2023 European Conference on Computing in Construction
40th International CIB W78 Conference
Heraklion, Crete, Greece
July 10-12, 2023

Towards fully automatic Scan-to-BIM: A prototype method integrating deep neural networks
and architectonic grammar
Yijie Wu1 , Maosu Li1 , and Fan Xue1
1 Faculty of Architecture, The University of Hong Kong, Pokfulam, Hong Kong SAR, China

This is the post-print version of the conference paper:
Wu, Y., Li, M., & Xue, F. (2023). Towards fully automatic Scan-to-BIM: A prototype method integrating deep neural networks and architectonic
grammar. Proceedings of the 2023 European Conference on Computing in Construction and the 40th International CIB W78 Conference, 257,
Heraklion, Crete, Greece. https://doi.org/10.35490/EC3.2023.257
The final version: https://ec-3.org/publications/conference/paper/?id=EC32023_257 .This use of document is licensed under CC-BY NC 4.0.

Abstract
Building Information Modeling (BIM) has presented great
potential in the construction industry. Scan-to-BIM is demanded to verify as-designed models and to digital twin
many existing buildings without BIMs. This paper focuses on an extreme case without human interferences like
data cleansing and partitioning – fully automatic Scan-toBIM, on which recent advances in 3D scanning and deep
neural networks (DNNs) shed light. This paper presents
a prototype FLKPP that integrates DNNs with architectonic grammar for fully automatic Scan-to-BIM. FLKPP
won the 2nd place in 3D reconstruction (F15cm = 0.316,
F110cm = 0.454, F120cm = 0.584) and the 3rd place in the
2D reconstruction in the 2nd International Scan-to-BIM
Challenge. Nevertheless, the results of all methods in the
challenge were still limited in training data quality, DNN
architecture, and inconsistency between DNN predictions
and BIM reconstruction; the observations indicate a long
way leading to the complete automation of Scan-to-BIM.

Introduction
Building Information Modeling (BIM) is becoming an
essential practice to optimize the whole building life
cycle from planning to demolition (NBIMS-US, 2022).
The adoption of BIMs at the urban scale can help with
fine-scale city information modeling and urban analytics (Batty, 2013; Xue et al., 2020; Li et al., 2022). However, the restricted efficiency of creating and updating the
as-is or as-built Building Information Models (BIMs) of
existing or even aged buildings is still a crucial hinder
against the trend of BIM (Esfahani et al., 2021). To create as-is/as-built BIMs, 3D scanning, in particular Light
Detection and Ranging (LiDAR), is nowadays prevalent to
capture the realistic and detailed surface status of building
interior rapidly (Bosché et al., 2015). Yet, it is still fairly
time-consuming to manually create BIMs from 3D scans
through burdensome editing. Hence, automatic Scan-toBIM is highly desired in the construction field for years,
which attracts significant efforts from both academia and
industry in developing this automation (Wang et al., 2019;
Xue et al., 2021; Wu et al., 2021).
However, the full automation of Scan-to-BIM remains
challenging due to the complexity and diversity of the
building interior as well as the data deficiencies of 3D
scanning. The automation of Scan-to-BIM can be roughly
decomposed into two fundamental tasks. The first task is
to understand the semantics of the 3D scans, such as recog-

