38th International Symposium on Automation and Robotics in Construction (ISARC 2021)

A Systematic Review of Automated BIM Modelling for

Existing Buildings from 2D Documentation
Cheng Zhanga, Yang Zoua and Johannes Dimyadib,c

a
Department of Civil and Environmental Engineering, University of Auckland, Auckland, New Zealand
b
School of Computer Science, University of Auckland, Auckland, New Zealand
c
Compliance Audit Systems Ltd, Auckland, New Zealand
E-mail: [email protected], [email protected], [email protected]

Abstract –
Building Information Model (BIM) with rich
geometric and semantic information of facilities has 1 Introduction
increasingly been used to establish the City An accurate 3D information model at the city scale
Information Model (CIM). Although BIMs for new has the potential to support a wide range of applications
buildings are becoming more available, BIMs for such as infrastructure planning, policy evaluation,
most existing buildings can only be modelled from 2D disaster management, energy demand estimation,
drawings and specifications. Manual BIM modelling situational awareness, and multiple domain integration
is an error-prone and time-consuming process, which [1]. Building information models (BIMs) of buildings
becomes more challenging for the city scale. Recently, and civil infrastructure have been widely acknowledged
automating the 3D BIM modelling process for as a key data source to establish the City Information
existing buildings from their 2D drawings has been an Model (CIM) [2]. Although the BIM of recently
emerging research trend. To understand the state-of- completed or new buildings has been modelled in the
the-art and guide future research, this paper presents design process and can be directly used for CIM, the BIM
a systematic review of automated BIM modelling for of most existing old facilities is often unavailable.
existing buildings from 2D drawings. Fifty-five Therefore, the automation of BIM modelling for existing
publications, including 34 journal articles and 21 buildings has drawn growing attention in the past years.
conference papers identified from Scopus from 1998 Currently, there are two main methods of BIM
to 2021, were reviewed and analysed. A chronological modelling for existing buildings, i.e., (1) on-site
distribution shows that most papers (60%) aimed to surveying and (2) as-built documentation.
generate 3D geometric models, and BIM modelling The first method is through on-site surveying. To
with semantic information appeared in 2015 and generate BIM for existing buildings, the collection of
increased dramatically. This review classified the sufficient spatial information is essential. Various
existing work into three aspects: geometric modelling, sensors (e.g., laser scanning [3], photogrammetry [4], etc.)
semantic modelling, and model quality checking. The have been adopted to collect 3D point cloud data, and
results show fully automated conversion of 2D then building components are detected and modelled.
drawings to semantically enriched BIM has not been Although the 3D model generated through on-site
eventuated, future work may consider overcoming surveying could represent the current state of existing
the following challenges: (1) height information is buildings, the modelling process is time-consuming and
either set as default or entered manually, and complex labour-intensive, especially when it comes to digital
components such as staircases are barely studied; (2) modelling at the city scale. Besides, those surveying
most research only focused on floor plans and ignored approaches can only detect the exterior and interior
semantic information contained in other drawings; (3) surface geometry. Structural components, such as beam
drawing errors have not been well addressed, and the and column, are usually hidden from view for aesthetic
validation of the generated model is still cumbersome. reasons, which makes measuring their geometry
impractical. Additionally, missing points are common in
Keywords – the data collection process, and, more importantly, the
Building Information Model (BIM); Existing point cloud does not contain any semantic information.
building; 2D drawings; City information model (CIM) There have been some strategies suggested to add
engineering rules in the modelling process using point
cloud, however it is still challenging to develop a

220
 38th International Symposium on Automation and Robotics in Construction (ISARC 2021)

