BSS-Indoor: Volumetric wall reconstruction using

1

Building Section Skeleton

2

Yijie Wu 1, Siyuan Meng 2, Fan Xue 3

3

This is the authors' version of the paper:
Wu, Y., Meng, S., & Xue, F. (2025). BSS-Indoor: Volumetric wall reconstruction using Building
Section Skeleton. Proceedings of the 29th International Symposium on Advancement of
Construction Management and Real Estate, Springer, in press.
This file is shared for personal and academic use only, under the license CC BY-NC-ND 4.0
(Non-Commercial, No Derivatives, and with an Attributed citation when you use). The final
published version of this paper can be found at: [LINK_TO_SPRINGERLINK]. Any uses other
than personal and academic purposes must obtain appropriate permissions from Springer first.

15

Abstract. Building Information Modeling (BIM) has become well recognized and increasingly mandated
in the building sector, where Scan-to-BIM automation shows great potential for BIM users. For the ‘last
mile’ or ‘last step’ of automatic Scan-to-BIM, parametric reconstruction, existing methods focus on
surface reconstruction. However, BIM practitioners have to handle architectural elements as volumetric
instances rather than connected surfaces. This paper proposes a volumetric reconstruction method called
BSS-Indoor. The method is based on a novel Building Section Skeleton (BSS) theory which pairs the
parallel and symmetric facades of the building exterior. By adapting BSS from building exteriors to
interiors, BSS-Indoor reconstructs internal walls as BSS segments. Our method simultaneously locates
walls and estimates their thicknesses, addressing common issues in existing solutions, such as fragmented
long walls and merged adjacent walls. Preliminary experiments on a subset of the CV4AEC (3D) dataset
demonstrate the effectiveness and potential of BSS-Indoor in reconstructing volumetric walls.

16

Keywords: Scan-to-BIM; Volumetric wall reconstruction; Point cloud processing

4
5
6
7
8
9
10
11
12
13
14

1 Yijie Wu

Department of Real Estate and Construction, The University of Hong Kong, China
Email: yijiewu@connect.hku.hk
2 Siyuan Meng
Department of Real Estate and Construction, The University of Hong Kong, China
Email: siyuanm@connect.hku.hk
3 Fan Xue
Department of Real Estate and Construction, The University of Hong Kong, China
Email: xuef@hku.hk

1

17

1

Introduction

18

The Building Information Modeling (BIM) has become significantly important throughout a building’s

19

entire lifespan, covering all phases from initial design to eventual demolition [1]. To streamline the

20

building sector with BIM, automating the creation of BIMs is extremely crucial [2]. Scan-to-BIM, a

21

popular task that converts 3D scans captured by laser scanners or photogrammetry, shows great potential

22

in the creation and uses of as-is BIM [3]. A typical Scan-to-BIM workflow involves four steps: scanning,

23

scan registration, semantic segmentation, and parametric reconstruction. In recent years, solutions and

24

datasets for registration and semantic segmentation of building interiors have rapidly evolving, which

25

significantly boosted the Scan-to-BIM popularity [4]. However, the parametric reconstruction, which is

26

the ‘last mile’ or ‘last step’ of the Scan-to-BIM pipeline, does not yet fulfill the actual requirements of

27

BIM practitioners in the industry [5].

28

Existing methods for reconstructing building interiors focus on surface reconstruction instead of vol-

29

umetric modeling. These methods reconstruct the surfaces of architectural elements, i.e., walls, floors,

30

and ceilings, into lightweight meshes [6, 7]. Simultaneously, rooms and their closed boundaries are

31

extracted [8, 9]. The resulting surface models or floorplans satisfy the requirements of indoor naviga-

32

tion [10] and fit the native geometric representation of CityGML for City Information Modeling [6].

33

However, BIM practitioners in the building sector pay more attention on the individual instances of ar-

34

chitectural elements [5]. Walls and slabs are preferably modeled as volumetric instances enriched with

35

joints and details inside the surfaces rather than separated facets that bound different rooms.

37

Volumetric reconstruction of architectural elements essentially involves locating the volume, outlines
the volumetric shape and/or estimates the dimensions of the identified volume. For walls and slabs,

38

thickness should be estimated. Meanwhile, walls are usually simplified as cuboids. Volumetric walls can

39

be reconstructed by clustering the wall points produced by semantic segmentation and then estimating

40

the bounding box of each cluster [4]. To achieve a higher reconstruction fitness to the scans, optimiza-

41

tion models were formulated to finetune the locations and dimensions of wall boxes [11]. Moreover,

42

wall thickness can also be predicted by deep learning after projected the point clouds onto horizontal

43

planes [5]. However, as the height information is missing after the projection, the method reconstructs

44

floorplans only. A common aspect among these solutions is that they separate the volume localization and

45

dimension estimation processes. As a result, long walls are prone to being cut into pieces and adjacent

46

walls with different dimensions are prone to being merged into one wall.

