Energy & Buildings 360 (2026) 117384

Contents lists available at ScienceDirect

Energy & Buildings
journal homepage: www.elsevier.com/locate/enbuild

Automated window detection for digital twin buildings: A generalized
‘sandwich’ model and practical guidelines
Siyuan Meng
a

a , Longyong Wu

a , Maosu Li

b , Anthony G. O. Yeh

c,d , Fan Xue

a,d,∗

Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong Kong, China

b Thrust of Urban Governance and Design, The Hong Kong University of Science and Technology (Guangzhou), Nansha, Guangzhou, China
c

Department of Urban Planning and Design, The University of Hong Kong, Pokfulam, Hong Kong, China

d National Center of Technology Innovation for Digital Construction Hong Kong Branch, The University of Hong Kong, Pokfulam, Hong Kong, China

a r t i c l e

i n f o

a b s t r a c t

Keywords:
Digital twin building
Openings
Window detection
Decision tree
Systematic review

Windows play a signiﬁcant role in the livability and sustainability of buildings, such as energy eﬃciency, natural
lighting, and ventilation; therefore, they are increasingly important subjects of surveys and analysis for both new
and existing buildings. Many automated window detection (AWD) workﬂows have been studied to extract windows’ properties; however, a generalized model of the underlying technologies and practical guidelines remains
lacking for both newcomers and industry practitioners. Therefore, this paper aims to address the following questions: (1) the scope and general workﬂow of AWD; (2) the trends regarding AWD key components; and (3) the
pathways of key component (e.g., data source, attribute, storage, and method) selections to ﬁt the application
scenario. Four decision trees, trained on 105 technical papers collected in accordance with the PRISMA standard,
achieved an average F1-score of 80.3% and accuracy of 88.1%. This paper contributes the ﬁrst comprehensive
review of AWD, with a general ‘sandwich’ model, a comprehensive analysis of advanced technologies, trends,
and future directions, as well as three-step guidelines for practitioners.

1. Introduction

have been increasingly incorporated into building life-cycle projects in
the past decade under the concept of Construction 4.0 [11]. Typical
DTBs encompass acquisition, transmission, modeling, data/model integration, as well as service layers [12]. DTBs provide dynamic data inputs that, in conjunction with digital models, further support reasoning
functionalities beyond those enabled by static models such as building
information model (BIM) [13,14]. These reasoning functionalities include project progress monitoring, construction information exchange,
building energy performance simulation and evaluation, and building
maintenance optimization [15–17]. Digital twin windows, the extension of DTB, can therefore empower window information modeling and
enable building life-cycle applications.
Automated window detection (AWD) refers to an automated workﬂow that aims to detect window-related information, such as position,
shape, and condition. The dynamic digitalized windows can serve as
the basis for model and data/model integration layers in DTBs, owing
to the intermediary role of windows in real-time interaction with the
built environment. AWD can further support DTB-based life-cycle reasoning in response to increasing sustainability demands, such as light

Sustainable development of buildings is increasingly recognized
worldwide. Buildings serve as primary hotspots of resource consumption and environmental impact [1], with windows constituting major
components of buildings that signiﬁcantly contribute to this overall impact. Windows are openings in walls, doors, or rooftops that serve both
physical (e.g., light exchange, air circulation, enhancement of indoor
thermal comfort, and sound transmission) and psychological (e.g., improvement of mental health, facade articulation, and aesthetic enjoyment) functions for residents [2–7]. Windows’ potential environmental impacts include energy consumption and greenhouse gas emissions
during the construction, operation, and demolition phases [1,8]. Additionally, windows are key elements in the analyses of ﬂood, wind, and
ﬁre damage risks when evaluating potential environmental damage [9].
Therefore, proper design, construction, and maintenance of windows are
essential for promoting sustainable building development and supporting the United Nations’ Sustainable Development Goal 11 (SDG 11) [10].
Digital twin buildings (DTBs), virtual replicas of physical buildings,

∗ Corresponding author.

E-mail addresses: siyuanm@connect.hku.hk (S. Meng), wulongyong@connect.hku.hk (L. Wu), maosuli@hkust-gz.edu.cn (M. Li), xuef@hku.hk (F. Xue).
https://doi.org/10.1016/j.enbuild.2026.117384
Received 8 December 2025; Received in revised form 6 February 2026; Accepted 24 March 2026
Available online 28 March 2026
0378-7788/© 2026 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 1. Examples of detected window attributes and DTB applications in the literature. (a) Window sizes from a building in the temperate zone [21]; (b) Window
types from a building in the arid zone [22], with the application of semantic enrichment modeling; (c) Window positions and opening statuses from a building in
the cold zone [23], with the application of window open behavior analysis.

and visual comfort design in design phase, construction robotics monitoring in construction phase, and monitoring carbon dioxide emissions
in O&M phase [18–20].
AWD has been studied in many cities using advanced remote sensing
and detection technologies. For example, window positions and opening status were examined in Harbin City, which is located in a cold region, as shown in Fig. 1c. Various data sources and technologies are employed in the AWD workﬂow due to the signiﬁcance of window-related
attributes, such as position, size, and type, for DTB applications. Typical data sources include imagery, point clouds, and thermal cameras,
which are used to collect information on windows’ open or closed status, geometric shapes, types, materials, and other relevant details. Both
rule-based and learning-based methods are applied across multiple disciplines to reﬁne AWD workﬂows.
Various applications of AWD have been implemented in the architecture, engineering, and construction (AEC) industry. Multidisciplinary research from other ﬁelds, such as geoscience, remote sensing, and instrumentation, also involves AWD applications. The applications of AWD
can be categorized into two types: semantic enrichment and window semantics for reasoning. Speciﬁcally, the shape of detected windows can
ﬁrst be enriched as semantic information in 3D BIM (i.e., Scan-to-BIM)
and city information models (CIMs) for managing and analyzing building information [24,25]. Furthermore, window semantics also enable
various DTB simulations and reasoning tasks. For example, the opening status of windows is utilized to monitor changes in window opening
behavior [23]; the position of windows is combined with facade area
to calculate the window-to-wall ratio [26]; the type of window is utilized to reﬁne BIM models [22]; moreover, the shapes of windows can
be utilized to monitor construction progress [27].
However, a new researcher or practitioner in related ﬁelds may ﬁnd
AWD perplexing despite its fundamental role in DTB-related functions.
Existing AWD workﬂows, including objectives, methods, and outcomes
of AWD tasks across various applications, remain fragmented. Besides,
they diﬀer substantially in their pipelines and level of automation, including data source selection, feature engineering, model storage, processing methods, and intended application. For example, data sources,
such as LiDAR point clouds and imagery, vary across the construction
project stages to support information management [28–30]. As for feature engineering, energy simulation primarily focuses more on window
open/close status [31], whereas building information extraction only requires the existence and quantity of windows [32]. Despite its growing
relevance, AWD lacks a systematic review to summarize the general information extraction process, and no structured guidelines are available
to assist newcomers in the ﬁeld. Existing literature usually considers
AWD as a secondary element embedded in other industry-speciﬁc processes or in narrowly scoped, task-speciﬁc scenarios [24]. Therefore,

this paper aims to contribute the ﬁrst review that summarizes the AWD
model, workﬂow, and guidelines in AEC industry, thereby ﬁlling a gap
in the literature.
The knowledge gap in the literature between AWD and related applications, therefore, can be summarized as the following research questions:
•

What is the scope and general workﬂow of AWD?
What are the technological advancements and trends regarding AWD
data sources, attributes, methods, and applications in the literature?
• How can a practitioner select appropriate data sources, attributes,
storage, detection methods, and pathways to ﬁt the application scenario?
•

This paper employs a mixed-methods research design to systematically answer the research questions. First, a conceptual ‘sandwich’
model of the AWD workﬂow is deﬁned to summarize the scope and
general processing steps. Then, a literature data collection process is
conducted based on the conceptual model. To answer the second question, the subsequent comprehensive literature review examines the development trends of AWD data sources, attributes, methods, and applications. Finally, a decision tree approach is applied, with the authors’
moderation, to derive concise and comprehensive AWD guidelines from
the existing workﬂows in the literature.
2. Research methods
2.1. Deﬁnition and conceptual ‘sandwich’ model of AWD
Window detection refers to the task of identifying interested
attributes–such as the position, shape, and condition of windows–within
a building, as an extension of object detection [33]. AWD denotes the automated workﬂows for collecting and processing information to accomplish window detection tasks. In DTB, AWD typically comprises workﬂows for deﬁning requirements, collecting data and information, and
executing techniques across three main layers [34]. Thus, requirements,
data and information, and execution constitute the key steps of AWD.
Fig. 2 presents a conceptual ‘sandwich’ model of requirement-driven
AWD studies as reviewed in this paper. The conceptual model outlines
the general steps of AWD identiﬁed in the literature, which comprise
three layers. The top layer is the DTB requirement layer, which includes
construction phases in the lifecycle and application needs. The ‘meat’
layer in the middle is the data and information layer, comprising data
sources, attributes, and model storage. The base layer is the execution
layer, which primarily involves rule-based and learning-based (including both traditional and deep learning) methods. In the model in Fig. 2,
the initial application block in the requirement layer determines the
2

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 2. Conceptual ‘sandwich’ model of AWD studies.
Table 1
Value and notes of the seven components in the ‘sandwich’ model’s three layers.
Layers

Components

Value

Notes

Example

Phase

Design
Construction
O&M
Building information extraction
Construction robotics
Damage simulation
Energy simulation
Heritage conservation
Semantic enrichment
Drawings
IRT
RGB camera
BIM
LiDAR
RGB-D
Dimension
Position
Shape
Status
Type
2D polygon
3D model
Raw
\
Classiﬁcation
GAN
Instance segmentation
Localization
Semantic segmentation

\
\
Operations and Management
Extract Building high-level information (e.g., ﬂoor elevation, ﬂoor number)
Use automated robotics for building construction tasks.
Simulate damage impact on buildings (e.g., ﬂood, winds).
Predict energy behaviour of buildings (e.g., thermal comfort, energy consumption)
Technical activity towards historical buildings
Interchange and interoperate on a semantic level
Floor plan images/vectors of buildings
Infrared thermography
Camera that captures light in red, green, and blue wavelengths (RGB)
Building information model
Light Detection And Ranging
Depth camera that provide both depth and RGB color information
Estimate size like width and height of window
Estimate the location of window
Estimate the geometry patch of window
Estimate whether window is close, open or half-open
Classify the window type (e.g., style)
Store in 2D image
Store in 3D data like BIM and Mesh
Not speciﬁed
Detect combined with rule and domain knowledge
Distinguish type or style of windows
Generative adversarial network
Annotate windows boundary of each window
Annotate windows position
Annotate windows boundary as a single category

[35]
[29]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[37]
[46]
[47]
[48]
[49]
[50]
[23]
[22]
[51]
[52]
[53]
[54]
[28]
[55]
[56]
[57]
[27]

Requirement
Application

Data source

Data & info.
Attribute

Storage
Rule-based
Execution

Learning-based

contextual factors that inform the middle information layer of DTB. The
bottom layer supports the data processing and the exchange of information between the three components in the middle.
Table 1 lists the possible values and notes for the components of the
‘sandwich’ model across the three layers depicted in Fig. 2. These values
represent the component options utilized across the AWD literature.
In the top requirement layers, design, construction, as well as operation and maintenance (O&M) are included in the phase component,
aligning with the building lifecycle. The application component encompasses six key analytical categories. These application scenarios utilize
the AWD results and further extend their relevance in published works,
including semantic enrichment, building information extraction, construction robotics, damage simulation, energy simulation, and heritage
conservation.
The data and information ‘meat’ layer comprises six data sources
(LiDAR, RGB camera, drawings, IRT, RGB-D, and BIM), ﬁve attribute

types (shape, position, status, type, and dimension), and three storage
formats (3D model, 2D polygon, and raw data). It is worth noting that
RGB-D and LiDAR represent distinct data sources, as RGB-D technology
is unsuitable for outdoor environments in AWD tasks.
In the base execution layer, the rule-based component refers to methods described in the literature that rely on domain knowledge and explicit rules (e.g., symmetry and similarity). In contrast, the learningbased component encompasses ﬁve prevalent deep learning approaches
identiﬁed in the literature: classiﬁcation, localization, semantic segmentation, instance segmentation, and generative adversarial networks
(GANs).
2.2. Literature search
The literature data collection was conducted according to the conceptual model and the Preferred Reporting Items for Systematic Reviews
3

