[Conference TBD]

HFX3D: A LiDAR Benchmark for
Semantic Segmentation of
Urban Building Facades

A large-scale LiDAR facade dataset annotated for semantic, instance, and functional understanding of real urban buildings.

Saint Mary's University, Halifax, Nova Scotia

Paper Dataset Code 3D Explorer

Buildings

Classes

492M

Points

Neighbourhoods

Seasons

Abstract

A New Benchmark for Facade Segmentation

We introduce HFX3D, a large-scale LiDAR benchmark for 3D semantic segmentation of urban building facades in Halifax, Nova Scotia, Canada. The dataset comprises 26 buildings captured with a ZEB Horizon handheld scanner, spanning commercial, residential, institutional, and waterfront contexts with facades ranging from historical masonry to contemporary glass curtain walls, annotated across 12 semantic classes totalling over 492 million points. Each building is enriched with semantic, instance, and functional labels, making HFX3D the first facade benchmark to provide all three annotation tiers.

Building facade segmentation is fundamental to urban digital twins, architectural surveying, heritage preservation, and infrastructure inspection, yet existing benchmarks predominantly target autonomous driving or indoor environments. We benchmark seven leading 3D segmentation architectures: KPConv, PTv3, PTv2, DGCNN, RandLA-Net, PointNet++, and PointNet, establishing strong baselines and revealing challenges unique to close range architectural LiDAR capture across five urban neighbourhoods, four seasons, and varied lighting conditions.

Dataset

Data Collection and Annotation

Split

18 train · 4 val · 4 test

Scanner

ZEB Horizon

Features

XYZ · RGB · Int · GPS

Dataset Statistics

Class Balance

Points & Classes per Building

Per-Class Split

Building × Class Heatmap

RGB & Semantic Annotation

BLD006 · VIC Suites

RGB

Semantic

BLD017 · 1801 Hollis (Sidewalk)

RGB

Semantic

BLD022 · Steele Ocean 1

RGB

Semantic

BLD024 · Sobeys Hub

RGB

Semantic

12-Class Taxonomy

Environmental Diversity

Spring

Summer

Fall

Winter

Daytime

Night

ZEB Horizon Scanner View

Winter

Fall

Summer

Urban

Dataset Comparison

Pts = total points (M) · Cls = semantic classes · Inst. = instance labels · Func. = functional attributes · Seas = seasons captured

Dataset	Year	Sensor	Scenes	Pts (M)	Cls	Inst.	Func.	Seas	Facade
General outdoor / indoor benchmarks
S3DIS	2016	RGB-D	272 rooms	696	13	✓	N/A	N/A	N/A
Semantic3D	2017	TLS	~30	4,000	8	N/A	N/A	1	N/A
Paris-Lille-3D	2018	MLS	2 km	143	50	N/A	N/A	1	N/A
Toronto-3D	2020	MLS	1 km	78	8	N/A	N/A	1	N/A
SensatUrban	2021	UAV	3 cities	3,000	13	N/A	N/A	1	N/A
KITTI-360	2022	MLS	73 km	N/A	13	✓	N/A	1	N/A
Façade-specific benchmarks
ArCH	2020	TLS	17 bldgs	102	10	N/A	N/A	N/A	✓
TUM-FAÇADE	2022	MLS	14 bldgs	118	17	N/A	N/A	1	✓
ZAHA	2024	MLS	66 bldgs	601	15	N/A	N/A	1	✓
City-Facade	2026	MLS	N/A	200	9	✓	N/A	N/A	✓
HFX3D (ours)	2026	Handheld	26 bldgs	492	12	✓	✓	4	✓

Interactive 3D Explorer

Explore the Dataset in 3D

Select a building to explore its point cloud · RGB and semantic side by side

Halifax Lib.