36

47

In this paper, we propose a volumetric reconstruction method for walls based on the Building Section

48

Skeleton (BSS) [12]. Our method aims to locate and estimate the thickness of walls simultaneously. BSS

49

is a shape skeleton defined to adapt the morphological characteristics of building exterior boundaries.

50

Each BSS segment corresponds to a pair of parallel vertical planes representing facades or a pair of

51

symmetric incline planes representing roofs. A BSS segment is parameterized with the radius and forms

52

a polyhedron, which is a similar concept of volume thickness. This study transfers the BSS for building

53

exteriors to interiors. Each internal walls are reconstructed as a BSS segment, while the external walls

54

are detected as the parallel facades that bound buildings. Preliminary experiments on the a subset of the

55

CV4AEC (3D) dataset [13] show that our method, named BSS-Indoor, can effectively reconstruct the

56

volumetric walls.

57
58

The remainder of this paper will introduce our method BSS-Indoor, report preliminary results, as
well as discuss the strengths and future improvement of BSS-Indoor.

2

59

2

Preliminaries: Building Section Skeleton for wall reconstruction

60

The fundamental concept of the Building Section Skeleton is to capture the parallel and symmetric planes

61

of buildings, building on the classical concept, shape skeleton, in computer graphics. The identified par-

62

allel and symmetric planes can be assembled into 3D volumes which represent the space inside buildings,

63

providing semantic information. These 3D volumes may corresponding to the empty space of building

64

interiors designated for specific functions, such as offices and corridors, or they could also be physical

65

structures, such as walls.

66

BSS are defined at three levels, i.e., atoms, segments, and the relations between segments. A BSS

67

atom is the center of the sphere inscribed by symmetric points on the exterior or interior surfaces of

68

buildings. If the surfaces are vertical, the radius of the sphere and the normals of the inscribed points

70

are attributes of a BSS atom. If the surfaces are not vertical, the intersection height of the two inscribed
directions and the normals of inscribed points serve as the parameters of a BSS atom. When two points

71

lie on a pair of parallel planes, then the corresponding BSS atom is equivalent to the midpoint of these

72

two points. At the second level in the definition of BSS, a segment is a group of BSS atoms distributed

73

on a same plane, forming a region, and sharing same radius or intersection height and normal. Hence, a

74

BSS segment represents the medial plane of a pair of parallel or symmetric planar regions. Furthermore,

75

by sweeping the medial vertical region of a BSS segment, there is a corresponding volume of a BSS seg-

76

ment, which represents the empty or physical space of buildings. The relations of BSS segments, which

77

constitute the third level of the BSS definition, refer to their topological relations, including intersections

78

and adjacency. For more details on BSS, please refer to [12].

69

80

In the practice application of BSS in Scan-to-BIM, there are two specific designs for wall reconstruction. Although a BSS segment could be the medial planar region of both vertical and incline planes which

81

could be parallel and non-parallel, respectively, the planes involved in wall reconstruction are usually par-

82

allel. Therefore, we solely use BSS for vertical and parallel planes. Moreover, different mechanisms are

83

employed reconstructing interior and exterior walls. As a BSS segment represents the medial plane of

84

a pair of parallel or symmetric planes, the use of BSS assumes the presence of dual planes to bound a

85

space. However, the scanned point clouds for BIM reconstruction may not always include the facades

86

of buildings, leaving only one single plane for an exterior wall. To reconstruct both interior and exterior

87

walls, our method captures dual wall surfaces as close as possible to each other for reconstructing BSS

88

segments for interior walls, and as far apart as possible for reconstructing the interior spaces of buildings.

89

Subsequently, we use the boundaries of the interior spaces to supplement the exterior walls.

90

3

Method

91

3.1

Overview

92

Fig. 1 shows an overview of our method. BSS-Indoor takes the segmented point clouds of architectural

93

elements as inputs and outputs the volumetric walls. Plane detection is performed first to group points as

94

planar primitives. Then, the method clusters the primitives in terms of their normals. In the case shown

95

in Fig. 1, two orthogonal orientations are clustered. After that, two stages, one for proposing, the other

96

for refining, are implemented to reconstruct the walls of each orientations.