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 3. PRISMA workﬂow.

and Meta-Analyses (PRISMA) standard, as illustrated in Fig. 3. The Web
of Science academic database was used to identify relevant documents.
The search strategy employed two groups of keywords: Group 1 targeted
papers focused on window or opening detection ((window* OR opening*) AND (detection OR recognition OR segmentation OR extract* OR
parsing OR reconstruct*)), while Group 2 was designed to narrow the
scope of the AWD domain to buildings only (architecture OR city OR
cities OR building* OR facade*). The literature was screened from 2000
to 2025 to reﬂect application trends.
To identify relevant literature, an initial screening was conducted
based on the title, abstract, and keywords to select articles that matched
the search criteria. In addition to sources from the AEC industry,
application-oriented literature from computer science and remote sensing ﬁelds related to buildings was included to provide a more comprehensive understanding of the diversity and trends of AWD across diﬀerent scenarios. Three duplicate entries and 1,743 out-of-scope publications were excluded. Furthermore, publications lacking one or more layers of the conceptual model–for example, remote sensing studies without any analysis or application–were removed, totaling 90 publications,
as illustrated in Fig. 3. The full texts of the remaining articles were examined to ensure comprehensive coverage. In total, 105 papers relevant to the proposed review topic were collected, with the earliest published in 2009; of these, 78 were journal articles and 27 were conference
papers.

This review examined the latest technological trends in applications,
data sources, attributes, and methods related to AWD. Each component,
along with all values presented in the Sankey chart, is depicted using a
stacked bar plot. These trends demonstrate the current choices for AWD
within the conceptual model.

2.4. Guidelines compilation
A set of AWD guidelines is compiled based on the conceptual
model and literature analysis. In the guidelines, each step was summarized based on the reviewed publications. The AWD guidelines provide practitioners with a comprehensive map to choose the optimal
methods for window detection. For researchers, the proposed guideline,
in conjunction with technological advancements and trends, serves as
a foundation for future research to explore more innovative research
concepts.
The selection procedures for components in the data and information
layer and the execution layer were illustrated using a series of decision
trees. Decision trees are interpretable methods widely used to derive
insights from complex datasets across various disciplines [58]. Four decision trees were constructed through in-depth analysis using OrangeML
(ver. 3.36) in Python (ver. 3.11). The decision tree algorithm’s primary
parameter was set to ‘min_samples_split = 2, ’ while the rest of the parameters were set to their default values. The input variables for the
data and information layer were deﬁned as requirement layer components (i.e., application and phase), while input variables for the execution layer were set as data source and storage. Literature published after
2020 was selected to ensure the methods reﬂect current practice. Unsuitable or outdated options were manually excluded based on ﬁndings
from the literature analysis. Methods, processing steps, and applications
were manually synthesized from the selected publications.
Literature published after 2020, indexed by Scopus but not by Web
of Science, was collected to validate the recommended decision trees.
Validation metrics, including accuracy, precision, recall, and F1-score,
were employed to demonstrate the accuracy and robustness of the four
decision trees.

2.3. Literature analysis based on the conceptual model
The 105 selected publications were coded according to the components of the conceptual model. A heat map based on the corresponding
author’s aﬃliated institution for each publications was generated to reﬂect the distribution of AWD applications across diﬀerent climate zones.
A Sankey chart was employed to visually represent the coded literature as an overview of the existing studies. The intertwined links
in the Sankey chart denote the systematic and organized technological
roadmaps in the literature. By examining the network representation of
the shortlisted papers in relation to the conceptual models, we summarized the general workﬂow of AWD.
4

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 4. Map of source cities with climate zone of the 105 publications in Table A.1.

3. Data extraction and analytical results

The majority of attribute detection focused on the shape, position,
and spatial dimensions of windows. Additionally, seven papers focused
on detecting the status of windows to predict energy consumption or
heat usage. Three papers focused on categorizing windows by type to
enhance the semantic enrichment of BIM.
The outcomes of AWD were integrated with 3D reconstruction techniques, such as shape grammar and scan-to-BIM, resulting in a 3D
model. 2D polygons are less utilized in the storage stage, with only limited applications. Raw data formats were exclusively used for storage
purposes in 30 articles, as the primary requirement was to infer higherlevel data from the existence of windows, such as building ﬂoors, point
cloud generation, and energy monitoring [45,59,60].
In the execution layer, window detection methods included 55 studies using rule-based approaches, while 50 studies employed deep learning. The majority of windows on building facades are characterized by a
high degree of repeatability and symmetry. Therefore, the predominant
and more established method utilizes these rules to detect the existence
of windows, which can be applied to various data types based on the
extracted primitives (e.g., lines, patches) [26,61]. Deep learning methods can be categorized into instance segmentation, semantic segmentation, localization, classiﬁcation, and generative models, depending on
the data inputs and application scenarios of AWD. The main advantage
of deep learning is that it can recognize objects in scenes in cases such
as occlusion and irregular windows [45,51].
The application of AWD encompasses six topics in the requirement
layer: building information extraction, heritage conservation, robotics,
damage simulation, semantic enrichment, and energy simulation. The
majority of applications were related to semantic enrichment for 3D
building models (totaling 50), which was addressed in both the design
phase and the O&M phase. The remaining applications pertained to the
O&M phase, except for robotics, which was exclusively utilized in the
construction phase. This suggests that AWD applications play a crucial
role in the O&M phase compared to other phases. Energy simulation was
the second most prevalent type of application, with a total of 25 papers,
which is reasonable given that the presence of windows signiﬁcantly
inﬂuences the thermal performance of buildings [57].

3.1. Overview of 105 AWD publications
A total of 105 records of AWD technological pipelines were collected,
of which the values of the seven components of the ‘sandwich’ model
were manually encoded. Fig. 4 shows the origin cities with climate zone
of the 105 publications. The majority of papers had original cities in
the temperate zone (totaling 78), potentially because of diverse building typologies and high-density population densities in these regions.
The most prominent cluster is concentrated in the temperate zone of
Europe. Arid, tropical, and cold zones contained 15, 8, and 4 papers,
respectively. No studies originated from polar zones, possibly because
of the small population size and low-rise, low-density buildings.
Fig. 5 presents a Sankey chart of the 105 technical pipelines, highlighting the frequency and relationships among the coded components.
The rectangles containing numbers within each subcategory indicate
the total number of publications. Scenarios related to construction and
building technology (C&BT) account for 60–over half of the 105 papers
on AWD workﬂows, whereas computer science (CS)-centric scenarios
contribute approximately 11% of the methods. This contrast demonstrates that domain-speciﬁc datasets and application scenarios are crucial to the success of AWD. Additionally, researchers from remote sensing and computer science have primarily focused on applications in the
operations and maintenance (O&M) phase, while publications from the
C&BT domain span the entire building lifecycle, from design and construction to O&M.
The input data sources can be classiﬁed into four categories: 2D
(e.g., RGB camera, drawings), 3D (e.g., LiDAR point clouds with or
without RGB information, and BIM), thermal (infrared thermography
[IRT]), and multi-source data (e.g., combinations of IRT and RGB camera). Among these, the most frequently used data sources were LiDAR
and RGB cameras, which supported a range of applications from semantic enrichment modeling to energy simulation. Drawings and BIM, as
manually created data, were primarily utilized for semantic enrichment
modeling in the O&M and design phases [28,35].
5

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 5. Sankey diagram of the relationships between Scenario, Phase, Application, Data Source, Method, Attribute, and Storage.

Fig. 6. General workﬂow of AWD.

3.2. Generalized workﬂow of AWD

conservation, and construction robotics, have remained relatively limited in number. The diversity of these application types highlights the
signiﬁcant potential of AWD within the AEC industry.
The trend of the input data sources for AWD is illustrated in Fig. 7b.
LiDAR point clouds have been the most popular data source for semantic enrichment modeling since the ﬁrst paper in 2009, due to their ability to capture the three-dimensional characteristics of windows. Prior
to 2022, the use of RGB cameras in AWD applications was less prevalent compared to LiDAR point clouds. However, after 2022, there was
a signiﬁcant increase in the use of RGB cameras, as evidenced by a rise
in related applications, ranging from semantic enrichment modeling to
energy simulation. Since 2021, IRT has been utilized for energy simulation, as reported in up to seven papers. In contrast, the use of drawings,
although considerably niche, has remained steady at approximately one
publication per year since 2014.
The yearly statistics and trend of applied methods, i.e., rule-based,
semantic segmentation, instance segmentation, localization, classiﬁcation, and generative adversarial networks, are demonstrated in Fig. 7c.
Before 2018, rule-based methods were the predominant method. For
instance, techniques such as utilizing symmetry and repetitiveness to
identify buildings in aerial images [65], employing top-down rules inference to parse facade components [66], and detecting holes and edges of
windows based on extracted primitives [49] were commonly adopted.
Rule-based methods continue to be widely used after 2018, but focus
more on 3D input data sources (19 out of 28). This shift was attributed
to the scarcity and high cost of annotations of 3D training datasets for
deep learning, in contrast to the abundance of publicly available 2D
datasets.
Since 2018, with the advancement of computer vision technology,
the application of deep learning in AWD has increased signiﬁcantly.

Fig. 6 shows the general workﬂow of AWD. Before detection, preprocessing steps are essential to tailor the process to the speciﬁc attributes of interest. For rule-based methods, the pre-processing of the 2D
and 3D data sources diﬀers slightly. 2D images, typically captured from
a human perspective of street view, require perspective correction to enhance the accuracy of AWD results [60]. 3D-based methods commonly
involve noise ﬁltering and plane detection as standard pre-processing
procedures [62]. Some 3D methods may also integrate 2D projection
since 2D-based methods are more mature. After pre-processing, the window detection algorithm is applied. Rule-based methods typically extract primitives such as lines and patches, while learning-based methods
directly utilize the preprocessed data. Once the windows are extracted,
they are further processed or analyzed to satisfy speciﬁc application requirements, such as 3D model reconstruction or window-to-wall ratio
calculation.
3.3. Current trends
Fig. 7a illustrates the trend of AWD applications in the requirement
layer. It is evident that AWD’s use in semantic enrichment modeling, including BIM and CIM, has consistently attracted the most attention in the
sector. This was primarily because semantic enrichment models provide
signiﬁcant value for subsequent research, such as energy simulation and
building asset management [63,64]. The growth trajectory of semantic
enrichment modeling slowed after 2018, while energy simulation has
exhibited a more obvious upward trend. This trend shift underscores
the increasing emphasis on the urban thermal eﬀect in current literature. The remaining applications, such as damage simulation, heritage
6

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 7. Trend of 105 publications from 2009 to 2025. (a) Application, (b) Data Source, (c) Method, (d) Attribute.