nizing the walls, slabs, and columns in the point cloud. It
could be extremely difficult and onerous to manually craft
rules for recognizing different architectural structures with
diverse styles or flexible designs. The second task aims at
the topologically-consistent and parametric reconstruction
of geometry, such as estimating the shapes, locations, orientations, and dimensions of wall solids without conflicts.
The challenges encountered by this task are to guarantee
geometric accuracy despite the inevitable noise, outliers,
nonuniform sampling, and missing parts of scans (Berger
et al., 2017).
In recent years, point cloud processing has been largely
promoted by deep learning by the community of computer
vision (CV), which led to advanced tools especially for
semantics understanding to equip the Scan-to-BIM. One
of the most significant advances is the 3D semantic segmentation based on deep neural networks (DNNs). On top
of the public and large-scale 3D datasets (Armeni et al.,
2016; Dai et al., 2017), DNNs, such as PointNet++ (Qi
et al., 2017) and KPConv (Thomas et al., 2019), has been
designed and trained to predict the semantic categories
of each point in the 3D scans of building interior, equivalently segmenting a scan into clusters belong to different physical objects. Another main track that also aims
at semantics understanding is the 3D object detection (Qi
et al., 2018) which would benefit the recognition and reconstruction of furniture or small objects in building interior. Meanwhile, the features of points can be learned
through DNNs for correspondence matching and registration between multiple scans (Aoki et al., 2019; Choy et al.,
2020). Besides, existing studies also parsed point clouds
from the top views to simultaneously extract the semantics
and reconstruct floor plans (Liu et al., 2018; Chen et al.,
2019).
Although the DNN-based point cloud processing developed by the CV community inspires Scan-to-BIM in semantics understanding and geometry reconstruction, there
are still crucial gaps in adopting DNNs for fully automatic
Scan-to-BIM. In this paper, we dissect these gaps from the
input and output sides. The common issue on the input
side is that the present large-scale datasets for benchmarking DNN-based algorithms are substantially simpler and
smaller than those used in AEC practice. More particular, the datasets for 3D semantic segmentation are limited
to room scale, although in practice, an initial scan may
include parts of multiple rooms and a finished scan may
span the entire building story. Meanwhile, there could

1. Preprocessing

Input:
Point clouds

2. Learning

3. Postprocessing

2D edges detection

2D edge repairing

2D output:
Floorplans

3D semantic segmentation

RANSAC plane fitting

Instance segmentation

3D output:
3D BIMs

Room clustering, layers
parsing, & clutter removal
3D box estimation

Figure 1: Overview of the prototype: Floor Layer-based Kernels and Pillars of Points (FLKPP).

be fewer noises, outliers, or missing parts in the current
datasets, which also contributes to the high accuracy reported by the current DNN-based 3D semantic segmentation. Hence, without careful preprocessing, there could
be accuracy gaps between the reported ones and the actual adoption in Scan-to-BIM. From the output side, the
target results as 3D BIMs or 2D floor plans presented different characteristics from the 3D point-wise prediction.
The end branches of DNNs should be modified to generate
line drawings or 3D parametric primitives, or postprocessing should be used to ’translate’ the point-wise prediction
into parametric 2D or 3D components.
To fill the above-mentioned gaps and maximize the learning ability of DNNs, this paper presents a prototype integrating DNNs and architectonic grammar for fully automatic Scan-to-BIM. Architectonic grammar, such as the
planarity, angle regularity, alignment, and adjacency of
architectural structures as well as the furniture distribution of building interior, could be harnessed as useful a
prior in the preprocessing, network design, and postprocessing of DNNs. In this paper, the proposed prototype
involves preprocessing to clean and simplify the input data
for DNNs based on furniture distribution. For the adoption
of DNNs, except for embedding state-of-the-art (SOTA)
DNNs directly in the prototype, the DNN to reconstruct
floor plans has been specially designed for estimating 2D
line drawings more efficiently. After the learning procedure, postprocessing has been designed to finalize or repair
the prediction of DNNs and output the floor planes and 3D
BIMs. The prototype, FLKPP, is named after the floorlayer based preprocessing, semantic segmentation through
KPConv (Thomas et al., 2019), and the inputs as point
pillars (Lang et al., 2019) for the DNN to reconstruct floor
plans.
The proposed FLKPP is also the entry that won the second and third places, respectively in the 3D and 2D track
of the 2nd International Scan-to-BIM challenge hosted
by the 2nd Workshop on Computer Vision in the Built

Environment, which was held in conjunction with IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR) 2022. Meanwhile, the workshop aims at connecting fields AEC and CV and devotes to benchmarking, promoting, and communicating the fully automatic solutions
of Scan-to-BIM. A large-scale dataset aligned to the practical scenarios in AEC has been released, and the Scanto-BIM challenge, including floor plan reconstruction (2D
track) and 3D building model reconstruction (3D track),
has been held based on the dataset.
The remainder of this paper introduces FLKPP, reports the
results of our entry on the 2nd Scan-to-BIM Challenge,
and discusses the integration of DNNs with architectonic
grammar for fully automatic Scan-to-BIM as well as the
limitations and future improvement of the current FLKPP.