79

97

Stage 1 locates the possible walls and outlines their vertical planar boundaries as the initial BSS

98

segments. This stage pairs the parallel planar primitives of each orientation whose distances between

99

each other are close or far apart enough in terms of given thresholds. After pairing, the proposed method

3

Stage 2: Alignment
Intersections

Planes to pair

Initial BSS segments
Internal

Projecting and Triangulating

Orientation #2

Plane fitting

Clustering

Planar groups

Intersecting

Orientation #1

Input:
Semantic points

between different orientations

Aligned BSS segments

External

Internal

Aligned polygons

Triangulation

Integer programming

Stage 1: Pairing and segmentation

Output: Volumetric walls

External

Figure 1: Overview of BSS-Indoor.

100

segments the common regions coarsely of each paired regions in an ascending order for close ones and

101

descending order for far away ones. Note that the common region is the vertical medial region of the

102

paired planes. However, as shown in Fig.1, the initial BSS segments have irregular boundaries. There

103

are also topological errors including intersections and gaps between different BSS segments.

104

Therefore, Stage 2 further refines the BSS segments by aligning the BSS segments of different orien-

105

tations to smooth the boundaries and remove the intersections and gaps. The aligning procedure projects

106

the vertical regions of a BSS segment onto a vertical plane. The method then triangulates the projected

107

extent. Triangles in the extent are labeled as inside or outside the BSS segment. Refining the boundary

108

of a BSS segments now can be conducted by flipping the inside/outside labels of the triangles. An in-

109

teger programming model is formulated to optimize the flipping for a smooth and topologically correct

110

boundary of each BSS segment.

111

After the two stages, internal walls are reconstructed from the BSS segments formed by close paired

112

planes, while external walls are from BSS segments composed by far apart parallel planes. The method

113

extrudes the vertical region of a BSS segment horizontally by the radius of a BSS segment for interior

114

walls or a standard thickness for exterior walls.

115

3.2

116

In Stage 1, parallel planes within each cluster are paired, and the common regions of the paired planes

117

are segmented for each orientation individually.

Stage 1: Pairing and Segmentation

118

Planes within a cluster that are either closer or farther than given thresholds are filtered as potential

119

pairs. Then, the method sorts the pairs with small distances in an ascending order for reconstructing inte-

120

rior walls, while pairs with large distances in descending order for exterior walls. The segmentation of the

121

common regions for interior and exterior walls are performed individually according to their respective

122

orders. Although a plane can be paired with multiple planes, a region of the plane can only be segmented

123

once and paired with one of its pairing candidates. Therefore, within a single orientation cluster, BSS

124

segments and their corresponding volumes cannot intersect with each other. By design, paired planes

4

125

with closer distances are segmented first for interior walls; whereas farther paired planes are prioritized

126

for exterior walls.

127

The segmentation of common regions are based on the Boolean operations of polygons. Our method

128

projects the points of each plane onto a vertical plane and then polygonizes the projected regions by

129

calculating their 2D Alpha shapes. When segmenting the common region for a pair of planes (A, B), the

130

two polygons of each plane are separated into three parts: A ∩ B, A \ B, and B \ A. The region of A ∩ B

131

is segmented out as one of the initial BSS segment. The regions of A \ B and B \ A are retained for the

132

subsequent segmentation with the remaining paired planes of A and B, respectively.

133

After segmenting the region of a BSS segment, our method extrudes the vertical region of a BSS

134

segment horizontally in both orientations by its radius to form a 3D polyhedron, representing the initial

135

interior walls and indoor spaces (presented in the two sub-figures on the left-hand side in the second row

136

in Fig. 1).

137

3.3

138

Stage 2 refines the initial BSS segments by aligning those of different orientations. This process includes

139

smoothing the boundaries of the BSS segments and eliminating intersections and gaps between them.

Stage 2: Alignment

140

The alignment procedure relies on the intersections of BSS segments from various orientations. The

141

method first intersects the medial planes of the initial BSS segments, filtering out intersection lines that

142

are too distant from the segmented regions. Next, for each orientation cluster, our method projects all

143

the segmented common regions from Stage 1, along with the intersection lines, onto the vertical plane

144

orthogonal to the orientation. Constrained triangulation is performed on the projection. The resulting

145

triangles are labeled based on whether they belong to a particular segment or not. Assume there are n

146

segments in a cluster; each triangle is labeled with a k-dimensional binary vector. The refinement of

147

initial BSS segments is then transformed into a problem of flipping the labels of these triangles.