Initially, the classiﬁcation methods, such as 2D convolutional neural
networks, were employed to identify and classify unknown objects as
windows [28]. Semantic segmentation methods emerged as the mainstream approach since 2021 alongside rule-based methods, because they
provide accurate pixel-level window detection results and have been applied in 26 studies [45]. In summary, there has been an increased interest in utilizing deep learning to detect windows using 2D input data
sources.
Fig. 7d demonstrates the trend of attributes for AWD. Despite the
continued popularity of 3D data sources for AWD, research focused on
detecting window dimensions has been relatively scarce in the past ﬁve
years. Only six papers were published during this period, addressing
3D thermal modeling, indoor building quality measurement, and energy
simulation [48,64,67]. Shapes and positions have consistently been the
primary attributes in this ﬁeld. These ﬁndings underscore the inherently
two-dimensional characteristics of windows in urban buildings.
In summary, AWD is being increasingly applied in building life-cycle
management, consistent with the concept of DTB. The growing number
of related publications and the diversity of application trends underscore the potential for developing novel applications of AWD. Input data
sources are becoming increasingly diverse. Learning-based methods are
gradually supplanting rule-based methods in the execution layer.

model. To support a target application scenario, one should carefully determine the three key factors–namely, data source, attributes, and storage format–within the data and information layer. Three decision trees
for guiding the selection of the factors are illustrated and explained as
follows.
Fig. 9 illustrates the process of selecting an input data source. The
primary recommended input data sources include RGB cameras, LiDAR
point clouds, and drawings. LiDAR point clouds are primarily recommended for applications that require highly accurate 3D positioning and
geometric information in the construction and O&M phases [67–69].
RGB cameras are recommended for energy simulation, damage assessment, and building information extraction, especially during the O&M
phase. They provide high-resolution, vivid color images that can capture
a window’s real-time position, shape, and status. Therefore, for applications that only require calculating window-related indicators or directly
classifying window status, 2D imagery is a cost-eﬀective choice [23,59].
One is recommended to use BIM, drawings, or IRT for updating the semantic information of BIM components, semantic enrichment modeling
during the design phase, and heat monitoring, respectively, as discussed
in the literature [28,70,71].
Fig. 10 illustrates the window attributes to be prioritized in AWD.
The typical detection attributes include shape, position, and status.
Detecting windows’ 2D/3D shapes is suﬃcient and recommended for
most applications, as indicated by the default ‘otherwise’ branches.
There are two main exceptions. In energy simulation applications,
status is employed to calculate window opening behaviors [72] and
shape is used to estimate building energy consumption [73]. For heritage conservation during the O&M phase and semantic enrichment
modeling at the design stage, detection of window position is recommended. However, if 3D reconstruction is required, shape detection should be prioritized [74]. Additionally, if higher-level semantic information is needed (e.g., window type), other studies should be
consulted [22,37,75].
Fig. 11 displays the selection of a storage plan for the AWD system.
Since 2009, the primary storage format has been 3D models, as indicated
by the default ‘otherwise’ branches in Fig. 11. This trend is expected to

4. Guidelines for AWD and further application
4.1. A three-step guideline
Fig. 8 presents the three-step guidelines for AWD proposed in this
paper. The guidelines were developed primarily based on the technical
workﬂows reviewed in the 105 publications. Additional moderation was
made by the authors to enhance clarity and to incorporate complementary aspects that were insuﬃciently addressed in the existing literature.
Step 1: To conﬁrm the data and information ‘meat’ layer to support
the target application. The data and information ‘meat’ layer should effectively (correctly and accurately) and eﬃciently (less cost and waste)
support the top application layer, according to the proposed sandwich
7

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 8. Three-step guidelines for selecting a posteriori AWD systems.

Fig. 9. Decision tree for AWD data source selection.

Fig. 10. Decision tree for AWD attributes selection.

continue due to the extensive application support oﬀered by semantically rich BIM and CIM models [25]. Additionally, 2D polygons and
raw formats are recommended in damage simulation, robotics, and energy simulation applications. For example, ﬂood risk possibilities can be
precisely calculated based on extracted window 2D polygons [9]; real-

time robot movement paths can be detected [27]; and window opening
behavior can be monitored using raw results [76].
Step 2: To build the ‘base’ execution layer. The method in the execution layer can be determined based on the chosen data source and
storage plan in Step 1. The guidelines for this step notably incorporate
8

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 11. Decision tree for AWD storage selection.

Fig. 12. Decision tree for AWD method selection.

the authors’ moderation. In the existing studies, rule-based methods account for 55 out of 105, which is more than 52%. However, learningbased methods have pushed the boundaries of computer vision [77] and
have become increasingly popular in the AEC industry over the past ﬁve
years (47 out of 63), as shown in 7c. Therefore, the guidelines below
give more credit to the learning-based methods.
Fig. 12 shows the decision tree for the method to execute. Based on
the data source and storage plan, the default recommendation method
is learning-based. The recommendations for data sources of drawings
and IRT vary according to the storage plan. More speciﬁcally, the use of
rule-based methods was advocated for both 2D polygon and 3D models,
whereas for raw format, the recommendation was to utilize learningbased methods. The recommended methods for LiDAR and RGBD cameras include both rule-based and learning-based. The ﬁnal choice between rule- or learning-based methods may vary depending on the availability of training datasets (e.g., Ma et al. [27], Wei et al. [78]) and the
complexity of the input data sources.
In general, rule-based methods oﬀer a practical choice when there is
a limited amount of training data. The edges of windows have stronger
image gradients compared to other facade elements, which can be utilized as a 2D-based data source [79]. Besides, pattern recognition based
on similarity and symmetry can eﬀectively detect windows [65]. Using
rule-based methods with 3D-based inputs is more challenging than with
2D-based inputs. This is because the invalid returns and reﬂections of
LiDAR data result in holes in the scanned facades [80]. Nevertheless,
several rule-based methods can still be applied to detect windows in 3D
data. For example, splitting the 3D point cloud into slices to estimate
the boundary points of windows [81], utilizing ray casting [82] to identify open areas in 3D space [83], and projecting 3D data onto 2D image
spaces followed by applying edge detection algorithms [84] to identify
window edges [52].
Learning-based methods have been well-proven for 2D data sources.
The existing AWD methods can be categorized into localization, se-

mantic segmentation, and instance segmentation, ranging from coarse
to ﬁne. The localization-based methods (e.g., Faster-RCNN [85],
YOLO [86]) can simultaneously provide the window category [22] and
its position(i.e., bounding box) on the facade. However, such methods
are insuﬃcient for representing the detailed geometry of windows. For
example, if the facade images are aﬀected by perspective projection,
rectangular windows may be represented inaccurately. The main reason is that the parallel edges of windows undergo perspective projection
and intersect at an invisible vanishing point [87]. In contrast, semantic
segmentation-based methods (e.g., U-Net [88], Deeplab [89]) can more
precisely estimate the pixels belonging to windows, though they may
fail to distinguish between individual adjacent windows. By combining
the above two approaches, instance segmentation-based methods (e.g.,
Mask-RCNN [90], SOLOv2 [91]) not only estimate the detailed geometry of windows but also identify distinct window instances by parsing
the window mask image.
The further selection for a speciﬁc method from the learning-based
category varies with the level of requirements and window attributes.
Localization-based methods are suitable for studies that need to monitor
the status of window openings (e.g., open, closed, or half-open) [72]
and analyze the relationship between window opening behavior and
environmental factors [23]. Semantic segmentation-based methods are
viable approaches to address tasks such as calculating the window-towall ratio, which do not require explicit estimation of the position and
shape of windows [26]. However, if one wants reverse engineering on
the facade [92], instance segmentation-based methods would be a better
choice [56,93]. Similarly, for both vectorized and rasterized drawings,
instance segmentation-based methods [94–96] can eﬀectively provide
accurate information on the position and geometric details of window
symbols.
Step 3: To implement and apply the constructed AWD. The third
step is to realize and apply the AWD workﬂow. The primary principle
of sensor placement is to obtain the best coverage of target windows,
9

Energy & Buildings 360 (2026) 117384

S. Meng et al.

4.2. Validation

such as capturing adequate lighting for the RGB camera and conducting
multiple scans for the LiDAR. Sensor carriers encompass vehicles for
large-scale low-rise building windows; drones for large-scale outdoor
windows; Internet of Things devices for monitoring window/building
status; and surveyors for individual buildings/scenes data collection [67,
97,98]. The dataset for training and execution can be obtained either
by utilizing oﬀ-the-shelf datasets or by annotating using tools, such as
labelImg [72].
The execution layer’s learning-based methods can often be obtained
from oﬃcial algorithm repositories, such as GitHub or HuggingFace,
which typically contain pre-trained learning models and auto-pipelines.
Training for 2D data sources can be conducted locally because of the
relatively low computational cost, whereas LiDAR-based training is recommended to utilize online GPU rental services to accommodate higher
computational demands. One should carefully examine the target attributes before selecting a performance metric from classiﬁcation metrics (primarily for instance segmentation), mean Intersection over Union
(for semantic segmentation), and error metrics (commonly reported in
rule-based methods).
Applying AWD results with post-processing DTB applications may
need further transformation. For heritage conservation and semantic
enrichment modeling, windows function as digital assets directly [73].
In energy simulation, raw window data facilitate the calculation of
window-to-wall ratios by orientation and climate-based daylight metrics [99]. Some scenes parsing applications convert detected results into
graphs to incorporate human knowledge [65] or classify room types
[100]. Therefore, the detected results must be further transformed according to the ﬁnal intended use. The detected results can be stored on
local/cloud disks or in BIM based on the transformed results.

Table 2 presents the validation results for the four decision trees. The
validation set shown in Table A.2 comprised 21 AWD papers from Scopus but excluded by Web of Science; so the validation set contained no
items in the training data. The 21 papers pertain to the DTB life-cycle
phases, too. The overall accuracy was satisfactory, with a mean accuracy of 88.1%, precision of 92.1%, and F1-score of 80.3%. Therefore,
the proposed guidelines and decision trees are reliable and robust for
practitioners conducting AWD and DTB applications.
Table 2
Validation results of the guidelines consisting of four decision trees on 21
more AWD papers from Scopus.
Metric

Data Source

Attribute

Storage

Method

Average

Accuracy (%)
Precision (%)
Recall (%)
F1-score (%)

85.7
88.9
92.3
89.0

81.0
88.9
71.4
73.8

85.7
90.7
60.0
58.4

100.0
100.0
100.0
100.0

88.1
92.1
80.9
80.3

The majority of cases were correctly classiﬁed. Some focused on
building energy in the O&M stage across ﬁve DTB applications. For example, for enhanced building energy simulation, Kim et al. [101] utilized aerial imagery to detect windows and store them as LoD-3 3DModel, as shown in Fig. 13d.
In data source selection, three cases mismatched our guidelines. Two
papers aimed at building information extraction in the construction
phase, and one for heritage conservation in the O&M phase [102–104].

Fig. 13. Demonstrative case of decision trees. (a) decision tree for data source; (b) decision tree for storage; (c) decision tree for method; (d) ﬁnal storage illustration
of building energy model [101].
10

Energy & Buildings 360 (2026) 117384

S. Meng et al.

A reason was the training data had no camera-based cases for construction monitoring. This gap highlights the diversity, growing demand, and
importance of AWD.
The decision tree for attribute selection achieved the lowest accuracy (81.0%) and F1-score (58.4%), primarily because the position information derived from localization-based methods remains relatively
infrequent across the 105 publications, especially in energy simulation
[101,105]. Further, three misclassiﬁed cases in storage selection predominantly involved semantic enrichment in the O&M phase, damage
simulation in the O&M phase, and building information extraction in the
construction phase [102,106,107]. These misclassiﬁed cases, nevertheless, provide valuable examples that enrich the potential data sources,
attributes, and storage options available to non-expert layman users.

