evaluation
MLDAS	synchronized multi-LiDAR frame packages	OS128 / OS64 / XT32 LiDAR scans	14 semantic classes; labels propagated from the 128-beam reference LiDAR
USCILab3D	raw ROS bags; processed point clouds pending	32-beam LiDAR + 5 cameras + poses	267 projected semantic categories reported; processed semantic release to verify
Industrial3D	TLS point-cloud release pending	dense terrestrial industrial MEP scans	12 semantic classes; full release and licence still pending
Point Cloud City / Open3D-ML PCC	collection packages plus Open3D-ML conversion scripts	indoor point clouds and associated project imagery where available	collection-specific labels; Open3D-ML PCC tools support SemanticKITTI-style conversion and unified-label experiments
City-Facade	.txt facade point-cloud segments	x, y, z, LiDAR intensity, instance label, semantic label	integer facade semantic labels plus instance IDs; full train/test package access to verify
ZAHA	TUM2TWIN / official download package	mobile-laser-scan facade points in local and global UTM coordinate frames	LoFG2 and LoFG3 semantic labels; verify exact loader/package layout before training
SensatUrban	PLY	per-point	integer class label
CUS3D	Point cloud + mesh + 2D imagery	RGB geometry from UAV reconstruction	10 semantic classes on 3D points, mesh triangles, and 2D images
SUM / SUM Parts	Textured mesh	mesh faces + texture pixels	6-class SUM; 21-class SUM Parts with face/pixel label variants
GOOSE / GOOSE-Ex	SemanticKITTI-compatible	point-wise	SemanticKITTI label files
SemanticRail3D	LAZ	(x,y,z,intensity,class,instance_id)	explicit class + instance fields
WHU-Railway3D	.npy	per-point	uint8 values 0–10
WHU-Urban3D	Benchmark upload uses .h5	coordinates + labels per scene	semantic and instance labels in labels key; raw package format to verify
S3DIS	.txt	(x,y,z,R,G,B,label)	per-point integer
ScanNet v2/200	mesh + voxel grid	per-vertex	2 cm voxel grid labels

Key compatibility point: SemanticKITTI-compatible format (SynLiDAR, GOOSE, RareBoost3D, S.MID) enables direct reuse of SemanticKITTI-style data loaders after label remapping. LAS/LAZ files (FRACTAL, DALES, ECLAIR, YUTO Semantic, SemanticRail3D, Toronto-3D, KITTI-360) require the laspy or pdal library. Mesh-based formats (ScanNet) require a separate voxelisation or subsampling step before point-wise inference. Release-maturity watch items (USCILab3D processed data, Industrial3D full data/code, and email-gated MLDAS) should be represented in experiments by a dataset manifest that records access date, licence, available modalities, and exact class-map evidence.

Taxonomies do not align across datasets. Cross-dataset training requires an explicit label-mapping layer. See the Cross-Dataset Taxonomy Mismatch section above and ../overview/aggregated-map-semantic-segmentation.md §5.2 for the selection-guide framing.