semantically-rich BIM using this approach [5]. areas, such as Mathematics, Material Science, and
The second method focuses on generating BIMs from Physics and Astronomy, were excluded. Journal articles
2D drawings since as-built documentation contains and conference papers are both included in this review.
abundant geometric and semantic information, which
better describe the building. Since manually extracting Table 1 Literature search methods
building information from 2D drawings and modelling
all facilities at the city scale is error-prone and Search attributes Values used in the search
challenging, automated BIM modelling from design or Database Scopus
as-built documentation has seen a growing research
interest. Gimenez et al. [6] provided an in-depth review Keywords and (“2D” AND (“drawing” OR
of advanced technologies toward each step of the Boolean “floorplan” OR “floor plan”)
generation of 3D building models from 2D scanned plans operators AND (“BIM” OR “building
in 2015. However, literature aimed to achieve 3D information model” OR “3D
modelling from drawings has not been well analysed, and model”) AND (“generation” OR
many semi-automated and automated methods have been “creation” OR “reconstruction”
proposed during the last few years. Kang et al. [7] OR “construction” OR
reviewed the recent development of 3D indoor “modelling”))
reconstruction but mainly focused on technologies based Search scope Article title, abstract, or
on on-site surveying. A comprehensive review of recent keywords
advances in BIM modelling from as-built documentation Published year From all years to present
is required urgently. Subject area Engineering, Computer science,
To fully understand the contributions and limitations Social Science, Environment
of current research studies, this paper provides a Science
systematic review of automated BIM modelling for Document type Journal article; Conference
existing buildings from 2D documentations. The 2D paper;
documentations refers to various types of drawings, Language English
including digital drawings, paper-based drawings,
architectural drawings, structural drawings, and floor The initial literature search has resulted in a total of
plans. The structure of this paper is organised as follows: 456 papers. Then, the title and abstract of the literature
Section 2 presents the overall design of the literature search results were analysed to identify relevant
search. Section 3 describes three aspects of BIM publications. The filtering criteria are as follows:
modelling: geometric modelling, semantic modelling, (1) Publications aimed to construct a 3D model from 2D
and model quality checking. Section 4 discusses the drawings were included.
limitations of existing research studies and points out (2) Articles that only mentioned creating a 3D model
possible further research opportunities. Finally, Section 5 from 2D documentation but did not focus on were
summarises the finding with a conclusion. excluded.
(3) Other publications, such as those focused on
converting the building design process (from CAD
2 Review methodology to BIM), were excluded.
This research adopted the five-step review In addition, the snowballing technique was employed
methodology to conduct a systematic review [8]. First, a to find additional papers through the reference and
keyword search-based approach was adopted to collect citation lists. As a result, a total of 55 papers, including
relevant publications. Searching attributes and their 34 journal articles and 21 conference papers, were
corresponding values are listed in Table 1. retained for in-depth review and analysis.
The main keywords were, for example, ‘2D’, Figure 1 shows the chronological distribution of all
“drawing”, “floor plan”, “floorplan”, “BIM”, “building included papers, which are classified into “Semantic BIM”
information model”, “3D model”, “construction”, or “Geometric model”. The group of “Geometric model”
“reconstruction”, “creation”, “generation”, and included publications focused on construct 3D geometry
“modelling”. The Scopus database was selected for such as indoor or surface model, other publications aimed
literature searching, and papers not published in English at creating a BIM with semantic information fell into the
were omitted. Since this review focused on the 3D model group of “Semantic BIM”. Research on 3D model
generation for existing buildings from as-built construction from 2D drawings started in 1998, and all
documentation, only four relevant subject areas, identified papers published before 2015 aimed to create
Engineering, Computer Science, Social Science, and the 3D geometric model, such as indoor or surface model.
Environment Science, were considered. Other subject The semantic BIM modelling from 2D drawings first

221
 38th International Symposium on Automation and Robotics in Construction (ISARC 2021)