The proposed FLKPP
Overview
Fig. 1 shows the overview of our prototype. The pipeline
takes the point cloud of a whole building story as input
and outputs 2D floor plans and 3D BIMs. Three stages
are designed to prepare the inputs of DNNs, i.e., Stage 1
‘Preprocessing’, involve DNNs, i.e., Stage 2 ‘Learning’,
and finalize their prediction into floor plans and 3D BIMs,
i.e., Stage 3 ‘Postprocessing’.
Preprocessing
The preprocessing is designed according to architectonic
grammar for cleaning and simplifying the input point
clouds by removing the furniture clutters and preserving
the architectural structures. More specifically, the preprocessing extracts the points of or close to the walls, ceilings,
and edges. Then, the remaining points can be grouped into
clusters of different rooms and removed to let the DNNs
focus on the architectural structures for Scan-to-BIM. For
computational efficiency, the point cloud is first voxelized
to label the occupied voxels close to the edges, ceilings,
and walls, as illustrated in the left part of Figure 2. The

labeling is performed iteratively in terms of whether the
neighborhoods on the top and sides of a voxel are also occupied. After removing the clutters, the remaining point
cloud mainly covering the architectural structures can be
separated into the upper, middle, and lower layers. The
point densities of the three layers are distributed according to the locations and dimensions of architectural structures, which provides hints for the following semantic understanding and geometry reconstruction.
Learning
In this stage, we explored different schemes to use DNNs
for 2D and 3D reconstruction. 3D reconstruction can be
derived naturally and directly from the 3D semantic segmentation based on DNNs, while 2D reconstruction does
not need to rely on the 3D segmentation. Hence, for 2D
reconstruction, the 3D scans are flattened features of 2D
grids from the top view. We then designed a more compact representation as the input of the DNN and chose the
proper top branches to predict the line drawings directly.
Learning for 2D reconstruction
As illustrated in Figure 3, a horizontal 2D grid is created
for the input point cloud. The input point cloud is then
converted into the feature of each cell in the grid, forming a compact representation that can be the input of any
standard 3D convolutional network architecture. The feature of each cell looks like a vertical column and is thus
named as point pillar (Lang et al., 2019), which presents
much higher efficiency in semantics understanding task
from the top view. As the preprocessing cleans and separate the point cloud into three floor layers, the normalized
distribution of the point density on each floor layer can
serve as a simple yet distinguishable feature to recognize
the architectural structures. Meanwhile, considering the
reconstruction of floor plans are equivalent to predicting
a set of lines and thanks to point pillars embedded in the
2D grids, the DNNs designed for the line detection of 2D
images can be naturally used in predicting the line drawings as floor plans, as shown at the far right of Figure 3.
Therefore, the ‘ConvNet’ in Figure 3 to predict the corners and edges from the point pillars can be replaced with
the SOTA DNNs for line detection, such as L-CNN (Zhou
et al., 2019) and deep Hough Tranformation-based line detection (Zhao et al., 2021). In the implementation of the
Challenge, L-CNN was used.
Learning for 3D reconstruction
For 3D reconstruction, we directly adopt the SOTA DNNs
of 3D semantic segmentation to label the points into different categories, including walls, doors, columns, and others. The input of the DNN is the point cloud cleaned by
preprocessing. Pilot tests on the Challenge dataset proved
that the IoU of 3D semantic segmentation could be significantly improved with data after cleaning. Furthermore,
KPConv (Thomas et al., 2019) was adopted in our prototype because it achieved the highest Intersection-of-Union
(IoU) in our pilot experiments on the dataset of the Scan-

to-BIM Challenge. Besides, the instance segmentation
among the same categories and the 3D geometric reconstruction are left to the postprocessing.
Postprocessing
Postprocessing for 2D reconstruction
Thanks to the DNN for 2D reconstruction predicting the
line drawings directly, only repairing is required in the
postprocessing for the reconstruction of floor plans. As
shown in the upper of Figure 4, there are gaps between
the predicted lines. The postprocessing reconnects them
based on the probability of the line map simultaneously
estimated by the DNN for 2D reconstruction and the room
map generated by the room clustering in the preprocessing
(as presented in the right-hand side of Figure 2). Figure 5
presents the mentioned maps of lines and rooms, while the
repaired floor plan is shown in the lower of Figure 4.
Postprocessing for 3D reconstruction
The postprocessing of 3D reconstruction should ‘translate’
the points labeled as wall, door, and column into 3D parametric instances. For points labeled as the wall, the postprocessing splits them into different groups according to
their orientations. Next, in each group, the points are further clustered by DBSCAN (Ester et al., 1996) to separate
them into individual walls. The instance segmentation of
doors and columns is conducted in a similar way as the
wall segmentation. For the doors connected to each other
as shown in the middle of Figure 6, we threshold the width
of the door instances and clip those larger than the threshold into parts. After the instance segmentation, the postprocessing estimates the Manhattan bounding box of each
instance and store the results in the IFC standard.

