Towards detailed building typology by
urban-scale 3D building decomposition
❌

1
2
1
Yijie Wu , Liangliang Nan , Fan Xue

1The University of Hong Kong, 2Delft University of Technology

Contact: yijiewu@connect.hku.hk, liangliang.nan@tudelft.nl, xuef@hku.hk

1. Introduction

• Urban-scale building shape analytics is crucial for identifying building
functions, analyzing urban morphology, and understanding how
building shapes and their distribution affect microclimate and energy
consumption.
• However, both traditional building shape analytics based on 2D
footprints and exploratory studies on 3D building models describe
building shape as a whole, i.e., a polygon in 2D or a polyhedron in 3D.
It fails to capture the shape primitives composing a building and
obscures the building massing customized for functionality and
performance.

0011.
Building as
parts?

4. Result gallery

5. Urban-scale experiment

2

Experiments have been conducted on
more than 8,000 buildings in Hong
Kong.

src: Ubuy NL

src: Wikipedia-Bank of China Tower

• We transfer the classic task of shape decomposition from the field of
computer graphics into urban analytics to address this limitation.
• Building shape analytics then separates a building into smaller primitive
parts. Adjacency between parts derives a graph for shape
characterization. We then can use this graph to identify and cluster the
shape typology of buildings.

Our method successfully detected the typical 8-wing buildings of Hong with
precision and recall above 90 %.

2. Data

Advances in reality capture technologies, such as LiDAR scanning and
photogrammetry, have significantly improved the accessibility and detail
of 3D building models. Recent developments in 3D reconstruction,
including neural radiance field (NeRF) and Gaussian splatting, further
broaden the potential for the widespread adoption of 3D building models
on urban and global scales. 3D building
0011. models are expected to become
a fundamental data source for a broad
range
of
3D
urban
analytics.
Data
2
used in this study include:

Reconstructed lightweight models

The visualization of multidimensional scaling on pairwise similarity between
the decomposition graphs of each building displays a distinct layering of the
complexity of building typology.

Manual CAD models

(from point clouds or photogrammetry meshes)

3. Method: BuildingCut
✂

✂

✂

✂

Atomic 2 steps to
decompose a shape

✂

2. Decompose the part by which plane?

1. Which part to decompose?

Framework of BuildingCut

0011.
2

Mul$ple cu)ng modes
Concave edges

In/out points

Sampled points

✂

✂
✂

In/out point sampling

Manifold repairing [op/onal]

Cuts
Vote in/out
by ray intersec0on

Itera1ve carving
Plane
Planedetec*on
detec/on

Cut to increase
1) in points in in cells; 2) out points in out cells
cut #1

Input

cut #4
…

cut #11
…

cut #30
…

Ini$al
convex cells

Mul$ple objec$ves

Objec*ves

❌
Surface cuts

Bisector cuts

❌
✅

Orthogonal cuts

❌

6. Conclusion

Ver0cal cuts

Our pilot results validate the feasibility of using 3D building decomposition for
more detailed building typology. The experiments also demonstrate a strong
potential for scaling our method to urban-scale 3D building shape analytics to
empower future applications with rapidly expanding 3D building models.

Fewer-part

Maximal-volume

Centralizing

Decentralizing

Tree search

Stage 1: Preprocessing

Acknowledgement

Stage 2: Decomposi:on framework

Funding: RGC, ECS, ITF@HKSAR, NSF@Guangdong; Compu+ng: HPC@HKU

References

Output

Wu et al. 2024 A novel Building Section Skeleton for compact 3D reconstruction from point clouds:
A study of high-density urban scenes (ISPRS J)
Wei et al. 2022 Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search (TOG)
Xue et al. 2019 A derivative-free optimization-based approach for detecting architectural symmetries from 3D point clouds (ISPRS J)
Nan & Wonka. 2017 Polyfit: Polygonal surface reconstruction from point clouds. (ICCV)

Will be Open! Our method and test samples will be released at https://github.com/eiiijiiiy/ after paper submission (Perhaps by Dec. 31 2024!)

首屆智慧遙感製圖國際研討會
The 1st Interna,onal Workshop on
Remote Sensing Intelligent Mapping (RSIM)