appeared in 2015, and a dramatic increasing trend can be classification need to be corrected in advance, and layer
observed. Then, all identified publications are reviewed information is unavailable for hand drafting. Another
critically according to three aspects of BIM modelling: solution is using geometric features and symbols to
geometric modelling, semantic modelling, and model identify building components. As the most common
quality checking. Key limitations regarding each aspect features in 2D building drawings, line segments were
are concluded, and future work aimed to address these used for identifying walls [13–18], columns [19], and
limitations are illustrated. rooms[20]. Symbols were adopted to detect grid lines,
then building elements (such as columns, beams, and
walls) can be further identified [21,22]. But these
methods are not applicable for building components with
irregular shape and will lead to false matches due to
missing or inconsistent information in as-built
documentation. With the rapid development of artificial
intelligence (AI), the adoption of AI for drawing analysis
shows a significant increase. Walls and openings are
detected from floor plan image based on a convolutional
neural network [23,24]. Rho et al. [25] developed a
machine learning-based classifier to extract text
information for component detection and localisation.
Zhao et al. [26] detected structural components (e.g.,
columns and beams) from the framing plans based on
Faster R-CNN, further created IFC BIMs for multi-story
Figure 1. Chronological distribution of included papers buildings. However, the existing application of AI only
focused on detecting specific part of information. A
unified solution for identifying all component
3 BIM Modelling from 2D documentation information and texts is preferred for automating 3D
In this section, the state-of-the-art BIM modelling for model generation.
existing buildings from 2D drawings is summarised in After components were identified, most research
terms of three aspects, namely geometric modelling, projects extruded the labelled 2D floor plan to generate
semantic modelling, and model quality checking. the 3D building model [9,10,12–19,21,23-26]. However,
Geometric modelling aims to construct the 3D geometry the floor elevation and height of openings were either set
of buildings. In contrast, semantic modelling focuses on as default or entered manually. Lu et al. [22] proposed to
extracting and attaching semantic information (e.g., room generate three orthogonal views (the top, side and front
identity, component dimension, construction material, view) of each component from the 2D architectural
etc.) contained in 2D drawings as properties of objects in drawings. Then, the 3D model of all components can be
the 3D model. Once the 3D model is constructed, the constructed and integrated to obtain the 3D building
model quality checking process is performed to verify its model. But this method only suits the modelling of
completeness and correctness. simple geometry due to the challenge regarding the
generation of three orthogonal views of building
3.1 Geometric modelling components from 2D drawings. In addition, the
combination of floor plans and elevation drawing was
3D building geometric model has been envisioned as proposed by Bortoluzzi et al. [20] to create BIM with
the data management platform for facility management. room layout and exterior openings for existing buildings.
However, manually modelling existing buildings is error- Yin et al. [11] constructed a façade BIM model by
prone and time-consuming [9]. In the past decades, there locating exterior components in floor plans and
has been a growing interest in automating 3D modelling extracting height information from elevation drawings.
based on 2D drawings. Existing research studies However, height information of interior building
regarding geometric modelling are summarised in Table components (e.g., interior openings, beam) cannot be
2. identified from elevation drawings. The combination of
The first step in the 3D model generation is more as-built documentation is preferred to obtain all
identifying and extracting building component required information for automated 3D modelling.
information from 2D drawings. Since the geometry of Another vital part of 3D building model generation is
different components is usually divided into different the matching and integration of different floors. Most
layers during computer-aided building design, layer existing research studies only considered 3D modelling
information in CAD drawings has been widely used for of one floor, and few papers studied the matching of
component recognition [9–12]. However, errors on layer

222
 38th International Symposium on Automation and Robotics in Construction (ISARC 2021)

Table 2 Summary of geometric modelling

Themes Approach Description of the approach Research
Drawing ● Based on layer The geometry of different components is stored in different [9–12]
analysis information layers of CAD drawings, but it is not applicable for manual
drafting;
● Based on Geometric features and symbols are used to identify building [13–22]
geometric features components, but irregular components cannot be detected;
● Artificial Using artificial intelligence to identify building components, [23–26]
intelligence which is suitable for all types of drawings and irregular
components;

Generating ● Extrusion of 2D Extrude the labelled floor plan to generate the 3D model, but [9,10,12–
the 3D drawing height information (e.g., floor elevation, the height of 19,21,23-26]
building openings) is either entered manually or set as default;
model ● Modelling from Generate 3D model of components from three orthogonal [22]
three orthogonal views, then integrate all component models to obtain a 3D
views building model, but it’s hard to extract the three orthogonal
views from 2D drawings;
● Combine floor Detect building components from floor plans, then extract [11,20]
and elevation heigh information of external components from elevation
drawings drawings, but internal components haven’t been considered;

Integrating ● Floorplan-based Align other floor plans to the first inputted floor plan based [11,15,16]
different matching on drawing features or use global coordinates to locate
floors components;
● 3Dfeature-based Match 3D model of different floors based on geometric [17]
matching features like pips, staircases, corners and bearing walls;