2) In the Data and Information layer: Lack of multimodal data fusion,
limited 3D dataset availability, low reusability of detected results,
and insuﬃcient contextual information regarding the window.
The detection accuracy of AWD is limited due to its reliance on
a single input data source and corresponding limitations (e.g., occlusion in 2D, glass reﬂection in 3D, and distortion in photorealistic
CIM). The existing literature does not suﬃciently address window
detection through data fusion [27]. Additionally, current research
primarily focuses on AWD utilizing 2D input; however, there is a
lack of available datasets for 3D semantic segmentation of windows
using deep learning techniques. The reusability of detected AWD results should be improved to facilitate the broader adoption of AWD.
Finally, the absence of supplementary contextual information about
the window, such as the associated unit and ﬂoor, poses a challenge
for accurately locating the window within the DTB application.
3) In the Execution layer: Low robustness and applicability across window types and dataset sizes, lack of zero- or few-shot learning, and
high computational costs.
The robustness of execution methods requires improvement. For
instance, both rule-based and learning-based approaches struggle to
reliably extract windows when window features (e.g., coordinates,
color) are incomplete. The applicability of these methods is further
challenged by small sample sizes and imbalanced data distributions,
such as windows in rural areas or with uncommon shapes and styles.
Furthermore, reducing the amount of required training data and minimizing dependence on domain expertise are necessary for advancing DTB applications. Finally, current methods are limited in terms
of computational eﬃciency and power consumption in window detection, which constrains the deployment of large-scale window detection platforms, such as drones.

5. Discussion
5.1. Comprehensive review of AWD
5.1.1. General observations
The diversity of DTB application categories does not inherently reﬂect a corresponding level of innovation in AWD techniques. Speciﬁcally, most studies have focused on the end-use aspects of DTB applications. However, AWD methodologies are often repeatedly applied or
yield only incremental advancements.
Input data sources are typically limited to unimodal and small-scale
datasets. Most studies primarily utilize a single modality for feature extraction, with any additional sources serving only as supplementary information, although multiple data sources are discussed in the literature.
Furthermore, input data are frequently limited to small-scale samples,
rather than being applied at the scale of urban buildings.
The literature predominantly emphasizes attributes intrinsic to the
window itself, rather than contextual information, such as the associated
unit, ﬂoor level, or material properties. The predominant storage format
is three-dimensional, which is favored due to its portability and versatility. In addition, several proprietary models, such as building energy
models augmented with window-based energy simulation data, have
been developed to address speciﬁc, niche applications [64].
Deep learning, as an emerging technology, is increasingly replacing traditional rule-based methods. These learning-based approaches are
primarily supervised and necessitate carefully annotated datasets. Additionally, it is noteworthy that the execution methods utilized in AWD
tend to lag behind the most recent advancements in computer science,
such as vision-language models.
The validation of detected windows is contingent upon the execution methods utilized and the role of windows within DTB applications.
Notably, 29% of studies did not report window-level evaluation metrics,
indicating that window detection is frequently regarded as a secondary
outcome in DTB applications. Nevertheless, AWD results continue to
impact the assessment of overall DTB application performance, as reﬂected by metrics such as the root-mean-square error of model distance
in semantic enrichment modeling.

5.1.3. Future directions
Fig. 14 illustrates three groups of future directions for AWD, based
on the ﬁndings from the literature analysis, practical guidelines, the conceptual model, and the identiﬁed challenges.
Mapping emerging needs and demands: The inﬂuence of windows
on urban residents is increasing due to the rapid and vertical expansion of urban environments. Drones are expected to signiﬁcantly
aﬀect the daily lives of city dwellers in the near future [108]. Window views, noise, privacy perception, and real-time instance-level
window operation represent potential DTB applications aimed at
measuring and enhancing residents’ comfort [109–111]. In addition,
building facade health monitoring and renovation also depend on the
digitization of windows [60]. These emerging directions can broaden
the requirements layer and further expand DTB applications.
• Cross-style window detection: Windows across diﬀerent building
types, such as commercial and rural buildings, represent feasible
detection targets. For instance, commercial buildings are typically
characterized by glass curtain walls with small openable panels. In
contrast, rural buildings may exhibit irregular window patterns that
are challenging for current execution methods to detect. Accurate
detection of windows across diverse styles will enhance DTB applications, including assessments of energy consumption and indoor
thermal insulation [112].
• Crowd-sourced city-wide cloud platform of windows: Although
CityGML has established standards for various levels of detail (LoD)
in building modeling, these standards are diﬃcult to modify and
have limited widespread adoption and application [113]. Advanced
technologies for semantic mapping, geometry reconstruction, and
specialized software integration can be incorporated into such cloud
platforms. Additionally, data sharing protocols will facilitate largescale integration and reusability of AWD results, promoting the adoption of DTBs throughout the entire building lifecycle.
• Multi-source, multi-modal data fusion: Both 2D and 3D window
features oﬀer unique advantages. For instance, 2D data captures con•

5.1.2. Challenges
This section highlights the common issues and challenges identiﬁed
in AWD across the reviewed studies, as informed by the proposed conceptual model.
1) In the Requirement layer: Limited to individual buildings, lack of
cross-style detection, and dependence on domain expertise.
The DTB applications of AWD are mainly focused on individual buildings. However, comprehensive window digital assets at the
city scale are required to support data-driven analysis and residentcentric applications at a ﬁne-grained level. Moreover, few studies
have investigated window detection across diﬀerent building types
and regions. The creation of AWD relies on domain expert knowledge, semantic annotation, and algorithm execution, which increases
the challenge of broadly adopting AWD in DTB applications.
11

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Fig. 14. Potential future directions of AWD.

tinuous and complete glass surfaces, while 3D data provides distinct
three-dimensional characteristics, such as window sills and frames.
Integrating both modalities through deep learning and 3D Gaussian
splatting may represent an eﬀective approach for enhancing window
detection.
• Version alignment of data sources and window assets: The same
window assets across diﬀerent data sources (e.g., photogrammetry
models from Google Earth and governmental databases) may exhibit
discrepancies, such as incomplete features in one source while being
complete in another. Aligning the detected window versions from
various platforms can yield more comprehensive and accurate window detection results.
• Encourage open-source 3D benchmark datasets: High-quality 3D
benchmark datasets for window detection are needed. Such datasets
will facilitate baseline model training and the evaluation of execution methods, thereby supporting the development of new algorithms and reducing the learning curve for non-experts.
• Reconstruction of CIM and BIM models with high LoD: High LoD
CIM or BIM models, with window details, serve as the foundational
geometry of DTB applications [114]. Moreover, in BIM’s automated
modeling, the low LoD of building components hampers building
simulations across the entire lifecycle [115]. Thus, developing automated and cost-eﬀective methods for high LoD CIM/BIM modeling
is necessary to promote AWD and DTB applications.
• Extraction of contextual information for windows: Diﬀerent
building data sources encompass a range of information, from intrinsic attribute data to external contextual information. For example, photorealistic CIMs represent the up-to-date and vivid appearances of building facades. In contrast, architectural drawings depict indoor layouts but are less digitized and may be misaligned
with the constructed buildings. The automated integration of window information based on indoor and outdoor data sources (such
as drawings and photorealistic CIM) represents a promising research avenue for achieving comprehensive semantic enrichment of
windows.
• Adoption of advanced window detection methods: New deep
learning models are emerging with the development of computer

vision. The hybrid method can consist of the following potential directions to improve the performance of window detection results:
(i) Utilizing up-to-date 2D semantic segmentation models, such as
Segment Anything [116]; (ii) Introducing vision-language models to
assist in identifying errors in post-data processing steps, because of
the multi-modality nature and large-scale pretraining on existing human digitalized knowledge; (iii) Utilizing interactive labeling to implement few-shot learning to dynamically improve the model’s segmentation capabilities, given the similarity of windows on a single
building [117].
• Improvement of window detection methods based on domain
knowledge: The 3D deep learning model architecture can be improved by domain-characteristic-based adjustments, such as loss
function design based on symmetry and repeatability, synthetic
data generation and training, and contrastive learning to enhance
3D scene understanding [118]. Additionally, 3D scene completion can be used to mitigate the eﬀects of occlusion and glass
reﬂections.
• Development of low-cost, low-power algorithms: The large-scale
window detection requires devices such as drones to collect citywide window information. Considering the high computational cost
of current execution methods, algorithms with low computational
cost and power consumption are needed.

5.2. The sandwich model and guidelines
5.2.1. Implications
The sustainable development of building life-cycle projects is increasingly emphasized across many ﬁelds. Window detection plays a
crucial role in various building life-cycle activities (e.g., design, construction, and management) that align with the Sustainable Development Goals (SDGs). The growing number of new buildings underscores
the necessity of Automated Window Detection (AWD) at both the microscale (e.g., indoor facilities management) and the macroscale (e.g.,
urban energy consumption), with attributes being essential for both indoor and outdoor windows.
12

Energy & Buildings 360 (2026) 117384

S. Meng et al.

A systematic analysis of the current literature was conducted based
on coded classiﬁcation results, general workﬂows, and emerging trends.
This paper further proposes a three-step guideline for newcomers
aiming to develop applications based on AWD results. Four decision
trees–covering data sources, attributes, storage plans, and methods–
demonstrate suitable options for users addressing speciﬁc application
tasks. This guideline, combined with the conceptual model, reveals potential connections between the three layers and seven components of
application-driven AWD. It also provides new insights relevant to ﬁelds
such as remote sensing, building engineering, building energy consumption, and urban planning.

To the best of our knowledge, this paper presents the ﬁrst comprehensive review of AWD applications in the AEC-related industry and
provides full coverage of the literature from 2009 to 2025. It delineates
the current status, prevailing trends, and key developments that inform
the selection of critical factors for AWD-based applications. The proposed three-step guidelines oﬀer a practical framework for future researchers and industry practitioners seeking to implement AWD-based
solutions. Recommended future research directions include: investigating novel application domains, enhancing window detection across diverse building typologies, developing cloud-based platforms, integrating
multi-source data fusion, harmonizing heterogeneous data versions, promoting the development of 3D public datasets, reﬁning high LoD CIM
and BIM models, expanding window information repositories, adopting
innovative detection methodologies, and advancing 3D learning through
domain-speciﬁc knowledge. This review is limited to literature published up to 2025 and relies exclusively on sources indexed in the Web of
Science and Scopus databases and written in English. It is recommended
that systematic reviews be conducted every ﬁve years to capture the
evolving landscape of AWD research and validate its applicability across
various stages of the building life cycle.

5.2.2. Limitations and future work
This paper has several limitations that should be addressed in future research. First, the literature review covers publications only up
to December 2025. Given the rapid evolution of large models, regular
updates–ideally every ﬁve years–are recommended to keep pace with
advancements in the ﬁeld. Second, the literature search was limited
to the Web of Science and Scopus databases, which may introduce selection bias. Third, only papers written in English were included, potentially overlooking signiﬁcant contributions from non-Anglophone research communities. Although the applications span the entire building
life cycle, there is a disproportionate focus on studies related to the Operation and Maintenance (O&M) phase, which limits insights into the integration of design and construction phases using the proposed decision
trees. The impact of rapid technological advancement on the conceptual
model and guidelines needs to be assessed. Future research should validate the robustness and accuracy of the proposed generic conceptual
model and guidelines across diverse building life-cycle scenarios and
applications in diﬀerent regions.

CRediT authorship contribution statement
Siyuan Meng: Writing – original draft, Visualization, Software,
Methodology, Data curation, Conceptualization; Longyong Wu: Writing
– original draft, Visualization, Software, Data curation; Maosu Li: Writing – review & editing, Visualization; Anthony G. O. Yeh: Resources,
Project administration, Writing - review & editing; Fan Xue: Writing –
review & editing, Supervision, Resources, Project administration, Funding acquisition, Conceptualization.
Data availability

6. Conclusion

No data was used for the research described in the article.

