Please tell us a little about yourself. Access is granted automatically. We collect this information only to track usage statistics and to understand the research impact of HFX3D. Your details will be handled responsibly, kept confidential, and not shared with third parties or used for any other purpose.

HFX3D — Urban Facade Point-Cloud Dataset

Large-scale mobile-LiDAR dataset of urban building facades in Halifax, Canada, with semantic, instance, and functional-attribute annotations. Captured with a GeoSLAM ZEB Horizon handheld scanner + ZEB Vision RGB camera.

26 buildings · ~492 M points · building-level train / val / test (18 / 4 / 4)
12 semantic classes · per-instance 15-dim functional attributes · 6,549 instances
One self-describing HDF5 per building
Paper: https://arxiv.org/abs/2611.08645
Code / benchmark: https://github.com/jijup/hfx3d-benchmark
Project page: https://jijup.github.io/hfx3d-benchmark/

Files

train/ val/ test/   HFX_BLD0NN_ZEB_CLEAN.h5   (18 / 4 / 4)
MANIFEST.csv                      # per building: split, #points, #instances, size
ontology_class_to_attributes.json # which attributes can apply to each class
README.md

Schema (per building, `N` points, `M` instances)

Group / dataset	shape	dtype	meaning
`point_cloud/xyz`	(N, 3)	float32	raw scanner coordinates, metres (local SLAM frame, not georeferenced)
`point_cloud/rgb`	(N, 3)	uint8	colour 0–255 (raw, lossless)
`point_cloud/intensity`	(N,)	uint16	raw LiDAR intensity (not normalized)
`labels/semantic`	(N,)	uint8	class id 0–11
`labels/instance`	(N,)	int32	instance id (≥1)
`functional/instance_id`	(M,)	int32	join key to `labels/instance`
`functional/semantic_id`	(M,)	uint8	instance class id
`functional/attributes`	(M, 15)	uint8	binary functional vector (human-validated)
`metadata/*`	—	—	class/attr names, counts, centroid, units

Semantic classes (`labels/semantic`, 12)

id	class	id	class
0	wall	6	terrain
1	window	7	roof
2	door	8	blinds
3	balcony	9	other
4	vegetation	10	column
5	stairs	11	arch

Functional attributes (`functional/attributes` column order, 15)

col	attribute	col	attribute
0	load_bearing	8	drainage
1	thermal_envelope	9	fall_protection
2	vegetation_support	10	aesthetic
3	operable	11	privacy_screening
4	solar_shading	12	circulation
5	ventilation	13	illumination
6	natural_lighting	14	surveillance
7	access

Attribute applicability per class is in ontology_class_to_attributes.json.

Usage

import h5py, numpy as np
with h5py.File("train/HFX_BLD001_ZEB_CLEAN.h5", "r") as f:
    xyz   = f["point_cloud/xyz"][:]        # (N,3) metres (raw scanner frame)
    rgb   = f["point_cloud/rgb"][:]        # (N,3) uint8
    inten = f["point_cloud/intensity"][:]  # (N,)  uint16 raw
    sem   = f["labels/semantic"][:]        # (N,)  0..11
    inst  = f["labels/instance"][:]        # (N,)  >=1
    fid   = f["functional/instance_id"][:] # (M,)
    attr  = f["functional/attributes"][:]  # (M,15) 0/1
# per-point attribute lookup: point -> instance -> 15-dim vector
row = {int(i): k for k, i in enumerate(fid)}
point_attr = attr[[row[int(i)] for i in inst]]   # (N,15)

Data is shipped raw. Typical preprocessing you apply yourself:

xyz — mean-center per sample, or subtract metadata/centroid; keep metric scale.
rgb — rgb / 255.0.
intensity — scale to [0, 1] (e.g. clip to train 1st/99th percentiles, then min-max).

Paper input is the 7-dim feature [x, y, z, r, g, b, intensity].

Tasks & splits

Semantic segmentation, instance segmentation, functional-attribute prediction. Splits are building-level (no leakage); see MANIFEST.csv. A full Datasheet (DATASHEET.md) accompanies the dataset.

Citation

@inproceedings{hfx3d2027,
  title     = {From Geometry to Function: A Multi-Task {LiDAR} Benchmark Dataset
               for Semantic and Functional Parsing of Urban Facades},
  author    = {Muse, Abdiaziz and Javed, Tooba and Pu, Tao and
               Poovvancheri, Jiju and Chen, Dong},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications
               of Computer Vision (WACV)},
  year      = {2027},
}

Downloads last month: 26

Total file size:

4.51 GB

Paper for SMU-GSCLab/hfx3d-benchmark

From Geometry to Function: A Multi-Task LiDAR Benchmark Dataset for Semantic and Functional Parsing of Urban Facades

Paper • 2611.08645 • Published 5 days ago • 1

Request access to HFX3D