different floors to generate a 3D model of the entire constructed 3D model. Lu et al. [21] generated Industry
building. The matching approaches could be divided into Foundation Class (IFC) model from 2D drawings and
two groups. further attached material information of components
(1) Floor plan-based matching. For example, Zhu et through on-site surveying. By analysing component lists
al. [15] and Li et al. [16] aligned other floor plans to the (including column list, beam list, slab list, and wall list)
first inputted floor plan by intersection points between and floor plans, Byun et al. [19] created the IFC model
axes; Yin et al. [11] used the global coordinate system to with the material property of concert and rebar for
locate components of each floor plan. reinforced concert structures. Yang et al. [9] focused on
(2) 3D feature-based matching. For instance, Dosch the semantically-rich 3D BIM modelling from 2D CAD
et al. [17] proposed to match models generated from drawings. Structural components and corresponding axis
different floors based on features like pipes, staircases, were first detected and generated, then semantic
corners and bearing walls. information of components (e.g., coding number,
element cross-section, element reinforcement
3.2 Semantic modelling information) was linked to the axis as parameters.
In addition to geometric modelling, semantic
3.3 Model quality checking
modelling is the next stage of BIM modelling for existing
buildings. Following the construction of the geometric Once the 3D building model is constructed, it should
models, a common approach is to classify building be checked manually, semi-automatically, or
components [10,11,13,18,24,25], or attach identities to automatically to verify its completeness and correctness.
space objects [14,15,17,20,22,23]. To generate Research on model quality checking can be divided into
semantically-rich BIM and further improve its three areas.
functionality, other semantic information contained in (1) Checking the drawing quality before the 3D
2D drawings are extracted and modelled accordingly. For modelling. Once any errors contained in 2D drawings are
example, Li et al. [16] built a profile containing building corrected, the 3D model generation can proceed using the
type, size and other semantic information along with the drawings, mitigating the risk of inheriting errors or

223
 38th International Symposium on Automation and Robotics in Construction (ISARC 2021)

inconsistencies [12,18]. Future efforts are needed to integrate all information in

(2) Comparing with ground truth. Gimenez et al. [14] multiple drawings and explore the modelling of a
proposed to compare the automatically labelled drawing semantically-rich BIM data model.
with manually annotated one to evaluate the component Errors in drawings are inevitable, and existing research
identification results. To evaluate the constructed 3D only focused on detecting and correcting specific types
model, the number of components in the input 2D of errors [12,18]. A comprehensive illustration of various
drawings and output building model are compared forms of errors in 2D drawings and their corresponding
[14,19]. Rho et al. [25] and Yang et al. [9] checked the influences in 3D modelling is preferred. To evaluate the
constructed BIM model by comparing it with a manually accuracy and quality of the constructed 3D model,
modelling result. comparing it with ground truth is reliable but requires
(3) Checking without ground truth. The philosophy is extensive manual work, especially when the reference
to evaluate consistency and conflicts according to basic model does not exist. Compliance checking could be
rules, which is mainly used for checking building design conducted fully automatically and is envisioned as an
results with building codes [27]. Nikoohemat et al. [28] ideal solution for model evaluation. However, existing
proposed an approach for checking the geometric, research only considered simple building component
semantic and typological consistency of the constructed compliance [28]. To provide a more reliable model
3D model based on formal grammars defined through evaluation report, a comprehensive analysis of building
existing international standards. compliance rules is required. In addition, methods aimed
at automatically checking all relevant building
compliance aspects are needed in future studies.
4 Discussion
Although a number of automatic and semi-automatic
5 Conclusion
approaches for BIM modelling from as-built
documentation have been proposed, further research is BIM modelling for existing buildings from 2D
still needed toward fully automated modelling of drawings and specifications has attracted growing
semantically-rich BIMs for existing buildings. This research interest. It has the potential to support the
section summarises the limitations and challenges of development of CIM. This paper provides a systematic
current research along with possible future research review of 55 relevant publications to understand the
directions. state-of-the-art to guide future research. A chronological
Although building geometric information is distribution is used to illustrate the research trend. It has
distributed in various drawings, most existing research been found that most papers (60%) aimed to construct 3D
only used floor plans for geometric modelling. The geometric models, such as indoor or surface models, and
height information, including floor elevation and height semantic BIM modelling first appeared in 2015 and
of openings, is either set as default or entered manually. increased dramatically.
Few publications have proposed to extract floor elevation Recent advances in three aspects of BIM modelling,
from building elevation drawings [11,20]. To namely geometric modelling, semantic modelling, and
automatically identify the height information of all model quality checking, have been reviewed and
building components, future research could focus on the analysed critically. Limitations and challenges of
combination of various 2D as-built documentation. existing research on each aspect have also been identified,
Another challenging problem is that only common as follows:
building components are modelled, and complex (1) For geometric modelling, the height information
components, such as staircases, are ignored in current (e.g., floor elevation, the height of openings, etc.) are
research. The modelling of these building components either set as default or entered manually. Besides, most
should be further studied to create a complete BIM model research studies only focused on few categories of
for existing buildings. building components. Modelling of complex components
In contrast to geometric modelling, semantic such as staircases has barely been studied;
information modelling for existing buildings has not been (2) In contrast to geometric modelling, semantic
given the attention it needs. Most research studies only modelling has not been given the attention it needs. In
detected room identity and component category from particular, since most research only considered floor
floor plans. The generated BIM can be used for energy plans, rich semantic information in other drawings has
consumption simulation [14] or indoor navigation [23]. been ignored;
However, to support facility management and disaster (3) Errors in drawings have not been well addressed,
simulation, semantic information like space function and and the validation of the generated 3D model is still
activities, occupancy characteristics, building materials, cumbersome.
and properties of various building elements are required. To overcome these drawbacks, the integration of all