148

A binary optimization model is formulated for the label flipping. The Boolean variables indicate

149

whether a label should be flipped, or equivalently, a triangle should be removed from or added to a BSS

150

segment. The smoothing and topology refinement are achieved by controlling the label distinctions of

151

triangles distributed across the intersection lines. Adjacent triangles sharing edges on the intersection

152

lines are prone to be labeled differently; that is, if one belong to a BSS segment, the other should not

153

belong to the same BSS segment. For adjacent triangles whose common edges are not on the intersection

154

lines, the optimization model tends to label the two triangles in a same way, i.e., both either belong to

155

or do not belong to a BSS segment. As a result, the optimization encourages continuity inside a BSS

156

segment and discontinuity along the intersection lines, which drives the boundaries of the BSS segments

157

to align with the intersection lines. Additionally, the refinement should not deviate significantly from the

158

initial BSS segments. Therefore, the optimization model also minimizes the number of label flipping.

159

4

Preliminary experiments

160

4.1

Data and implementation details

161

To validate the initial effectiveness of BSS-Indoor, six samples from the training set of the CV4AEC 3D

162

dataset were tested. Note that this dataset was captured by laser scanners. As a result, there are missing

163

points for walls and outliers outside the buildings due to windows. The six samples include two typical

164

scenarios: office and parking. For the following experiments, the point clouds were uniformly sampled
5

165

at an interval of 2cm.

166

The semantic point clouds of architectural elements were segmented by PointContrast [14] which

167

were trained on ScanNet [15] and S3DIS [16]. Plane primitives of the semantic point clouds were de-

168

tected by RANSAC [17]. We employed K-Means for clustering the plane orientations, where users need

169

to provide the number of clusters k in advance. In Stage 1, the Alpha shape polygonization of the plane

170

primitives was implemented based on GEOS [18]. The hyper-parameter α was set to a range from 1 to

171

3 m. In Stage 2, the constrained triangulation was implemented based on CGAL [19]. The optimization

172

of integer programming was solved by Gurobi [20] with a weight factor provided by users to balance the

173

geometric consistency and the degree of alignment.

174

4.2

175

Results were evaluated by the official 3D evaluation code [21] of CV4AEC’s Scan-to-BIM competition.

176

The evaluation metrics include 3D Intersection-over-Union (IoU) and the matching precision, recall, and

177

F1-score of the 2D wall segments. Note that the matching of the 2D segments were evaluated at three

178

different resolutions with matching thresholds of 5, 10, and 20 cm, respectively.

179

4.3

180

The reconstructed volumetric walls and the evaluation results are presented in Fig. 2 and Tab. 1. The

181

average 3D IoU is 20.4 %, with a range of 7.4 % to 28.7 %. At the 10-cm resolution, the average

182

precision, recall, and F1 scores are 41.4 %, 41.2 %, and 40.6 %. The precision and recall show a good

183

balance. The F1-scores at 10-cm resolution range from 21.2 % to 56.1 %. As shown in Fig. 2, some

184

internal walls and their thicknesses are accurately estimated, as BSS-Indoor captured their dual planes

185

correctly. The external walls were also partially reconstructed. However, there are still some noticeable

Evaluation

Results

limitations, particularly for missing walls.
Table 1: Evaluated results of the 6 samples in CV4AEC 3D dataset. #gt, #pred, #m refer to the
number of ground truth, prediction (reconstruction), and matched walls, respectively. The matching
between 2D wall segments were evaluated with three matching thresholds of 5, 10, and 20 cm.
thresh: 5 cm

thresh: 10 cm

thresh: 20 cm

sample

#gt

#pred

3D IoU

#m

prec.

rec.

F1

#m

prec.

rec.

F1

#m

prec.

rec.