Licensing

Benchmark	Licence	Production use
SemanticKITTI	CC BY-NC-SA	Non-commercial only
nuScenes-lidarseg	CC BY-NC-SA	Non-commercial only
Waymo Open Dataset	Waymo custom	Non-commercial; no redistribution
Semantic3D	Custom academic	Non-commercial research
Paris-Lille-3D	CC BY-NC-ND 3.0	Non-commercial, no derivatives
Toronto-3D	Open (research)	Research use
KITTI-360	CC BY-NC-SA	Non-commercial only
DALES	Open (research)	Research use
GridNet-HD	CC-BY-4.0	Commercial + research with attribution; verify downstream model/data redistribution rules
ECLAIR	CC BY-NC-SA 4.0 for data; code MIT	Dataset is non-commercial; contact owner for commercial cases
YUTO Semantic	CC BY-NC 4.0	Non-commercial only
S.MID	CC BY-NC-SA 4.0	Non-commercial research only
OpenTrench3D	CC BY-NC 4.0	Non-commercial only
MLDAS	Custom email-gated licence	Non-commercial, non-transferable; no commercial model training
USCILab3D	Processed-data licence unclear	Verify before redistribution or model release
Industrial3D	Dataset licence pending; repo GPL-3.0 for code	Treat as watchlist until full release/licence is explicit
Point Cloud City / Open3D-ML PCC	Collection-specific public-safety data access; repo licence visible on GitHub	Verify each collection's access terms; Open3D-ML PCC code is no longer NIST-maintained
City-Facade	Open-access paper; dataset/full-release terms to verify	Treat as research proxy until full train/test access and license terms are explicit
ZAHA	Official repo is CC0-1.0; dataset/download terms to record per package	Strong research proxy for facade taxonomy; verify dataset terms before commercial training or redistribution
SensatUrban	Academic (registration)	Research use; check before commercial
CUS3D	Open-access paper; data terms to verify	Research use until data licence is confirmed
SUM / SUM Parts	Project/code/data released; data terms to verify	Research use until data licence is confirmed
FRACTAL	French open-data (permissive)	Commercial + research
CITYLID	DL-DE	Commercial + research + redistribution
GOOSE / GOOSE-Ex	Open (goose-dataset.de)	Research use
SynLiDAR	MIT	Commercial permissible
S3DIS	Non-commercial academic	Research only
ScanNet v2/200	Non-commercial academic	Research only
WHU-Railway3D	Not specified	Verify before use
WHU-Urban3D	Not specified on landing page	Verify before use
SemanticRail3D	Not confirmed	Verify before use
SemanticTHAB	Zenodo open dataset	Verify exact Zenodo licence metadata before production

Almost every benchmark is released for research / non-commercial use. They are appropriate for pre-training, architecture selection, and benchmarking, but a model whose weights are shipped in a product should have its supervised fine-tuning grounded in owned in-domain data. Verify each licence before production use — see ../overview/aggregated-map-semantic-segmentation.md §5.2.

Airside Transfer

No airside aggregated-map benchmark exists — the 2023–2026 additions above confirm rather than close that gap: none is airport-specific. For an airside pipeline the transfer-relevant analysis starts from acquisition paradigm.

MLS ground-level, accumulated-map group (Paris-Lille-3D, Toronto-3D, KITTI-360, SemanticRail3D, WHU-Railway3D, WHU-Urban3D): These datasets share a survey-drive viewpoint — a vehicle-mounted multi-beam LiDAR driving a linear corridor, building a registered multi-scan map — that most closely matches a taxiway or apron survey drive. SemanticRail3D and WHU-Railway3D add a constrained linear ODD with structured trackside infrastructure (catenary poles, masts, overhead lines, track-bed marking) that is a closer analog to airside taxiway corridor furniture than open urban road geometry. WHU-Urban3D adds a mixed MLS/ALS urban-district case for non-road campus-style mapping. SemanticKITTI multi-scan task adds explicit static/moving reasoning useful for mixed-traffic apron scenarios.

Per-domain transfer assessment:

Dataset	Paradigm match	Transfer value for airside
KITTI-360	MLS accumulated, registered	High — best viewpoint + density match for taxiway survey drive
Paris-Lille-3D / Toronto-3D / WHU-Urban3D	MLS accumulated or MLS+ALS urban	High — survey-drive viewpoint; urban-district furniture and marking classes
SemanticRail3D / WHU-Railway3D	MLS accumulated, constrained corridor	High — corridor geometry, trackside infrastructure analog
SemanticTHAB	High-resolution MLS single-cloud	Moderate-high — modern OS2-128 density proxy; not map-scale
GOOSE-Ex	Aggregated clouds, large-vehicle platform	Moderate-high — relevant for GSE/construction-zone ODD
City-Facade	MLS facade-level point clouds	Moderate — strong for terminal/building-frontage vertical structures, weak for ground, markings, movers, and open areas
Point Cloud City / Open3D-ML PCC	Indoor managed-building point clouds	Moderate — useful for terminal-interior, public-safety, small-object, and label-unification lessons; wrong outdoor geometry
SemanticKITTI (multi-scan)	MLS per-scan sequences + motion	Moderate — per-scan, but 4D labelling available
FRACTAL	ALS nadir, large-area	Moderate for site surveys; wrong viewpoint for ground-vehicle
DALES	ALS nadir	Low for ground-vehicle; nadir density profile
ZAHA	MLS facade-only point clouds	Moderate for terminal/frontage slices — strong LoFG facade hierarchy, weak for full-site map validation
GridNet-HD	UAV LiDAR + oblique imagery, utility corridor	Low-moderate for airside; strong for thin overhead/edge infrastructure and LiDAR-image fusion stress
CUS3D / SUM Parts / SensatUrban / STPLS3D	UAV photogrammetric or textured mesh	Low-moderate — excellent for RGB/mesh/tiling design; no LiDAR pulse noise
nuScenes / Waymo	Per-frame, no map accumulation	Low — per-scan paradigm only

Urban-district / non-road proxy matrix:

Target condition	Best public proxies	Use in the pipeline
Railway, taxiway, or service-road corridor	SemanticRail3D, WHU-Railway3D	Stress linear-infrastructure classes, overhead/edge structures, corridor tiling, and constrained-route geometry
Dense urban district, campus, or depot map	WHU-Urban3D, KITTI-360, Paris-Lille-3D, Toronto-3D, SemanticTHAB, MLDAS, USCILab3D	Pre-train MLS backbones and validate markings, poles, wires, road/driveway, building, and low-vegetation confusion; use MLDAS for sensor-transfer stress and USCILab3D only after release-maturity checks
Managed-building, terminal-interior, or public-safety facility map	Point Cloud City / Open3D-ML PCC, S3DIS, ScanNet200, USCILab3D	Stress indoor/managed-site label harmonization, emergency-response safety features, small minority classes, and point-cloud-to-map annotation workflows; do not transfer outdoor dynamics from these sources
Building frontage, facade, or vertical-structure-heavy district	ZAHA, City-Facade, SUM Parts, CUS3D, WHU-Urban3D, Toronto-3D	Validate LoFG-style facade hierarchy, wall/window/door/roof/balcony/HVAC-like sublabels, facade continuity, vertical tiling, BIM/digital-twin handoff, and terminal-frontage QA
Utility, trench, overhead-line, perimeter, or gantry infrastructure	GridNet-HD, ECLAIR, OpenTrench3D, DALES, Toronto-3D, WHU-Railway3D	Stress pylon/cable/insulator/pole/wire/trench classes, LiDAR-image projection, and rare long-thin recall; OpenTrench3D is photogrammetric-only
Construction, quarry, apron works, industrial plant, and large equipment	GOOSE-Ex, SIP, S.MID, Industrial3D, CUS3D, STPLS3D	Cover unstructured terrain, temporary equipment, substation/MEP clutter, works-zone geometry, and simulator/synthetic rare-class augmentation
Aerial/site-survey layer	FRACTAL, DALES, ECLAIR, YUTO, CITYLID, CUS3D	Train or validate nadir/site-survey products that complement the ground survey map
Mesh or digital-twin release product	SUM, SUM Parts, CUS3D, H3D	Validate point-to-mesh transfer, textured-map annotation, road-marking/cycle-lane/sidewalk surfaces, and facade/roof class splits

Three additions are worth singling out for airside work specifically:

• GOOSE-Ex is the most directly relevant new resource — built on odometry-merged aggregated clouds (same accumulation regime as an airside apron map) and covers unstructured large-vehicle scenes (excavator, quadruped) sharing more with apron/movement-area operations than any urban-road benchmark. It is the primary public pre-training source for the mining and construction ODDs and a reasonable off-road transfer source for airside.
• SemanticRail3D adds a large MLS corridor reference; railway-domain classes (catenary poles, rails, trackbed, masts, overhead lines) directly parallel airside ground-lighting arrays, taxiway edge structures, and jet-bridge infrastructure.
• GridNet-HD adds the strongest current LiDAR-image utility-infrastructure proxy. It does not match airside geometry, but its pylon/cable/insulator split is a useful stress test for whether a taxonomy and model can preserve long, thin, safety-relevant static infrastructure instead of merging it into generic pole or vegetation classes.
• FRACTAL is the strongest ALS pre-training source — a 250 km², statistically balanced, geometry-only corpus useful for nadir-scanned site-survey models, though its airborne viewpoint still differs from a ground survey-drive.
• Point Cloud City / Open3D-ML PCC adds the managed-building/public-safety proxy that was missing from this page. Its value is not outdoor transfer; it is label-unification, minority safety-feature behavior, and reproducible Open3D-ML training/format-conversion evidence for terminal interiors or facility maps.
• City-Facade adds an MLS facade proxy where the existing facade evidence was mostly mesh or photogrammetry. It is useful when the release map needs finer building-frontage semantics or digital-twin handoff, but it is not a substitute for full-site road/yard/apron validation.
• ZAHA adds the current largest public point-cloud facade benchmark and a standards-oriented LoFG hierarchy. It complements City-Facade: City-Facade is a city-road frontage benchmark with semantic and instance labels; ZAHA is a facade-generalization stress test with two nested class levels. Use both for terminal-frontage and vertical-structure taxonomy design, not as substitutes for site-wide map acceptance evidence.

Product-Mode Benchmark Coverage

Public proxy datasets support pre-training, stress testing, and architecture selection, but product acceptance still needs owned held-out maps. Use this matrix to decide which public bundles to pair with local evidence.

Product mode	Public proxy bundle	Owned evidence required	Rejection signal
Runtime semantic map	Paris-Lille-3D, Toronto-3D, KITTI-360, SemanticRail3D, WHU-Urban3D, DALES/FRACTAL for aerial layers	Same-site held-out tiles with source-map QA, class-balanced mIoU, seam disagreement, localization regression, and release-state confusion	High public mIoU but false deletion of markings, kerbs, poles, signs, or other localization-useful thin structure
Map-derived training export	SemanticKITTI/nuScenes/Waymo for scan-taxonomy compatibility; MLS proxies for map-to-scan density gap	Back-projection audit from owned maps to contributing scans, with permanent_static eligibility masks and scan-level split isolation	Pseudo-label batches contain movable-static, static-transient, artifact, FOD-candidate, or unknown-review points as positives
Map hygiene and change monitoring	MOS/moving-static benchmarks, dynamic-map-cleaning benchmarks, multi-session map-maintenance datasets where available	Repeated local surveys with reason-coded release-state labels for dynamic residual, movable-static, static-transient, FOD, artifact, and true infrastructure change	Cleaner preserves dynamic ghosts or deletes protected permanent thin structures under the same threshold
Digital twin / mesh / facade	City-Facade, ZAHA, Point Cloud City / Open3D-ML PCC, CUS3D, SUM/SUM Parts, STPLS3D	Point-to-mesh or point-to-surface transfer QA, visual review, asset-inventory checks, and facade/interior held-out labels	Facade or mesh labels look good visually but lack point-cloud provenance, source confidence, or consumer-specific class mapping
Local benchmark and acceptance set	Non-road bundle matched to the target slice: apron/depot, campus, port, utility, facade, terminal/interior, construction/industrial	Locked owned train/validation/test split with semantic labels, release-state labels, product-mode manifest, and reviewer/audit evidence	Public proxy score is used as acceptance evidence without matching local ODD, sensor, class, and hygiene-layer coverage

The full selection rationale and the proposed airside benchmark specification are in ../overview/aggregated-map-semantic-segmentation.md §5.3–§5.4.