Entry results on 2nd Scan-to-BIM Challenge
Dataset, tracks, and evaluation metrics
The whole dataset of this Challenge, including those for
the training and validation of both the 2D and 3D tracks,
contains about 80 point clouds, each covering one building story. Examples are shown in Figure 7. The number of points in each point cloud could be up to a billion. Office and parking lots are the top 2 scenes in the
dataset. Besides, as presented in Figure 8, floor plans as
line drawings and the parametric 3D instances with semantic tags are provided as the ground truth for the 2D and 3D
tracks respectively. The evaluation of 2D reconstruction
involves both the geometric and topological metrics, including the IoU, endpoint accuracy and orientation deviation of the predicted segments, warping error (Jain et al.,
2010), and Betti number error. Meanwhile, the current 3D
evaluation metrics only concern the geometry, including
the 3D IoU of the walls’ bounding boxes and the accuracy
of endpoints. Readers can refer to the Challenge website
(https://cv4aec.github.io/) for more details about
the dataset, tracks, and evaluation.

1. Space voxels
(closer to Edge, Ceiling, Walls)

2. Room clustering
(using voxels 1m to ceilings)

(Voxels occupied
by scan data)
(Indoor
space
voxels)

E C C C
W E C C
W W E C
Figure 2: Illustration of the preprocessing.

ConvNet

Input: Point cloud

Point pillars of 2D grid

Predicted corners and edges

Output: Floorplan

Figure 3: Pipeline of the learning for 2D reconstruction.

cision and recall of endpoints at all the evaluation resolutions, i.e., 5, 10, and 20 cm, are still far from satisfactory. The highest value is still lower than 0.4, regardless of
those metrics less than 0.1 at the highest resolution which
is still much more coarse than the mm accuracy requirement common in AEC. The remaining three metrics in Table 1 mainly reflect the closeness of space split by the lines
of floor plans. The IoU of FLKPP is 0.374, while the winner’s achieved 0.657. Fig. 9 presents selected results of our
2D reconstruction. FLKPP has captured most main architectural structures on the floor plans. Yet, there still are
some gaps between the reconstructed segments, resulting
in an insufficient IoU. Meanwhile, the clusters of messy
lines can be spotted in the upper two examples in Fig. 9
due to the unstable line prediction at stairs by the DNN.
Figure 4: Postprocessing of the 2D reconstruction.

3D reconstruction
The evaluation results of the top 4 entries are reported in
Table 2, and FLKPP ranks second. Compared with the 2D
reconstruction, the precision and recall of endpoints, reported as F1-measure in Table 2, are much higher. Yet,
the 3D IoUs of architectural structures reconstructed by
FLKPP and other entries still should be much improved
in the future. As shown in Fig. 10, despite fairly reconstructing most architectural structures, some small structures and connections between the adjacent walls are still
missing in the current reconstruction.
Figure 5: Line map and room map for repairing the floor plans
predicted by the DNN.

Discussion

2D reconstruction

The rapid development of DNNs presents spectacular potential in promoting point cloud processing and automatic
Scan-to-BIM. However, careful integration is needed to

Table 1 shows the evaluation results of FLKPP and other
entries. Our prototype came 3rd in the 2D track. The pre-

Integrating DNNs with architectonic grammar for fully
automatic Scan-to-BIM

Figure 6: Postprocessing of the 3D reconstruction.
Table 1: Top 4 entries of the 2D reconstruction.
Precision

Recall

Method

5cm

10cm

20 cm

5cm

10cm

20cm

IoU

Warping Error

Betti Error

Seg2Plan
S2FP
FLKPP
VecIM

0.052
0.02
0.016
0

0.203
0.085
0.068
0

0.335
0.146
0.132
0

0.015
0.048
0.032
0

0.065
0.22
0.129
0

0.114
0.375
0.253
0

0.657
0.517
0.374
0

0.249
0.188
0.258
0.731

1.076
1.14
1.128
1.646

Figure 8: Groundtruth examples of the 2D (left) and 3D (right)
tracks.