F1

06 B1
06 F3
07 F5
32 F1
32 F2
32 F3

496
976
904
528
656
688

712
856
544
640
688
728

7.4%
7.5%
21.0%
31.3%
26.3%
28.7%

68
169
198
214
249
301

9.6%
19.7%
36.4%
33.4%
36.2%
41.3%

13.7%
17.3%
21.9%
40.5%
38.0%
43.8%

11.3%
18.4%
27.3%
36.6%
37.1%
42.5%

128
217
290
311
333
397

18.0%
25.4%
53.3%
48.6%
48.4%
54.5%

25.8%
22.2%
32.1%
58.9%
50.8%
57.7%

21.2%
23.7%
40.1%
53.3%
49.6%
56.1%

172
243
368
340
376
433

24.2%
28.4%
67.6%
53.1%
54.7%
59.5%

34.7%
24.9%
40.7%
64.4%
57.3%
62.9%

28.5%
26.5%
50.8%
58.2%
56.0%
61.2%

Average

/

/

20.4%

/

29.4%

29.2%

28.9%

/

41.4%

41.2%

40.6%

/

47.9%

47.5%

46.9%

186

187

The current BSS-Indoor implementation failed to reconstruct some external walls. As shown in

188

Fig .2, such failures are more significant for Samples 06-B1, 06-F3, and 32-F3. This issue occurred

189

because there might be large empty regions on the external walls where there are large windows. As a

190

result, holes could be observed in the original scans, semantic point clouds, and the initial alpha polygons

191

of the plane primitives, making it difficult to reconstruct the external walls.

192

Significant incompleteness in the semantic point clouds and initial alpha polygons also led to the

193

missing internal walls. These failures could be common due to the heavy occlusions caused by clutter

194

near the walls.

6

06_B1

06_F3

07_F5

F1 = 11.3 %@5 cm, 21.2%@1 cm, 28.5%@20 cm

F1 = 18.4 %@5 cm, 23.7 %@1 cm, 26.5%@20 cm

F1 = 27.3 %@5 cm, 40.1%@1 cm, 50.8%@20 cm

32_F1

32_F2

32_F3

F1 = 36.6 %@5 cm, 53.3%@1 cm, 58.2%@20 cm

F1 = 37.1 %@5 cm, 49.6%@1 cm, 56.0%@20 cm

F1 = 42.5%@5 cm, 56.1%@1 cm, 61.2%@20 cm

Figure 2: Results of the 6 samples. The right sub-figure for each case displays the matching between
the ground truth (in blue) and reconstruction (in red).

195

5

Discussion

196

We believe that BSS-Indoor has significant potential in calling for research interest on the volumes of 3D

197

objects above points and triangles. Our method separates the reconstruction of the dual-plane walls (usu-

198

ally internal walls) and single-plane walls (usually external walls), offering higher flexibility in thickness

199

estimation. By distinguishing these two situations, the thickness of a dual-plane wall is calculated as

200

the distance between its paired planes, while the thickness of a single-plane wall can be estimated or as-

201

signed by user-specific, average, or standard wall thickness. Furthermore, BSS-Indoor can clearly detect

203

single-plane walls, allowing users to manually review and adjust them if necessary. Importantly, BSSIndoor does not rely on any learning techniques requiring large-scale manual annotations for thickness

204

estimation, which makes it easier for BIM practitioners to adopt our method in the future.

202

205

There are also a lot more details should be improved for the current BSS-Indoor, including (1) fine-

206

tuning the parameters of PointContrast on CV4AEC dataset; (2) adjusting the hyper-parameters of plane

207

detection, Alpha shape polygonization, and the optimization for aligning the BSS segments; (3) debug-

208

ging the current implementation; (4) testing the method on more samples. Addressing these limitations

209

could lead to enhanced performance and more accurate results.

210

6

211

In this paper, we introduced BSS-Indoor, a novel volumetric reconstruction method for walls based on

212

Building Section Skeleton. Our method successfully adapts BSS for building interiors and simultane-

213

ously locates and estimates wall thicknesses. BSS-Indoor differentiates between the reconstruction of

214

dual-plane walls and single-plane walls, enhancing the accuracy and flexibility of wall thickness esti-

215

mation. Preliminary experiments on the CV4AEC (3D) dataset show the effectiveness of BSS-Indoor

216

in reconstructing volumetric walls. Although BSS-Indoor demonstrates potential, further refinement is

217

required to meet the standards necessary for real-world Scan-to-BIM practice in the building sector. Fu-

218

ture work will focus on improving the implementation and addressing the limitations identified in the

219

preliminary experiments.

Conclusion

7

220

Acknowledgments

221

The work described in this paper was supported by the Hong Kong Innovation and Technology Fund

222

(ITF) [grant numbers ITP/004/23LP].

223

References

224
225

[1] NBIMS-US. National BIM Standard-United States V4. https://www.nationalbimstandard.org/,
2022. Accessed: 2022-12-29.