The digitalization of building assets provides a foundational platform
for enabling applications throughout the building lifecycle in alignment
with the Sustainable Development Goals (SDGs). Windows, as integral
elements of the building envelope, play a pivotal role in functions such
as air exchange and heat transfer. Consequently, accurate window detection is essential for deriving high-level building information, estimating energy consumption, and monitoring human activities, among other
applications. Although numerous studies have explored advanced applications that leverage automated window detection (AWD) results, the
nature of the relationship between these applications and AWD remains
insuﬃciently addressed.
Therefore, this paper establishes a clear deﬁnition of automated window detection (AWD) and introduces a conceptual ‘sandwich’ model
comprising three layers–requirement as ‘top’, data and information as
‘meat’, and execution as ‘base’–and six components: application, phase,
data source, attribute, storage, and method. Utilizing the PRISMA standard and the proposed ‘sandwich’ model, 105 relevant publications were
systematically collected, analyzed, and evaluated to assess a posteriori
status and trends of each component. Furthermore, a set of three-step
guidelines is presented to assist prospective users: (1) selecting the input
data source, attributes of interest, and information storage strategy; (2)
determining the appropriate method to employ; and (3) processing and
transforming the results. The four decision trees achieved an average
accuracy of 88.1% in the validation set, pinpointing the reliability and
robustness of our proposed guidelines and decision trees.

Declaration of interests
The authors declare that they have no known competing ﬁnancial
interests or personal relationships that could have appeared to inﬂuence
the work reported in this paper.
Declaration of generative AI in writing
The authors declare the following ways of interactions with generative AI which may be considered as potential uses in writing or editing
the content in this paper:
GenAI-assisted proofreading. A university self-hosted GenAI LLM
(ChatGPT-4.1) was used to assist with proofreading the work to eliminate grammatical and connection errors in the main text, and to improve
readability of Abstract, Introduction, and Conclusion sections. We declare no use of GenAI to produce any new content in the work.
Acknowledgments
The work presented in this paper was supported by the Hong Kong
Research Grants Council (Nos. 17201325, T22-504/21-R), and in part
by the Department of Science and Technology of Guangdong Province
(GDST) (No. 2020B1212030009).

13

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Appendix A. Literature database

Table A.1
List of 105 cases of application-driven AWD.
No.

Reference

Scenarios∗

Phase†

Application‡

Data Source§

Method

Attribute

Storage

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69

Becker [66]
Chung et al. [65]
Meixner and Leberl [53]
Meixner and Leberl [119]
Truong-Hong et al. [120]
Demisse et al. [121]
Dumitru et al. [122]
Truong-Hong et al. [123]
Tuttas and Stilla [21]
Wang et al. [124]
Xiong et al. [83]
Riedinger et al. [125]
Frommholz et al. [126]
Hong et al. [127]
Liu and Qin [128]
Sadeghi et al. [129]
Arnaud et al. [130]
Colleu and Benitez [131]
Díaz-Vilariño et al. [132]
Gimenez et al. [133]
Jung et al. [134]
Lachat et al. [135]
Nguatem et al. [136]
Paravolidakis et al. [137]
Valero et al. [52]
Oskouie et al. [22]
Ruwanthika et al. [35]
Bacharidis et al. [74]
Chen et al. [138]
Franz et al. [139]
Hao et al. [49]
Jung et al. [140]
Lu et al. [75]
Previtali et al. [141]
Wu et al. [142]
Cui et al. [143]
Kim et al. [28]
Masiero et al. [68]
Shi et al. [62]
Tang et al. [144]
Van Ackere et al. [51]
Zheng et al. [79]
Adán et al. [50]
Dochev et al. [36]
Yin et al. [70]
Arjoune et al. [73]
Fan et al. [145]
Pexman et al. [61]
Touzani et al. [97]
Vor̆ís̆ek et al. [146]
Wang et al. [147]
Feng et al. [9]
Golombek and Marshall [54]
Gu et al. [55]
˛
Jarzabek-Rychard
and Maas [148]
Jiang et al. [149]
Li et al. [150]
Luong et al. [31]
Murtiyoso et al. [59]
Pantoja-Rosero et al. [151]
Sun et al. [23]
Szcześniak et al. [26]
Tien et al. [57]
Wang et al. [38]
Wysocki et al. [152]
Xiao et al. [37]
Zhang and Aliaga [45]
Zhao et al. [100]
Adán et al. [64]

R.S.
C.S.
R.S.
C.&B. T.
C.&B. T.
C.S.
R.S.
R.S.
R.S.
C.&B. T.
C.&B. T.
C.S.
R.S.
C.&B. T.
R.S.
R.S.
C.S.
R.S.
R.S.
C.&B. T.
C.&B. T.
R.S.
R.S.
C.S.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
R.S.
R.S.
R.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
R.S.
R.S.
R.S.
C.S.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
R.S.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
C.S.
C.&B. T.
C.&B. T.

O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
Design
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
Design
O&M
O&M
O&M
O&M
O&M
O&M
O&M
Design
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
Con.
O&M
O&M
O&M
O&M
O&M

Semantic Enr.
Building Info.
Semantic Enr.
Building Info.
Semantic Enr.
Energy Sim.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Energy Sim.
Semantic Enr.
Heritage Cons.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Building Info.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Heritage Cons.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Heritage Cons.
Semantic Enr.
Semantic Enr.
Damage Sim.
Energy Sim.
Semantic Enr.
Energy Sim.
Semantic Enr.
Energy Sim.
Semantic Enr.
Heritage Cons.
Energy Sim.
Semantic Enr.
Energy Sim.
Damage Sim.
Building Info.
Damage Sim.
Semantic Enr.
Semantic Enr.
Energy Sim.
Energy Sim.
Building Info.
Semantic Enr.
Energy Sim.
Energy Sim.
Energy Sim.
Robotics
Semantic Enr.
Building Info.
Building Info.
Building Info.
Energy Sim.

LiDAR
LiDAR
RGB Camera
RGB Camera
LiDAR
LiDAR
LiDAR
LiDAR
LiDAR
LiDAR
LiDAR
Drawings
RGB Camera
LiDAR
LiDAR-RGB
LiDAR
RGB-D
LiDAR
LiDAR
Drawings
LiDAR
RGB-D
LiDAR
RGB Camera
LiDAR
RGB Camera
Drawings
RGB Camera
LiDAR
RGB-D
LiDAR
LiDAR
RGB Camera
LiDAR
Drawings
LiDAR
BIM
LiDAR
LiDAR
RGB-D
LiDAR-RGB
RGB Camera
LiDAR
RGB Camera
Drawings
IRT
LiDAR
LiDAR
RGB Camera
LiDAR
RGB Camera
LiDAR
LiDAR
RGB Camera
RGB Camera
RGB Camera
IRT
RGB Camera
RGB Camera
RGB Camera
RGB Camera
RGB Camera
RGB Camera
LiDAR
LiDAR
BIM
RGB Camera
Drawings
LiDAR-IRT

Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Rule
Classiﬁcation
Rule
Rule
Rule
Classiﬁcation
Rule
Rule
Rule
Instance
Rule
Rule
Rule
Rule
Instance
Rule
Rule
Semantic
Rule
Semantic
Instance
Rule
GAN
Instance
Semantic
Semantic
Rule
Semantic
Semantic
Classiﬁcation
Rule
Localization
Rule
Semantic
Classiﬁcation
Semantic
Semantic
Rule

Shape
Shape
Position
Position
Shape
Status
Position
Shape
Shape
Shape
Shape
Position
Shape
Shape
Shape
Shape
Position
Position
Shape
Position
Shape
Shape
Shape
Shape
Shape
Type
Position
Shape
Shape
Shape
Position
Shape
Shape
Position
Position
Shape
Type
Position
Shape
Shape
Dimension
Status
Shape
Position
Shape
Shape
Shape
Position
Shape
Position
Shape
Shape
Shape
Shape
Shape
Shape
Shape
Status
Shape
Shape
Status
Position
Status
Shape
Shape
Type
Shape
Shape
Dimension

3D-Model
3D-Model
Raw
3D-Model
3D-Model
Raw
Raw
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
2D-Polygon
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
2D-Polygon
3D-Model
Raw
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
Raw
3D-Model
3D-Model
Raw
3D-Model
3D-Model
2D-Polygon
Raw
3D-Model
3D-Model
3D-Model
Raw
3D-Model
3D-Model
3D-Model
3D-Model
Raw
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
Raw
Raw
Raw
3D-Model
Raw
3D-Model
Raw
Raw
3D-Model
3D-Model
Raw
Raw
3D-Model

14

Energy & Buildings 360 (2026) 117384

S. Meng et al.

Table A.1
Continue.
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105

Biswanath et al. [44]
Johns et al. [29]
Liu et al. [72]
Lu et al. [60]
Walter et al. [71]
Wei et al. [78]
Zeng et al. [69]
Drobnyi et al. [47]
Forth et al. [30]
Gu et al. [98]
Knechtel et al. [43]
Ma et al. [27]
Ogawa et al. [56]
Pérez and Sánchez [41]
Surojaya et al. [96]
Tan et al. [67]
Wang et al. [76]
Xia and Gong [32]
Yoon et al. [153]
Hu et al. [154]
Wang et al. [155]
Chen et al. [156]
Chang et al. [157]
Su et al. [158]
Yoon et al. [159]
Suppa et al. [40]
Xie et al. [160]
Mirzabeigi et al. [161]
Liang et al. [46]
Ariss et al. [39]
Zhang et al. [48]
Yavariabdi et al. [162]
Hany et al. [163]
Fan et al. [164]
Zhao et al. [42]
Mehraban et al. [165]

R.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
C.S.
C.S.
R.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
C.S.
C.S.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
C.&B. T.
C.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.

O&M
Con.
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
Con.
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M

Energy Sim.
Robotics
Energy Sim.
Energy Sim.
Energy Sim.
Semantic Enr.
Semantic Enr.
Building Info.
Energy Sim.
Damage Sim.
Semantic Enr.
Robotics
Semantic Enr.
Heritage Cons.
Damage Sim.
Building Info.
Energy Sim.
Building Info.
Energy Sim.
Building Info.
Semantic Enr.
Building Info.
Semantic Enr.
Energy Sim.
Energy Sim.
Energy Sim.
Building Info.
Energy Sim.
Semantic Enr.
Damage Sim.
Energy Sim.
Damage Sim.
Semantic Enr.
Damage Sim.
Semantic Enr.
Semantic Enr.

IRT
IRT
RGB Camera
RGB Camera
IRT
LiDAR
LiDAR
RGB-D
LiDAR-RGB
RGB Camera
Drawings
LiDAR-RGB
RGB Camera
RGB Camera
RGB Camera
LiDAR
RGB Camera
RGB Camera
RGB Camera
RGB Camera
RGB Camera
RGB Camera
Drawings
RGB Camera
RGB Camera
RGB Camera
RGB Camera
IRT
LiDAR
RGB Camera
RGB Camera
RGB Camera
LiDAR
IRT
LiDAR
RGB Camera

Rule
Rule
Localization
Instance
Semantic
Semantic
Rule
Rule
Semantic
Semantic
Semantic
Semantic
Instance
Instance
Instance
Rule
Localization
Localization
Localization
Semantic
Semantic
Localization
Semantic
Semantic
Localization
Localization
Semantic
Semantic
Semantic
Localization
Localization
Semantic
Semantic
Semantic
Localization
Semantic

Shape
Shape
Status
Position
Position
Shape
Shape
Shape
Shape
Shape
Position
Shape
Shape
Shape
Position
Dimension
Status
Position
Shape
Shape
Shape
Position
Dimension
Shape
Shape
Dimension
Position
Dimension
Shape
Shape
Dimension
Position
Shape
Position
Shape
Shape

3D-Model
Raw
2D-Polygon
Raw
Raw
3D-Model
3D-Model
3D-Model
3D-Model
3D-Model
Raw
Raw
3D-Model
3D-Model
2D-Polygon
3D-Model
Raw
2D-Polygon
3D-Model
Raw
3D-Model
Raw
3D-Model
3D-Model
3D-Model
2D-Polygon
Raw
Raw
Raw
3D-Model
3D-Model
Raw
3D-Model
Raw
3D-Model
3D-Model

∗ RS = Remote Sensing, CS = Computer Science, C&BT = Construction & Building Technology
†O&M = Operations and Management, Con. = Construction
‡Enr. = Enrichment, Info. = Information extraction, Sim. = Simulation, Cons. = Conservation
§IRT = Infrared Temperature Sensor

Table A.2
List of 21 cases for validation collected from Scopus.
No.