Figure 7: Screenshots of the dataset. The right column presents
the zoom-in views of their left.

handle the issues raised by the data quality, choose and
adjust the network architectures, and link the prediction of
networks to the final reconstruction. The proposed FLKPP
is a prototype that explores solutions for optimal integration.
1. Training data quality. The SOTA DNNs in point
cloud processing, in particular the 3D semantic segmentation, are mainly trained on datasets that are
much simpler than the practical scans in AEC. Many
of them are confined to the room scale, while the practically finalized scans usually cover the whole building story. Meanwhile, those datasets have much less

noise, outliers, and occlusions. Such data condition
is difficult for the real scanning in AEC to guarantee,
resulting in the accuracy gaps between the reported
metrics by SOTA DNNs and real applications. Besides, the relevant DNNs usually need supervision
with large-scale labeling. The existing CAD drawings and 3D models of buildings can serve as an offthe-shelf data source to significantly ease the labeling
of 3D scans. However, due to the building innovation and own errors in the drawings or models, there
could be inevitable deviations from actual situations
captured by the scans, which would undermine the
prediction of DNNs.
2. Network architecture. Although existing DNNs on
both 3D and image processing provide numerous options to embed the deep learning in Scan-to-BIM, it is
beneficial for efficiency, robustness, and simplicity to

Table 2: Top 4 entries of the 3D reconstruction.
IoU
Method
Seg2BIM
FLKPP
PointToBIM
BoxDetector

Average F1

10cm F1

Average

Column

Door

Wall

5cm

10cm

20cm

Column

Door

Wall

0.309
0.231
0.17
0.024

0.47
0.372
0.396
0.038

0.26
0.23
0.061
0.006

0.266
0.152
0.15
0.033

0.417
0.316
0.276
0.109

0.515
0.454
0.366
0.171

0.577
0.584
0.448
0.258

0.618
0.608
0.633
0.167

0.494
0.367
0.165
0.144

0.477
0.452
0.415
0.197

Figure 10: Examples of 3D reconstruction by FLKPP.

3D parametric instances directly or 2) postprocessing DNNs’ prediction into floor plans or 3D parametric instances according to appropriate architectonic
grammar. Note that there may still be a need for further refinement based on architectonic grammar even
choosing the first option, which depends on the quality of network prediction and the requirement on the
reconstruction quality.
Figure 9: Examples of 2D reconstruction by FLKPP.

make an optimal choice and integration. 3D semantic
segmentation and object detection are the most direct
choices for Scan-to-3DBIM. For the 2D reconstruction, however, it is not necessarily the same case. We
believe that it could be more efficient to flatten the 3D
data as 2D grids and use the DNNs of image processing, which is recognized as more mature than that of
3D processing.
3. Inconsistency between DNN predictions and BIM reconstruction. Without altering the output layers of
DNNs, there could be ‘a last mile’ from the prediction of DNNs to the final 2D and 3D reconstruction.
Two options to complete this last mile are 1) reframing the last layers of DNNs to generate floor plans or

FLKPP presents a prototype taking the above three aspects
into account. The preprocessing removes many of the unwanted clutters of the point clouds to simplify the input
of DNN-based 3D semantic segmentation. Meanwhile,
the preprocessing also separates the cleaned point cloud
into floor layers, which leads to shorter point pillars for
the learning of 2D reconstruction. In the learning stage,
to improve the efficiency of learning, FLKPP converts the
3D scans into the point pillars of a 2D grid and frames the
2D reconstruction as the line prediction based on DNN,
leading to the end-to-end learning for Scan-to-Floorplan.
But refinement to fill the gaps between segments predicted
by the DNN is still needed in our implementation. For 3D
reconstruction, FLKPP uses a SOTA DNN of 3D semantic segmentation directly with the postprocessing to further
segment and estimate the bounding boxes of the wall, column, and door instances.