Sources

• SemanticKITTI: Behley et al., ICCV 2019 — http://semantic-kitti.org · https://github.com/PRBonn/semantic-kitti-api · arXiv:1904.01416
• nuScenes-lidarseg: Caesar et al., arXiv:2109.03805 — https://github.com/nutonomy/nuscenes-devkit
• Waymo 3D Seg: waymo.com/open · 2024 challenge 1st place arXiv:2407.15282 · 2nd place arXiv:2501.05472
• Waymo Panoramic Panoptic (4DSeg): arXiv:2206.07704
• SAM4D: arXiv:2506.21547
• Semantic3D: Hackel et al., ISPRS 2017 — http://semantic3d.net · arXiv:1704.03847
• Paris-Lille-3D: Roynard et al., IJRR 2018 — https://npm3d.fr/paris-lille-3d
• Toronto-3D: Tan et al., CVPRW 2020 — arXiv:2003.08284
• KITTI-360: Liao et al., PAMI 2022 — https://www.cvlibs.net/datasets/kitti-360 · arXiv:2109.13410
• DALES: Varney et al., CVPRW 2020 — arXiv:2004.11985
• GridNet-HD: Carreaud et al., 2026 — https://arxiv.org/abs/2601.13052 · https://huggingface.co/datasets/heig-vd-geo/GridNet-HD · GridNet-HD page
• ECLAIR: Melekhov et al., CVPRW 2024 — https://openaccess.thecvf.com/content/CVPR2024W/USM/html/Melekhov_ECLAIR_A_High-Fidelity_Aerial_LiDAR_Dataset_for_Semantic_Segmentation_CVPRW_2024_paper.html · https://github.com/SharperShape/eclair-dataset
• YUTO Semantic: ISPRS GSW 2023 dataset card — https://huggingface.co/datasets/ausmlab/yuto-semantic · https://yutosemantic.ausmlab.com/
• S.MID / SFPNet: Wang et al., ECCV 2024 — https://github.com/Cavendish518/SFPNet · https://www.semanticindustry.top/dataset
• OpenTrench3D: Hansen et al., 2024 — https://arxiv.org/abs/2404.07711 · https://github.com/SimonBuusJensen/OpenTrench3D
• MLDAS: Chen et al., IJCAI 2024 — https://sychen320.github.io/projects/MLDAS/ · https://www.ijcai.org/proceedings/2024/0072.pdf
• USCILab3D: Lekkala et al., NeurIPS 2024 Datasets and Benchmarks — https://proceedings.neurips.cc/paper_files/paper/2024/hash/628433f240414517fd95164b4275f5cc-Abstract-Datasets_and_Benchmarks_Track.html · https://sites.google.com/usc.edu/uscilab3d/
• Industrial3D: Yin et al., 2026 — https://arxiv.org/abs/2603.28660 · https://github.com/pointcloudyc/Industrial3D
• SensatUrban: Hu et al., CVPR 2021 — arXiv:2201.04494 · https://github.com/QingyongHu/SensatUrban · https://point-cloud-analysis.cs.ox.ac.uk/
• H3D (Hessigheim): arXiv:2102.05346 · ISPRS Open Journal 2021
• STPLS3D: Chen et al., BMVC 2022 — arXiv:2203.09065
• CUS3D: Gao et al., Remote Sensing 2024 — https://www.mdpi.com/2072-4292/16/6/1079
• SUM Parts: Gao, Nan, and Ledoux, CVPR 2025 — https://arxiv.org/abs/2503.15300 · https://tudelft3d.github.io/SUMParts/
• S3DIS: Armeni et al., CVPR 2016 — http://buildingparser.stanford.edu/dataset.html · arXiv:1702.01105
• ScanNet v2: Dai et al., CVPR 2017 — http://www.scan-net.org · arXiv:1702.04405