Reference

Scenarios∗

Phase†

Application‡

Data Source§

Method

Attribute

Storage

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

Zhu et al. [166]
Lu et al. [167]
Pan et al. [168]
Xing et al. [106]
Yu et al. [169]
Haznedar et al. [170]
Pantoja-Rosero et al. [171]
Chua and Cheah [102]
Dai et al. [172]
Escalada [173]
Fotsing et al. [174]
Li et al. [175]
Mehranfar et al. [176]
Ro and Gong [107]
Chowdhury et al. [105]
Duan et al. [177]
Hou et al. [104]
Lai et al. [103]
Zhong et al. [178]
Forouzandeh et al. [179]
Kim et al. [101]

C.S.
C.S.
C.&B. T.
C.&B. T.
R.S.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.
C.&B. T.

O&M
Design
O&M
Design
O&M
O&M
O&M
Con.
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
O&M
Con.
O&M
O&M
O&M

Semantic Enr.
Semantic Enr.
Semantic Enr.
Building Info.
Building Info.
Heritage Cons.
Damage Sim.
Building Info.
Building Info.
Building Info.
Semantic Enr.
Semantic Enr.
Semantic Enr.
Damage Sim.
Energy Sim.
Semantic Enr.
Heritage Cons.
Building Info.
Building Info.
Semantic Enr.
Energy Sim.

RGB Camera
Drawings
LiDAR-RGB
Drawings
LiDAR
LiDAR-RGB
RGB Camera
RGB Camera
RGB Camera
RGB Camera
LiDAR-RGB
LiDAR
RGB Camera
RGB Camera
RGB Camera
RGB Camera
RGB Camera
RGB Camera
RGB Camera
LiDAR
RGB Camera

Semantic
Semantic
Semantic
Localization
Localization
Semantic
Semantic
Localization
Semantic
Semantic
Rule
Rule
Localization
Semantic
Localization
Localization
Semantic
Localization
Semantic
Semantic
Localization

Shape
Position
Shape
Position
Position
Shape
Shape
Shape
Shape
Shape
Shape
Shape
Shape
Shape
Position
Position
Shape
Position
Shape
Shape
Position

Raw
3D-Model
3D-Model
Raw
Raw
Raw
3D-Model
2D-Polygon
Raw
Raw
3D-Model
3D-Model
3D-Model
Raw
3D-Model
3D-Model
Raw
Raw
Raw
3D-Model
3D-Model

∗ RS = Remote Sensing, CS = Computer Science, C&BT = Construction & Building Technology
†O&M = Operations and Management, Con. = Construction
‡Enr. = Enrichment, Info. = Information extraction, Sim. = Simulation, Cons. = Conservation
§IRT = Infrared Temperature Sensor

15

Energy & Buildings 360 (2026) 117384

S. Meng et al.

References

[25] F. Xue, L. Wu, W. Lu, Semantic enrichment of building and city information models:
a ten-year review, Adv. Eng. Inf. 47 (2021) 101245. https://doi.org/10.1016/j.aei.
2020.101245
[26] J.T. Szcześniak, Y.Q. Ang, S. Letellier-Duchesne, C.F. Reinhart, A method for using street view imagery to auto-extract window-to-wall ratios and its relevance
for urban-level daylighting and energy simulations, Build. Environ. 207 (2022)
108108. https://doi.org/10.1016/j.buildenv.2021.108108
[27] J. Ma, X. Jiang, D. Zheng, X. Liao, J. Lu, Y. Zhao, TransWallNet: high-performance
semantic segmentation of large-scale and multifeatured point clouds of building
gables, J. Constr. Eng. Manage. 150 (8) (2024) 04024092. https://doi.org/10.
1061/JCEMD4.COENG-14827
[28] J. Kim, J. Song, J.-K. Lee, Recognizing and classifying unknown object in
BIM using 2D CNN, in: J.-H. Lee (Ed.), Computer-Aided Architectural Design.
"Hello, Culture", Springer, Singapore, 2019, pp. 47–57. https://doi.org/10.1007/
978-981-13-8410-3_4
[29] B. Johns, E. Abdi, M. Arashpour, Crane payload localisation for curtain wall installation: a markerless computer vision approach, Measurement 221 (2023) 113459.
https://doi.org/10.1016/j.measurement.2023.113459
[30] K. Forth, F. Noichl, A. Borrmann, LCA calculation of retroﬁtting scenarios using geometric model reconstruction and semantic enrichment of point clouds
and images, in: Computing in Civil Engineering 2023, 2023, pp. 390–397. https:
//doi.org/10.1061/9780784485231.047
[31] D. Luong, R. Richman, M. Touchie, Towards window state detection using image
processing in residential and oﬃce building facades, Build. Environ. 207 (2022)
108486. https://doi.org/10.1016/j.buildenv.2021.108486
[32] J. Xia, J. Gong, Computer vision based ﬁrst ﬂoor elevation estimation from mobile LiDAR data, Autom. Constr. 159 (2024) 105258. https://doi.org/10.1016/j.
autcon.2023.105258
[33] Z. Zou, K. Chen, Z. Shi, Y. Guo, J. Ye, Object detection in 20 years: a survey, Proc.
IEEE 111 (3) (2023) 257–276. https://doi.org/10.1109/JPROC.2023.3238524
[34] S. Yoon, Building digital twinning: data, information, and models, J. Build. Eng.
76 (2023) 107021. https://doi.org/10.1016/j.jobe.2023.107021
[35] R.G.N. Ruwanthika, P.A.D.B.M. Amarasekera, R.U.I.B. Chandrasiri, D.M.A.I. Rangana, A. Nugaliyadde, Y. Mallawarachchi, Dynamic 3D model construction using
architectural house plans, in: 2017 6th National Conference on Technology and
Management (NCTM), 2017, pp. 181–184. https://doi.org/10.1109/NCTM.2017.
7872850
[36] I. Dochev, P. Gorzalka, V. Weiler, J. Estevam Schmiedt, M. Linkiewicz, U. Eicker,
B. Hoﬀschmidt, I. Peters, B. Schröter, Calculating urban heat demands: an analysis
of two modelling approaches and remote sensing for input data and validation, Energy Build. 226 (2020) 110378. https://doi.org/10.1016/j.enbuild.2020.110378
[37] M. Xiao, Z. Chao, R.F. Coelho, S. Tian, Investigation of classiﬁcation and anomalies
based on machine learning methods applied to large scale building information
modeling, Appl. Sci. 12 (13) (2022) 6382. https://doi.org/10.3390/app12136382
[38] Y. Wang, L. Xie, H. Wang, W. Zeng, Y. Ding, T. Hu, T. Zheng, H. Liao, J. Hu,
Intelligent spraying robot for building walls with mobility and perception, Autom.
Constr. 139 (2022) 104270. https://doi.org/10.1016/j.autcon.2022.104270
[39] M. Ariss, B.G. Pantoja-Rosero, F. Duarte, M. Klimenka, C. Ratti, Seismic assessment of unreinforced masonry façades from images using macroelementbased modeling, Commun. Eng. 4 (1) (2025) 155. https://doi.org/10.1038/
s44172-025-00487-2
[40] A.R. Suppa, A. Aliberti, M.C. Bottero, V. Corrado, Detecting window-to-wall ratio
for urban-scale building simulations using deep learning with street view imagery
and an automatic classiﬁcation algorithm, Build. Simul. 18 (8) (2025) 2175–2199.
https://doi.org/10.1007/s12273-025-1301-3
[41] F.N. Pérez, I.B. Sánchez, Automatic detection and 3D reconstruction of buildings
from historical maps, Appl. Artif. Intell. 38 (1) (2024) 2378274. https://doi.org/
10.1080/08839514.2024.2378274
[42] T. Zhao, T. Xiong, M. Li, Z. Li, Automatic reconstruction of 3D building models
from ALS point clouds based on façade geometry, ISPRS Int. J. Geo-Inf. 14 (12)
(2025). https://doi.org/10.3390/ijgi14120462
[43] J. Knechtel, P. Rottmann, J.-H. Haunert, Y. Dehbi, Semantic ﬂoorplan segmentation
using self-constructing graph networks, Autom. Constr. 166 (2024) 105649. https:
//doi.org/10.1016/j.autcon.2024.105649
[44] M.K. Biswanath, L. Hoegner, U. Stilla, Thermal mapping from point clouds to 3D
building model facades, Remote Sens. 15 (19) (2023) 4830. https://doi.org/10.
3390/rs15194830
[45] X. Zhang, D. Aliaga, RFCNet: enhancing urban segmentation using regularization,
fusion, and completion, Comput. Vision Image Understanding 220 (2022) 103435.
https://doi.org/10.1016/j.cviu.2022.103435
[46] X. Liang, Y. Liu, J. Wang, Z. Qin, L. Chen, H. Jia, C. Dai, Automatic 3D reconstruction of indoor building point clouds under occlusion based on gestalt rules, Int.
J. Digital Earth 18 (1) (2025) 2545582. https://doi.org/10.1080/17538947.2025.
2545582
[47] V. Drobnyi, S. Li, I. Brilakis, Digitization of existing buildings with arbitrary shaped
spaces from point clouds, J. Comput. Civ. Eng. 38 (5) (2024) 04024027. https:
//doi.org/10.1061/JCCEE5.CPENG-5853
[48] C. Zhang, P.-J. Hoes, B. Tian, R. Zhang, R. Loonen, Enhancing neighborhood-scale
building performance simulation through building classiﬁcation and automated
data acquisition: Supporting the Dutch heating transition, Sustainable Cities Soc.
132 (2025) 106811. https://doi.org/10.1016/j.scs.2025.106811
[49] W. Hao, Y. Wang, W. Liang, Slice-based building facade reconstruction from 3D
point clouds, Int. J. Remote Sens. 39 (20) (2018) 6587–6606. https://doi.org/10.
1080/01431161.2018.1463113
[50] A. Adán, B. Quintana, S.A. Prieto, F. Bosché, An autonomous robotic platform for
automatic extraction of detailed semantic models of buildings, Autom. Constr. 109