Limitations of the current FLKPP
Results reported on the Challenge in Table 1 and 2 reveal
that there is much space for the full automation of Scanto-BIM to improve its accuracy. The accuracy bottlenecks
partly originated from the DNNs themselves and also
partly from the integration for Scan-to-BIM. For FLKPP,
the limitations and further improvements are listed below.
1. Closeness and topology. Both 2D and 3D reconstruction of FLKPP lack explicit mechanisms to guarantee
the closeness of the split space in a building story and
the topological consistency around adjacent architectural structures. This issue results in the missing connections between adjacent walls and thus a lower IoU
in the 2D track. Inspired by the presentations of other
entries especially the winner, Seg2BIM or Seg2Plan,
an initial prediction can be extended to intersect those
other non-parallel predictions, and a selection based
on integer optimization can be conducted to repair the
disconnection.
2. Clutter removal. Although the preprocessing largely
removes the unwanted clutter for the following reconstruction, the heuristic rules used could be problematic if the ceiling scans do not form large planes, resulting in unstable predictions of the 2D learning as
shown in the bottom row in Fig. 9. More robust preprocessing should be designed to handle such cases.
3. Endpoint accuracy. As shown in Table 1 and 2, the
precision, recall, and F1-measure of endpoints in both
tracks are far from satisfactory, especially in the 2D
reconstruction. Except for improving the postprocessing in the future, this issue can be traced back to
the deviation between the 2D ground truth, i.e., existing CAD drawings of the scanned buildings, and the
actual situations in 3D scans. Meanwhile, there could
be no physical correspondences of the segment endpoints in CAD drawings, which confuses the DNN
training for 2D reconstruction. Specific mechanisms
to handle the issues of using existing CAD drawings
for supervision should be designed.

Conclusion
The efficiency and automation of creating as-is/as-built
BIMs are highly desired for BIM popularization in AEC.
Despite the rapid advance of 3D scanning and deep neural networks, fully automatic Scan-to-BIM still encounters
many challenges. To adopt DNNs in this automation, careful designs are needed to handle the issues caused by data
quality in real scenarios in the industry. Meanwhile, a robust pipeline of Scan-to-BIM based on DNNs also relies
on the appropriate selection or adjustment of network architectures as well as the linking from the prediction of
DNNs to the target BIM reconstruction.
This paper presents a prototype, FLKPP, successfully integrating DNNs with architectonic grammar and aiming
at the full automation of Scan-to-BIM. FLKPP consists of

three stages, i.e., preprocessing, learning, and postprocessing, and performs both floor plan (2D) and 3D BIM reconstruction. The preprocessing cleans and tiers input point
clouds to ease and improve the prediction quality of the
following learning. Next, FLKPP explores two different
network architectures for 2D and 3D reconstruction. We
directly adopt SOTA DNN-based 3D semantic segmentation for 3D reconstruction, while the 2D counterpart converts 3D scans into point pillars of 2D girds and frames
floor plan reconstruction as DNN-based line prediction.
The postprocessing then turns the 3D segmentation into
parametric walls, doors, and columns.
FLKPP won the 2nd and 3rd places on the 2nd International Scan-to-BIM Challenge, which validated the prototype on a large-scale dataset close to real situations in
AEC. However, the reported metrics of all entries in the
Challenge are still far from satisfactory. The precision and
recall of the endpoints in floor plans could be lower than
0.1. Significant efforts towards the accurate automation of
Scan-to-BIM are required in the future to improve both the
direct predictions of DNNs and the integration of learning
and architectonic grammar.

Acknowledgments
The work described in this paper was supported by Hong
Kong Research Grants Council (No. 27200520) and in part
by Hong Kong Innovation and Technology Commission
(No. ITP/004/23LP) and Natural Science Fund of Guangdong (2023A1515010757).

References
Aoki, Y., Goforth, H., Srivatsan, R. A., and Lucey, S.
(2019). Pointnetlk: Robust & efficient point cloud registration using pointnet. In Proceedings of the IEEE/CVF
conference on computer vision and pattern recognition,
pages 7163–7172.
Armeni, I., Sener, O., Zamir, A. R., Jiang, H., Brilakis, I.,
Fischer, M., and Savarese, S. (2016). 3d semantic parsing of large-scale indoor spaces. In Proceedings of the
IEEE conference on computer vision and pattern recognition, pages 1534–1543.
Batty, M. (2013). The New Science of Cities. MIT press.
Berger, M., Tagliasacchi, A., Seversky, L. M., Alliez, P.,
Guennebaud, G., Levine, J. A., Sharf, A., and Silva,
C. T. (2017). A survey of surface reconstruction from
point clouds. In Computer graphics forum, volume 36,
pages 301–329. Wiley Online Library.
Bosché, F., Ahmed, M., Turkan, Y., Haas, C. T., and
Haas, R. (2015). The value of integrating Scan-to-BIM
and Scan-vs-BIM techniques for construction monitoring using laser scanning and BIM: The case of cylindrical MEP components. Automation in Construction,
49:201–213.