224
 38th International Symposium on Automation and Robotics in Construction (ISARC 2021)

as-built documentation should be solved for automatic [7] Z. Kang, J. Yang, Z. Yang, S. Cheng. A review
semantically-rich BIM modelling. In addition, more of techniques for 3D reconstruction of indoor
research efforts are recommended to explore the quality environments. ISPRS International Journal of
evaluation of the generated BIM model. Geo-Information. 9, 2020.
https://doi.org/10.3390/ijgi9050330.
Acknowledgements [8] Z. Zhou, Y.M. Goh, Q. Li. Overview and
analysis of safety management studies in the
This research is supported by the University of construction industry. Safety Science, 72:337–
Auckland. 350, 2015.
https://doi.org/10.1016/j.ssci.2014.10.006.
References [9] B. Yang, B. Liu, D. Zhu, B. Zhang, Z. Wang, K.
Lei. Semiautomatic Structural BIM-Model
[1] E. Shahat, C.T. Hyun, C. Yeom. City digital
Generation Methodology Using CAD
twin potentials: A review and research agenda.
Construction Drawings. Journal of Computing
Sustainability, 13, 2021.
in Civil Engineering, 34, 2020.
https://doi.org/10.3390/su13063386.
https://doi.org/10.1061/(ASCE)CP.1943-
[2] Q. Lu, A.K. Parlikad, P. Woodall, G. Don 5487.0000885.
Ranasinghe, X. Xie, Z. Liang, E. Konstantinou,
[10] R. Lewis, C. Séquin. Generation of 3D building
J. Heaton, J. Schooling. Developing a Digital
models from 2D architectural plans. CAD
Twin at Building and City Levels: Case Study
Computer Aided Design, 30:765–779, 1998.
of West Cambridge Campus. Journal of
https://doi.org/10.1016/S0010-4485(98)00031-
Management in Engineering, 36, 2020.
1.
https://doi.org/10.1061/(ASCE)ME.1943-
5479.0000763. [11] M. Yin, L. Tang, T. Zhou, Y. Wen, R. Xu, W.
Deng. Automatic layer classification method-
[3] L. Sanhudo, N.M.M. Ramos, J.P. Martins,
based elevation recognition in architectural
R.M.S.F. Almeida, E. Barreira, M.L. Simões, V.
drawings for reconstruction of 3D BIM models.
Cardoso. A framework for in-situ geometric
Automation in Construction, 113, 2020.
data acquisition using laser scanning for BIM
https://doi.org/10.1016/j.autcon.2020.103082.
modelling. Journal of Building Engineering, 28,
2020. [12] C.Y. Cho, X. Liu. An automated reconstruction
https://doi.org/10.1016/j.jobe.2019.101073. approach of mechanical systems in building
information modeling (BIM) using 2D
[4] H. Fathi, F. Dai, M. Lourakis. Automated as-
drawings. In Computing in Civil Engineering
built 3D reconstruction of civil infrastructure
2017, pages 236–244, 2017.
using computer vision: Achievements,
https://doi.org/10.1061/9780784480823.029.
opportunities, and challenges. Advanced
Engineering Informatics, 29:149–161, 2015. [13] Q. Wen, R.-G. Zhu. Automatic generation of 3D
https://doi.org/10.1016/j.aei.2015.01.012. building models based on line segment
vectorisation. Mathematical Problems in
[5] R. Sacks, A. Kedar, A. Borrmann, L. Ma, I.
Engineering, 2020.
Brilakis, P. Hüthwohl, S. Daum, U. Kattel, R.
https://doi.org/10.1155/2020/8360706.
Yosef, T. Liebich, B.E. Barutcu, S. Muhic.
SeeBridge as next generation bridge inspection: [14] L. Gimenez, S. Robert, F. Suard, K. Zreik.
Overview, Information Delivery Manual and Automatic reconstruction of 3D building models
Model View Definition. Automation in from scanned 2D floor plans. Automation in
Construction, 90:134–145, 2018. Construction, 63:48–56, 2016.
https://doi.org/10.1016/j.autcon.2018.02.033. https://doi.org/10.1016/j.autcon.2015.12.008.
[6] L. Gimenez, J.-L. Hippolyte, S. Robert, F. [15] J. Zhu, H. Zhang, Y. Wen. A New
Suard, K. Zreik. Review: Reconstruction of 3D Reconstruction Method for 3D Buildings from
building information models from 2D scanned 2D Vector Floor Plan. Computer-aided design
plans. Journal of Building Engineering, 2:24– and applications, 11:704–714, 2014.
35, 2015. https://doi.org/10.1080/16864360.2014.914388.
https://doi.org/10.1016/j.jobe.2015.04.002.
[16] T. Li, B. Shu, X. Qiu, Z. Wang. Efficient
reconstruction from architectural drawings.