[1] M. Lotteau, P. Loubet, M. Pousse, E. Dufrasnes, G. Sonnemann, Critical review of
life cycle assessment (LCA) for the built environment at the neighborhood scale,
Build. Environ. 93 (2015) 165–178. https://doi.org/10.1016/j.buildenv.2015.06.
029
[2] A. Baranzini, C. Schaerer, A sight for sore eyes: Assessing the value of view and
land use in the housing market, J. Hous. Econ. 20 (3) (2011) 191–199. https://doi.
org/10.1016/j.jhe.2011.06.001
[3] R.K. Raanaas, G.G. Patil, T. Hartig, Health beneﬁts of a view of nature through
the window: a quasi-experimental study of patients in a residential rehabilitation center, Clin. Rehabil. 26 (1) (2012) 21–32. https://doi.org/10.1177/
0269215511412800
[4] L. Wang, S. Greenberg, Window operation and impacts on building energy consumption, Energy Build. 92 (2015) 313–321. https://doi.org/10.1016/j.enbuild.
2015.01.060
[5] C. Rubino, S. Liuzzi, G. Fusaro, F. Martellotta, C. Scrosati, M. Garai, Balancing
ventilation and sound insulation in windows by means of metamaterials: a review
of the state of the art, Build. Environ. (2025) 112780. https://doi.org/10.1016/j.
buildenv.2025.112780
[6] F. Abd-Alhamid, M. Kent, Y. Wu, Quantifying window view quality: a review
on view perception assessment and representation methods, Build. Environ. 227
(2023) 109742. https://doi.org/10.1016/j.buildenv.2022.109742
[7] Y. Chen, N. Li, X. Tian, X. Hu, Improving thermal comfort and working performance in windowless environment: relieving thermal response by introducing artiﬁcial view windows, Energy Build. 348 (2025) 116444. https://doi.org/10.1016/
j.enbuild.2025.116444
[8] J. Souviron, G. Van Moeseke, A.Z. Khan, Analysing the environmental impact of
windows: a review, Build. Environ. 161 (2019) 106268. https://doi.org/10.1016/
j.buildenv.2019.106268
[9] Y. Feng, Q. Xiao, C. Brenner, A. Peche, J. Yang, U. Feuerhake, M. Sester, Determination of building ﬂood risk maps from LiDAR mobile mapping data, Comput. Environ. Urban Syst. 93 (2022) 101759. https://doi.org/10.1016/j.compenvurbsys.
2022.101759
[10] UNDESA, Synergy Solutions for Climate and SDG Action: Bridging the Ambition
Gap for the Future we Want, Technical Report, United Nations, Department of
Economic and Social Aﬀairs, 2024. https://sdgs.un.org/sites/default/ﬁles/202408/UN%20Synergy%20Solutions%20for%20Climate%20and%20SDG%20Action4_0.pdf#page=58.08.
[11] S. Kosse, P. Hagedorn, M. König, Semantic digital twins in construction: developing
a modular system reference architecture based on information containers, Adv.
Eng. Inf. 67 (2025) 103483. https://doi.org/10.1016/j.aei.2025.103483
[12] Q. Lu, A.K. Parlikad, P. Woodall, G. Don Ranasinghe, X. Xie, Z. Liang, E. Konstantinou, J. Heaton, J. Schooling, Developing a digital twin at building and city levels:
case study of West Cambridge campus, J. Manage. Eng. 36 (3) (2020) 05020004.
https://doi.org/10.1061/(ASCE)ME.1943-5479.0000763
[13] L. Su, H. Yan, Y. Sun, D. Cao, X. Luo, Deciphering the trajectory and challenges of
BIM research in China: a data-driven review of topics and innovation trends, Adv.
Eng. Inf. 68 (2025) 103567. https://doi.org/10.1016/j.aei.2025.103567
[14] K. You, K. Chen, F. Xue, Automated scan-to-BIM for construction digital transformation: Conceptual framework, processing methods and best-practice guidelines, Autom. Constr. 182 (2026) 106755. https://doi.org/10.1016/j.autcon.2025.
106755
[15] P. Pauwels, R. de Koning, B. Hendrikx, E. Torta, Live semantic data from building
digital twins for robot navigation: Overview of data transfer methods, Adv. Eng.
Inf. 56 (2023) 101959. https://doi.org/10.1016/j.aei.2023.101959
[16] I.C. Akyol, Y.M. Karadag, S. Ucar, I. Talaz, F.E. Gursoy, I.G. Dino, S. Kalkan, Transfer learning and parameter-eﬃcient ﬁne-tuning for heating energy consumption
prediction using urban building energy models (UBEM), Adv. Eng. Inf. 68 (2025)
103576. https://doi.org/10.1016/j.aei.2025.103576
[17] W. Xie, Y. Zeng, X. Zhang, H.Y. Wong, T. Zhang, Z. Wang, X. Wu, J. Shi, X. Huang,
F. Xiao, et al., AIoT-powered building digital twin for smart ﬁreﬁghting and super
real-time ﬁre forecast, Adv. Eng. Inf. 65 (2025) 103117. https://doi.org/10.1016/
j.aei.2025.103117
[18] I. Acosta, M.a. Campano, J.F. Molina, Window design in architecture: analysis of
energy savings for lighting and visual comfort in residential spaces, Appl. Energy
168 (2016) 493–506. https://doi.org/10.1016/j.apenergy.2016.02.005
[19] A. Arsiwala, F. Elghaish, M. Zoher, Digital twin with machine learning for predictive monitoring of CO2 equivalent from existing buildings, Energy Build. 284
(2023) 112851. https://doi.org/10.1016/j.enbuild.2023.112851
[20] K. Guerrero Ramírez, C. Nuevo-Gallardo, J.M. Santamaría Ulecia, B. Montalbán Pozas, C. Fernández Bandera, Digital twin implementation based on a whitebox building energy model: a case study on blind control for passive heating, Energy Build. 349 (2025) 116454. https://doi.org/10.1016/j.enbuild.2025.116454
[21] S. Tuttas, U. Stilla, Reconstruction of façades in point clouds from multi aspect
oblique ALS, ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci. II-3-W3
(2013) 91–96. https://doi.org/10.5194/isprsannals-II-3-W3-91-2013
[22] P. Oskouie, B. Becerik-Gerber, L. Soibelman, Automated recognition of building
façades for creation of as-is mock-up 3D models, J. Comput. Civ. Eng. 31 (6) (2017)
04017059. https://doi.org/10.1061/(ASCE)CP.1943-5487.0000711
[23] C. Sun, X. Guo, T. Zhao, Y. Han, Real-time detection method of window opening behavior using deep learning-based image recognition in severe cold regions, Energy
Build. 268 (2022) 112196. https://doi.org/10.1016/j.enbuild.2022.112196
[24] A. Klimkowska, S. Cavazzi, R. Leach, S. Grebby, Detailed three-dimensional building façade reconstruction: a review on applications, data and technologies, Remote
Sens. 14 (11) (2022) 2579. https://doi.org/10.3390/rs14112579
16

Energy & Buildings 360 (2026) 117384

S. Meng et al.

(2020) 102963. https://doi.org/10.1016/j.autcon.2019.102963
[51] S. Van Ackere, J. Verbeurgt, L. De Sloover, A. De Wulf, N. Van de Weghe, P.
De Maeyer, Extracting dimensions and locations of doors, windows, and door
thresholds out of mobile lidar data using object detection to estimate the impact
of ﬂoods, Int. Ann. Photogramm. Remote Sens. Spatial Inf. Sci. XLII-3-W8 (2019)
429–436. https://doi.org/10.5194/isprs-archives-XLII-3-W8-429-2019
[52] E. Valero, A. Adán, F. Bosché, Semantic 3D reconstruction of furnished interiors
using laser scanning and RFID technology, J. Comput. Civ. Eng. 30 (4) (2016)
04015053. https://doi.org/10.1061/(ASCE)CP.1943-5487.0000525
[53] P. Meixner, F. Leberl, Interpreting Building Façades from Vertical Aerial Images
Using the Third Dimension, Int. Ann. Photogramm. Remote Sens. Spatial Inf. Sci.
38 (4) (2010).
[54] Y. Golombek, W. Marshall, High-density mobile LiDAR for measuring urban
streetscape features, URBAN DESIGN International 27 (1) (2022) 3–17. https:
//doi.org/10.1057/s41289-021-00176-4
[55] D. Gu, W. Chen, X. Lu, Automated assessment of wind damage to windows of
buildings at a city scale based on oblique photography, deep learning and CFD, J.
Build. Eng. 52 (2022) 104355. https://doi.org/10.1016/j.jobe.2022.104355
[56] Y. Ogawa, R. Nakamura, G. Sato, H. Maeda, Y. Sekimoto, End-to-end framework
for the automatic matching of omnidirectional street images and building data
and the creation of 3D building models, Remote Sens. 16 (11) (2024) 1858. https:
//doi.org/10.3390/rs16111858
[57] P.W. Tien, S. Wei, J.K. Calautit, J. Darkwa, C. Wood, Real-time monitoring of occupancy activities and window opening within buildings using an integrated deep
learning-based approach for reducing energy demand, Appl. Energy 308 (2022)
118336. https://doi.org/10.1016/j.apenergy.2021.118336
[58] Y.Y. Song, Y. Lu, Decision tree methods: applications for classiﬁcation and prediction, Shanghai Arch. Psychiatry 27 (2) (2015) 130–135. https://doi.org/10.11919/
j.issn.1002-0829.215044
[59] A. Murtiyoso, E. Pellis, P. Grussenmeyer, T. Landes, A. Masiero, Towards semantic photogrammetry: generating semantically rich point clouds from architectural
close-range photogrammetry, Sensors 22 (3) (2022) 966. https://doi.org/10.3390/
s22030966
[60] Y. Lu, W. Wei, P. Li, T. Zhong, Y. Nong, X. Shi, A deep learning method for building
façade parsing utilizing improved SOLOv2 instance segmentation, Energy Build.
295 (2023) 113275. https://doi.org/10.1016/j.enbuild.2023.113275
[61] K. Pexman, D.D. Lichti, P. Dawson, Automated storey separation and door and
window extraction for building models from complete laser scans, Remote Sens.
13 (17) (2021) 3384. https://doi.org/10.3390/rs13173384
[62] W. Shi, W. Ahmed, N. Li, W. Fan, H. Xiang, M. Wang, Semantic geometric modelling
of unstructured indoor point cloud, ISPRS Int. J. Geo-Inf. 8 (1) (2018) 9. https:
//doi.org/10.3390/ijgi8010009
[63] M. Al-Kasasbeh, O. Abudayyeh, H. Liu, An integrated decision support system for
building asset management based on BIM and work breakdown structure, J. Build.
Eng. 34 (2021) 101959. https://doi.org/10.1016/j.jobe.2020.101959
[64] A. Adán, A. Ramón, J.L. Vivancos, A. Vilar, C. Aparicio-Fernández, Automatic
generation of as-is BEM models of buildings, J. Build. Eng. 73 (2023) 106865.
https://doi.org/10.1016/j.jobe.2023.106865
[65] Y.-C. Chung, T.X. Han, Z. He, Building recognition using sketch-based representations and spectral graph matching, in: 2009 IEEE 12th International Conference
on Computer Vision, 2009, pp. 2014–2020. https://doi.org/10.1109/ICCV.2009.
5459444
[66] S. Becker, Generation and application of rules for quality dependent façade reconstruction, ISPRS J. Photogramm. Remote Sens. 64 (6) (2009) 640–653. https:
//doi.org/10.1016/j.isprsjprs.2009.06.002
[67] Y. Tan, X. Liu, S. Jin, Q. Wang, D. Wang, X. Xie, A terrestrial laser scanning-based
method for indoor geometric quality measurement, Remote Sens. 16 (1) (2023)
59. https://doi.org/10.3390/rs16010059
[68] A. Masiero, F. Chiabrando, A.M. Lingua, B.G. Marino, F. Fissore, A. Guarnieri,
A. Vettore, 3D modeling of Girifalco fortress, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XLII-2-W9 (2019) 473–478. https://doi.org/10.5194/
isprs-archives-XLII-2-W9-473-2019
[69] R. Zeng, J.J.S. Shi, C. Wang, T. Lu, Integrating as-built BIM model from point
cloud data in construction projects, Eng. Constr. Archit. Manage. 31 (9) (2024)
3557–3574. https://doi.org/10.1108/ECAM-12-2022-1196
[70] M. Yin, L. Tang, T. Zhou, Y. Wen, R. Xu, W. Deng, Automatic layer classiﬁcation
method-based elevation recognition in architectural drawings for reconstruction
of 3D BIM models, Autom. Constr. 113 (2020) 103082. https://doi.org/10.1016/
j.autcon.2020.103082
[71] I. Walter, M. Tanasković, M. Stanković, IR building analysis with extraction of
elements using image segmentation and RetinaNet, Buildings 13 (1) (2022) 109.
https://doi.org/10.3390/buildings13010109
[72] Y. Liu, W.T. Chong, Y.H. Yau, D. Han, M. Qin, F. Deng, X. Wang, S. Pan, A new
algorithm for real-time detection of window opening area in residential buildings, Build. Environ. 244 (2023) 110817. https://doi.org/10.1016/j.buildenv.
2023.110817
[73] Y. Arjoune, S. Peri, N. Sugunaraj, A. Biswas, D. Sadhukhan, P. Ranganathan, An
instance segmentation and clustering model for energy audit assessments in built
environments: a multi-stage approach, Sensors 21 (13) (2021) 4375. https://doi.
org/10.3390/s21134375
[74] K. Bacharidis, F. Sarri, V. Paravolidakis, L. Ragia, M. Zervakis, Fusing georeferenced and stereoscopic image data for 3D building façade reconstruction, ISPRS
Int. J. Geo-Inf. 7 (4) (2018) 151. https://doi.org/10.3390/ijgi7040151
[75] Q. Lu, S. Lee, L. Chen, Image-driven fuzzy-based system to construct as-is IFC BIM
objects, Autom. Constr. 92 (2018) 68–87. https://doi.org/10.1016/j.autcon.2018.
03.034