Chen, J., Liu, C., Wu, J., and Furukawa, Y. (2019). Floorsp: Inverse cad for floorplans by sequential room-wise
shortest path. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661–
2670.
Choy, C., Dong, W., and Koltun, V. (2020). Deep global
registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages
2514–2523.
Dai, A., Chang, A. X., Savva, M., Halber, M., Funkhouser,
T., and Nießner, M. (2017). Scannet: Richly-annotated
3d reconstructions of indoor scenes. In Proceedings
of the IEEE conference on computer vision and pattern
recognition, pages 5828–5839.
Esfahani, M. E., Rausch, C., Sharif, M. M., Chen, Q.,
Haas, C., and Adey, B. T. (2021). Quantitative investigation on the accuracy and precision of scan-to-bim under different modelling scenarios. Automation in Construction, 126:103686.
Ester, M., Kriegel, H.-P., Sander, J., Xu, X., et al. (1996). A
density-based algorithm for discovering clusters in large
spatial databases with noise. In kdd, volume 96, pages
226–231.
Jain, V., Bollmann, B., Richardson, M., Berger, D. R.,
Helmstaedter, M. N., Briggman, K. L., Denk, W., Bowden, J. B., Mendenhall, J. M., Abraham, W. C., et al.
(2010). Boundary learning by optimization with topological constraints. In 2010 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, pages 2488–2495. IEEE.
Lang, A. H., Vora, S., Caesar, H., Zhou, L., Yang, J., and
Beijbom, O. (2019). Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 12697–12705.
Li, M., Xue, F., Wu, Y., and Yeh, A. G. O. (2022). A room
with a view: Automatic assessment of window views
for high-rise high-density areas using City Information
Models and deep transfer learning. Landscape and Urban Planning, 226:104505.
Liu, C., Wu, J., and Furukawa, Y. (2018). Floornet: A
unified framework for floorplan reconstruction from 3d
scans. In Proceedings of the European Conference on
Computer Vision (ECCV), pages 201–217.
NBIMS-US (2022). National BIM Standard-United States
V4. https://www.nationalbimstandard.org/.
Accessed: 2022-12-29.
Qi, C. R., Liu, W., Wu, C., Su, H., and Guibas, L. J. (2018).
Frustum pointnets for 3d object detection from rgb-d
data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918–927.

Qi, C. R., Yi, L., Su, H., and Guibas, L. J. (2017). Pointnet++: Deep hierarchical feature learning on point sets
in a metric space. Advances in neural information processing systems, 30.
Thomas, H., Qi, C. R., Deschaud, J.-E., Marcotegui, B.,
Goulette, F., and Guibas, L. J. (2019). Kpconv: Flexible and deformable convolution for point clouds. In
Proceedings of the IEEE/CVF international conference
on computer vision, pages 6411–6420.
Wang, Q., Guo, J., and Kim, M.-K. (2019). An application oriented scan-to-bim framework. Remote sensing,
11(3):365.
Wu, Y., Shang, J., and Xue, F. (2021). Regard: Symmetrybased coarse registration of smartphone’s colorful point
clouds with cad drawings for low-cost digital twin buildings. Remote Sensing, 13(10):1882.
Xue, F., Lu, W., Chen, Z., and Webster, C. J. (2020). From
lidar point cloud towards digital twin city: Clustering
city objects based on gestalt principles. ISPRS Journal
of Photogrammetry and Remote Sensing, 167:418–431.
Xue, F., Wu, L., and Lu, W. (2021). Semantic enrichment
of building and city information models: A ten-year review. Advanced Engineering Informatics, 47:101245.
Zhao, K., Han, Q., Zhang, C.-B., Xu, J., and Cheng, M.-M.
(2021). Deep hough transform for semantic line detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4793–4806.
Zhou, Y., Qi, H., and Ma, Y. (2019). End-to-end wireframe
parsing. In Proceedings of the IEEE/CVF International
Conference on Computer Vision, pages 962–971.