225
 38th International Symposium on Automation and Robotics in Construction (ISARC 2021)

International journal of computer applications https://doi.org/10.1080/13467581.2020.1806071

in technology, 38:177–184, 2010. .
https://doi.org/10.1504/IJCAT.2010.034154.
[26] Y. Zhao, X. Deng, H. Lai. Reconstructing BIM
[17] P. Dosch, K. Tombre, C. Ah-Soon, G. Masini. from 2D structural drawings for existing
A complete system for the analysis of buildings. Automation in Construction,
architectural drawings. International Journal on 128:103750, 2021.
Document Analysis and Recognition, 3:102– https://doi.org/10.1016/j.autcon.2021.103750.
116, 2000. https://doi.org/10.1007/PL00010901.
[27] J. Zhang, N.M. El-Gohary. Integrating semantic
[18] S. Horna, D. Meneveaux, G. Damiand, Y. NLP and logic reasoning into a unified system
Bertrand. Consistency constraints and 3D for fully-automated code checking. Automation
building reconstruction. Computer-Aided in Construction, 73:45–57, 2017.
Design, 41:13–27, 2009. https://doi.org/10.1016/j.autcon.2016.08.027.
https://doi.org/10.1016/j.cad.2008.11.006.
[28] S. Nikoohemat, A.A. Diakité, V. Lehtola, S.
[19] Y. Byun, B.-S. Sohn. ABGS: A system for the Zlatanova, G. Vosselman. Consistency grammar
automatic generation of building information for 3D indoor model checking. Transactions in
models from two-dimensional CAD drawings. GIS, 25:189–212, 2021.
Sustainability, 12(17):6713, 2020. https://doi.org/10.1111/tgis.12686.
https://doi.org/10.3390/SU12176713.
[20] B. Bortoluzzi, I. Efremov, C. Medina, D.
Sobieraj, J.J. McArthur. Automating the
creation of building information models for
existing buildings. Automation in Construction,
105, 2019.
https://doi.org/10.1016/j.autcon.2019.102838.
[21] Q. Lu, L. Chen, S. Li, M. Pitt. Semi-automatic
geometric digital twinning for existing buildings
based on images and CAD drawings.
Automation in Construction, 115, 2020.
https://doi.org/10.1016/j.autcon.2020.103183.
[22] T. Lu, C. Tai, L. Bao, F. Su, S. Cai. 3D
reconstruction of detailed buildings from
architectural drawings. Computer-Aided Design
and Applications, 2:527–536, 2005.
https://doi.org/10.1080/16864360.2005.1073840
2.
[23] H. Jang, K. Yu, J. Yang. Indoor reconstruction
from floorplan images with a deep learning
approach. ISPRS International Journal of Geo-
Information, 9, 2020.
https://doi.org/10.3390/ijgi9020065.
[24] Y. Wu, J. Shang, P. Chen, S. Zlatanova, X. Hu,
Z. Zhou. Indoor mapping and modeling by
parsing floor plan images. International Journal
of Geographical Information Science, 10:1–27,
2020.
https://doi.org/10.1080/13658816.2020.1781130
.
[25] J. Rho, H.-S. Lee, M. Park. Automated BIM
generation using drawing recognition and line-
text extraction. Journal of Asian Architecture
and Building Engineering, 14:1–13, 2020.

226