BLD009

Goldberg

BLD001

Mona Camp. A

BLD002

Mona Camp. B

BLD003

Kenneth Rowe

BLD004

Schulich Law

BLD005

VIC Suites

BLD006

Engineering

BLD007

Mona Camp. C

BLD008

Emera Place

BLD010

Gerard Hall

BLD011

The Margarita

BLD012

1801 Hollis #1

BLD013

Summit Place

BLD014

1801 Hollis #2

BLD015

1801 Hollis #3

BLD016

1801 Hollis #4

BLD017

The Exchange

BLD018

Sexton Camp. 1

BLD019

Sexton Camp. 2

BLD020

TD Building

BLD021

Steele Ocean 1

BLD022

Steele Ocean 2

BLD023

Sobeys Hub

BLD024

Belmont House

BLD025

The Avery

BLD026

RGB

Semantic

Loading point cloud…

PLY preview not found Generate: python tools/export_ply.py --all

Loading…

Benchmark Results

Baseline Performance

Overall Metrics

#	Model	OA	μP	μR	μF1	mIoU ↓
1	KPConv	82.2%	49.7%	42.4%	43.2%	34.08%
2	PointNet++ (SSG)	78.3%	39.1%	34.3%	34.8%	28.47%
3	PointNet	74.4%	39.0%	31.9%	32.1%	25.34%
4	PTv2 (Point Transformer V2)	78.8%	24.1%	24.7%	23.6%	19.90%
5	PTv3 (Point Transformer V3)	69.5%	28.2%	20.9%	19.4%	15.79%
6	DGCNN	69.3%	31.3%	16.6%	15.7%	13.04%
7	RandLA-Net	41.5%	9.9%	10.8%	9.0%	5.91%

Per-Class IoU

Bold = best per class

Model	Wall	Win.	Door	Balc.	Veg.	Stairs	Terr.	Roof	Blinds	Other	Col.	mIoU
KPConv	77.5	28.5	0	19.6	91.2	9.7	87.4	6.6	41.0	23.5	23.9	34.08%
PointNet++ (SSG)	72.7	21.8	0	0	83.8	56.6	82.5	0.04	8.4	15.1	0.7	28.47%
PointNet	67.9	16.8	0.2	0	81.0	40.5	72.4	0.4	2.8	20.3	1.8	25.34%
PTv2	75.4	16.8	0	0	62.5	0	84.1	0	0	0	0	19.90%
PTv3	73.5	11.5	0	8.9	0.1	9.0	85.5	0	0	1.0	0	15.79%
DGCNN	61.6	4.3	0	0	6.9	0	83.8	0	0	0	0	13.04%
RandLA-Net	41.2	6.7	0	0	0	0	23.0	0	0	0	0	5.91%

Acknowledgments

Support & Resources

Saint Mary's University Faculty of Science

This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) and Canada Foundation for Innovation (CFI) . Computational resources were provided by the Digital Research Alliance of Canada. We gratefully acknowledge Saint Mary's University for institutional support. We thank the building owners and facility managers across Halifax for generously granting scanning access to their properties. The quality of the HFX3D annotations is owed to the dedication of our collection, curation and labelling team: Tooba Javed, Prachi Kudeshia, Jack Greenlaw, Avery Cao, and Khushal Das.

Citation

BibTeX

If you use HFX3D in your research, please cite:

BibTeX

@inproceedings{hfx3d,
  title     = {{HFX3D}: A {LiDAR} Benchmark for Semantic Segmentation
               of Urban Building Facades},
  author    = {[Authors TBD]},
  booktitle = {[Conference]},
  year      = {[Year]},
}

HFX3D: A LiDAR Benchmark for
Semantic Segmentation of
Urban Building Facades

A New Benchmark for Facade Segmentation

Data Collection and Annotation

Explore the Dataset in 3D

Baseline Performance

Overall Metrics

Per-Class IoU

Predictions vs. Ground Truth

Halifax, Nova Scotia

Download HFX3D

Support & Resources

BibTeX

Who's Exploring HFX3D?

HFX3D: A LiDAR Benchmark for Semantic Segmentation of Urban Building Facades

A New Benchmark for Facade Segmentation

Data Collection and Annotation

Explore the Dataset in 3D

Baseline Performance

Overall Metrics

Per-Class IoU

Predictions vs. Ground Truth

Halifax, Nova Scotia

Download HFX3D

Support & Resources

BibTeX

Who's Exploring HFX3D?

HFX3D: A LiDAR Benchmark for
Semantic Segmentation of
Urban Building Facades