226

[2] Mansour Esnaashary Esfahani, Christopher Rausch, Mohammad Mahdi Sharif, Qian Chen, Carl Haas, and

227

Bryan T Adey. Quantitative investigation on the accuracy and precision of scan-to-bim under different mod-

228

elling scenarios. Automation in Construction, 126:103686, 2021.

229

[3] Frédéric Bosché, Mahmoud Ahmed, Yelda Turkan, Carl T Haas, and Ralph Haas. The value of integrating

230

Scan-to-BIM and Scan-vs-BIM techniques for construction monitoring using laser scanning and BIM: The

231

case of cylindrical MEP components. Automation in Construction, 49:201–213, 2015.

232

[4] Yijie Wu, Maosu Li, and Fan Xue. Towards fully automatic scan-to-bim: A prototype method integrating deep

233

neural networks and architectonic grammar. In EC3 Conference 2023, volume 4, pages 0–0. European Council

234

on Computing in Construction, 2023.

235
236

[5] Weilian Song, Jieliang Luo, Dale Zhao, Yan Fu, Chin-Yi Cheng, and Yasutaka Furukawa. A-scan2bim: Assistive scan to building information modeling. British Machine Vision Conference (BMVC), 2023.

237

[6] Jiali Han, Mengqi Rong, Hanqing Jiang, Hongmin Liu, and Shuhan Shen. Vectorized indoor surface recon-

238

struction from 3d point cloud with multistep 2d optimization. ISPRS Journal of Photogrammetry and Remote

239

Sensing, 177:57–74, 2021.

240

[7] Jiali Han, Yuzhou Liu, Mengqi Rong, Xianwei Zheng, and Shuhan Shen. Floorusg: Indoor floorplan recon-

241

struction by unifying 2d semantics and 3d geometry. ISPRS Journal of Photogrammetry and Remote Sensing,

242

196:490–501, 2023.

243
244

245
246

[8] Hao Fang, Florent Lafarge, Cihui Pan, and Hui Huang. Floorplan generation from 3d point clouds: A space
partitioning approach. ISPRS Journal of Photogrammetry and Remote Sensing, 175:44–55, 2021.
[9] Hao Fang, Cihui Pan, and Hui Huang. Structure-aware indoor scene reconstruction via two levels of abstraction.
ISPRS Journal of Photogrammetry and Remote Sensing, 178:155–170, 2021.

247

[10] Pawel Boguslawski, Sisi Zlatanova, Dariusz Gotlib, Michał Wyszomirski, Miłosz Gnat, and Piotr Grzem-

248

powski. 3d building interior modelling for navigation in emergency response applications. International

249

Journal of Applied Earth Observation and Geoinformation, 114:103066, 2022.

250

[11] Mansour Mehranfar, Alexander Braun, and André Borrmann. From dense point clouds to semantic digital mod-

251

els: End-to-end ai-based automation procedure for manhattan-world structures. Automation in Construction,

252

162:105392, 2024.

253

[12] Yijie Wu, Fan Xue, Maosu Li, and Sou-Han Chen. A novel building section skeleton for compact 3d recon-

254

struction from point clouds: A study of high-density urban scenes. ISPRS Journal of Photogrammetry and

255

Remote Sensing, 209:85–100, 2024.

256

[13] Iro Armeni, Erzhuo Che, Martin Fischer, Yasutaka Furukawa, Daniel Hall, Jung Jaehoon, Fuxin Li, Michael

257

Olsen, Marc Polleyfeys, and Yelda Turkan. Computer visionin the built environment. https://cv4aec.

258

github.io/, 2023. Accessed: 2024-3-28.

8

259

[14] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised

260

pre-training for 3d point cloud understanding. In Computer Vision–ECCV 2020: 16th European Conference,

261

Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 574–591. Springer, 2020.

262

[15] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner.

263

Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on

264

computer vision and pattern recognition, pages 5828–5839, 2017.

265
266

267
268

269
270

271
272

[16] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017.
[17] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. In
Computer graphics forum, volume 26, pages 214–226. Wiley Online Library, 2007.
[18] GEOS contributors. GEOS coordinate transformation software library. Open Source Geospatial Foundation,
2021.
[19] Mariette Yvinec. 2D triangulations. In CGAL User and Reference Manual. CGAL Editorial Board, 5.6.1
edition, 2024.

273

[20] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com, 2023.

274

[21] Sayan Deb Sarkar. 3d building models matching and evaluation code. https://github.com/cv4aec/

275

3d-matching-eval, 2023. Accessed: 2024-3-28.

9