• ScanNet200: Rozenberszki et al., ECCV 2022 — arXiv:2204.07761 · https://rozdavid.github.io/scannet200
• FRACTAL: Gaydon et al., 2024 — arXiv:2405.04634 · HuggingFace IGNF/FRACTAL
• GOOSE: Mortimer et al., 2023 — arXiv:2310.16788 · https://goose-dataset.de
• GOOSE-Ex: Hagmanns et al., 2024 — arXiv:2409.18788
• SemanticRail3D: Nature Scientific Data, Dec 2025 — https://www.nature.com/articles/s41597-025-06392-9 · https://github.com/Arshia-Gha/SemanticRail3D_Dataset
• WHU-Railway3D: IEEE TITS 2024 — https://dl.acm.org/doi/10.1109/TITS.2024.3469546 · https://github.com/WHU-USI3DV/WHU-Railway3D
• WHU-Urban3D: https://whu3d.com/dataset/ · https://whu3d.com/benchmark.html
• SemanticTHAB: https://zenodo.org/records/14906179 · https://github.com/kav-institute/SemanticLiDAR
• Point Cloud City / Open3D-ML PCC: https://www.nist.gov/services-resources/software/point-cloud-city-open3d-ml-repository · https://www.nist.gov/publications/cross-dataset-semantic-segmentation-performance-analysis-unifying-nist-point-cloud-city · https://github.com/alexdimopoulos/PointCloudCity-Open3D-ML
• City-Facade: Chen et al., ISPRS JPRS 2026 — https://doi.org/10.1016/j.isprsjprs.2026.01.003 · https://github.com/SYSU-3DSTAILab/City-Facade
• ZAHA: Wysocki et al., WACV 2025 — https://openaccess.thecvf.com/content/WACV2025/html/Wysocki_ZAHA_Introducing_the_Level_of_Facade_Generalization_and_the_Large-Scale_WACV_2025_paper.html · https://github.com/OloOcki/zaha · https://tum2t.win/datasets/pc-mls · https://tum2t.win/benchmarks/pc-fac
• Turin3D: CVPR Workshops 2025 — arXiv:2504.05882
• CITYLID: Environment and Planning B (SAGE), 2025 — https://journals.sagepub.com/doi/full/10.1177/23998083241312273 · HuggingFace Deepank/CITYLID
• SIP: arXiv:2512.09062
• YUTO Semantic: https://www.ausmlab.com/datasets/dataset · https://huggingface.co/datasets/ausmlab/yuto-semantic
• Industrial3D: https://arxiv.org/abs/2603.28660 · https://github.com/pointcloudyc/Industrial3D
• ECLAIR: arXiv:2404.10699
• RELLIS-3D: ICRA 2021 — https://dl.acm.org/doi/10.1109/ICRA48506.2021.9561251
• SynLiDAR: arXiv:2107.05399 (AAAI 2022) · https://github.com/xiaoaoran/SynLiDAR
• RareBoost3D: arXiv:2510.10876
• SynthmanticLiDAR: arXiv:2501.19035
• KITTI-CARLA: ResearchGate 2021
• M3Net (union taxonomy): arXiv:2405.01538
• PPT (language-guided multi-dataset training): CVPR 2024 — https://openaccess.thecvf.com/content/CVPR2024/papers/Wu_Towards_Large-scale_3D_Representation_Learning_with_Multi-dataset_Point_Prompt_Training_CVPR_2024_paper.pdf
• OmniLiDAR / TerraSeg: arXiv:2603.27344
• Panoptic nuScenes: arXiv:2109.03805
• Local context: ../overview/aggregated-map-semantic-segmentation.md §5 (dataset selection guide), §13.3 (benchmarking rigor)