[76] Z. Wang, H. Tang, H. Zhang, D. Jiang, Y. Leng, Y. Wu, J. Calautit, L. Xia,
H.N. Chaudhry, Multi-objectives occupant-centric control of thermostats and natural ventilation systems in cold climate conditions using real-time occupantrelated information, Build. Environ. 263 (2024) 111882. https://doi.org/10.1016/
j.buildenv.2024.111882
[77] N. ÓMahony, S. Campbell, A. Carvalho, S. Harapanahalli, G.V. Hernandez, L. Krpalkova, D. Riordan, J. Walsh, Deep learning vs. traditional computer vision, in:
Advances in Computer Vision: Proceedings of the 2019 Computer Vision Conference (CVC), Volume 1 1, Springer, 2020, pp. 128–144. https://doi.org/10.1007/
978-3-030-17795-9_10
[78] J. Wei, H. Wu, H. Yue, S. Jia, J. Li, C. Liu, Automatic extraction and reconstruction of a 3D wireframe of an indoor scene from semantic point clouds, Int. J.
Digital Earth 16 (1) (2023) 3239–3267. https://doi.org/10.1080/17538947.2023.
2246937
[79] H. Zheng, F. Li, H. Cai, K. Zhang, Non-intrusive measurement method for the window opening behavior, Energy Build. 197 (2019) 171–176. https://doi.org/10.
1016/j.enbuild.2019.05.052
[80] T.S. Fong, W.Y. Yan, Signs on glasses: LiDAR data voids, hotspot eﬀect, and reﬂection artifacts, Autom. Constr. 169 (2025) 105877. https://doi.org/10.1016/j.
autcon.2024.105877
[81] S.M.I. Zolanvari, D.F. Laefer, A.S. Natanzi, Three-dimensional building façade segmentation and opening area detection from point clouds, ISPRS J. Photogramm. Remote Sens. 143 (2018) 134–149. https://doi.org/10.1016/j.isprsjprs.2018.04.004
[82] S.D. Roth, Ray casting for modeling solids, Comput. Graphics Image Process. 18
(2) (1982) 109–144. https://doi.org/10.1016/0146-664X(82)90169-1
[83] X. Xiong, A. Adan, B. Akinci, D. Huber, Automatic creation of semantically rich
3D building models from laser scanner data, Autom. Constr. 31 (2013) 325–337.
https://doi.org/10.1016/j.autcon.2012.10.006
[84] J. Canny, A computational approach to edge detection, IEEE Trans. Pattern Anal.
Mach. Intell. (6) (1986) 679–698. https://doi.org/10.1109/TPAMI.1986.4767851
[85] S. Ren, K. He, R. Girshick, J. Sun, Faster R-CNN: towards real-time object detection with region proposal networks, IEEE Trans. Pattern Anal. Mach. Intell. 39 (6)
(2016) 1137–1149. https://doi.org/10.1109/TPAMI.2016.2577031
[86] J. Redmon, You only look once: uniﬁed, real-time object detection, in: Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp.
779–788. https://doi.org/10.1109/CVPR.2016.91
[87] C. Rother, A new approach to vanishing point detection in architectural environments, Image Vis. Comput. 20 (9-10) (2002) 647–655. https://doi.org/10.1016/
S0262-8856(02)00054-9
[88] O. Ronneberger, P. Fischer, T. Brox, U-Net: convolutional networks for biomedical
image segmentation, in: International Conference on Medical Image Computing
and Computer-assisted Intervention, Springer, 2015, pp. 234–241. https://doi.org/
10.1007/978-3-319-24574-4_28
[89] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A.L. Yuille, Deeplab: semantic
image segmentation with deep convolutional nets, atrous convolution, and fully
connected crfs, IEEE Trans. Pattern Anal. Mach. Intell. 40 (4) (2017) 834–848.
https://doi.org/10.1109/TPAMI.2017.2699184
[90] K. He, G. Gkioxari, P. Dollár, R. Girshick, Mask R-CNN, in: Proceedings of the
IEEE International Conference on Computer Vision, 2017, pp. 2961–2969. https:
//doi.org/10.1109/ICCV.2017.322
[91] X. Wang, R. Zhang, T. Kong, L. Li, C. Shen, SOLOv2: dynamic and fast instance
segmentation, Adv. Neural Inf. Process. Syst. 33 (2020) 17721–17732. https://
doi.org/10.5555/3495724.3497211
[92] A.D. Styliadis, Historical photography-based computer-aided architectural design:
demolished buildings information modeling with reverse engineering functionality, Autom. Constr. 18 (1) (2008) 51–69. https://doi.org/10.1016/j.autcon.2008.
04.010
[93] Y. Sun, S. Malihi, H. Li, M. Maboudi, Deepwindows: windows instance segmentation through an improved mask R-CNN using spatial attention and relation modules, ISPRS Int. J. Geo-Inf. 11 (3) (2022) 162. https://doi.org/10.3390/
ijgi11030162
[94] S. Dou, X. Jiang, L. Liu, L. Ying, C. Shan, Y. Shen, X. Dong, Y. Wang, D. Li, C. Zhao,
Hierarchical recognizing vector graphics and a new chart-based vector graphics
dataset, IEEE Trans. Pattern Anal. Mach. Intell. (2024). https://doi.org/10.1109/
TPAMI.2024.3394298
[95] Y. Zhao, X. Deng, H. Lai, Reconstructing BIM from 2D structural drawings for existing buildings, Autom. Constr. 128 (2021) 103750. https://doi.org/10.1016/j.
autcon.2021.103750
[96] A. Surojaya, N. Zhang, J.R. Bergado, F. Nex, Towards fully autonomous UAV:
damaged building-opening detection for outdoor-indoor transition in urban
search and rescue, Electronics 13 (3) (2024) 558. https://doi.org/10.3390/
electronics13030558
[97] S. Touzani, M. Wudunn, S. Fernandes, A. Zakhor, R. Najibi, J. Granderson, A machine learning approach to estimate windows-to-wall ratio using drone imagery,
in: N. Chrysoulakis, T. Erbertseder, Y. Zhang (Eds.), Remote Sensing Technologies
and Applications in Urban Environments VI, SPIE, Online Only, Spain, 2021, p. 20.
https://doi.org/10.1117/12.2602157
[98] D.L. Gu, Q.W. Shuai, N. Zhang, N. Jin, Z.X. Zheng, Z. Xu, Y.J. Xu, Multi-view street
view image fusion for city-scale assessment of wind damage to building clusters,
Comput.-Aided Civ. Infrastruct. Eng. 40 (2) (2025) 198–214. https://doi.org/10.
1111/mice.13324
[99] K.A. Al-Sallal, A.R. AbouElhamd, M.B. Dalmouk, UAE heritage buildings converted
into museums: evaluation of daylighting eﬀectiveness and potential risks on artifacts and visual comfort, Energy Build. 176 (2018) 333–359. https://doi.org/10.
1016/j.enbuild.2018.06.067

17

Energy & Buildings 360 (2026) 117384

S. Meng et al.

[100] X. Zhao, T.-H. Wang, C. Peng, Automatic room type classiﬁcation using machine
learning for two-dimensional residential building plans, in: Proceedings of the 40th
Conference on Education and Research in Computer Aided Architectural Design in
Europe (eCAADe 2022), 2, Ghent, Belgium, 2022, pp. 593–600. https://doi.org/
10.52842/conf.ecaade.2022.2.593
[101] Y. Kim, J. Yoon, S. Lee, M. Shin, Automated building geometry modeling using
UAV imagery and deep learning for energy simulation, J. Build. Eng. 117 (2026)
114949. https://doi.org/10.1016/j.jobe.2025.114949
[102] W.P. Chua, C.C. Cheah, Deep-learning-based automated building construction
progress monitoring for prefabricated preﬁnished volumetric construction, Sensors
24 (21) (2024). https://doi.org/10.3390/s24217074
[103] Y. Lai, Z. Lin, Z. Ye, J. Zhang, Development of advanced progress recognition
algorithms for construction monitoring, PLOS ONE 20 (10) (2025) e0333262.
https://doi.org/10.1371/journal.pone.0333262
[104] J. Hou, J. Zhou, Y. He, B. Hou, J. Li, Automatic reconstruction of semantic façade
model of architectural heritage, Heritage Sci. 12 (1) (2024) 400. https://doi.org/
10.1186/s40494-024-01506-9
[105] S. Chowdhury, S.N. Kandelan, K.S. Grewal, Harnessing machine learning for rapid
and cost-eﬃcient 3D geometry generation in neighborhood energy modeling, Results Eng. 28 (2025) 107599. https://doi.org/10.1016/j.rineng.2025.107599
[106] J. Xing, Q. Luo, Y. Wu, J. Shang, Self-optimization for parsing ﬂoor plans, J.
Comput. Civ. Eng. 36 (6) (2022) 04022037. https://doi.org/10.1061/(ASCE)CP.
1943-5487.0001048
[107] S.H. Ro, J. Gong, Scalable approach to create annotated disaster image
database supporting AI-driven damage assessment, Nat. Hazards 120 (13) (2024)
11693–11712. https://doi.org/10.1007/s11069-024-06641-x
[108] S. Xiang, A. Xie, M. Ye, X. Yan, X. Han, H. Niu, Q. Li, H. Huang, Autonomous
eVTOL: a summary of researches and challenges, Green Energy Intell. Transp. 3
(1) (2024) 100140. https://doi.org/10.1016/j.geits.2023.100140
[109] T. Van Renterghem, D. Botteldooren, View on outdoor vegetation reduces noise
annoyance for dwellers near busy roads, Landscape Urban Plann. 148 (2016)
203–215. https://doi.org/10.1016/j.landurbplan.2015.12.018
[110] M. Li, F. Xue, Y. Wu, A.G.O. Yeh, 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 Urban Plann. 226 (2022) 104505. https://doi.
org/10.1016/j.landurbplan.2022.104505
[111] M. Li, S. Guo, F. Duarte, A. Kumar, N. Kobori, F. Xue, W. Zhuang, A.G.O. Yeh,
C. Ratti, Inﬂuence of objective and perceived exposures to urban nature on people’s happiness, npj Urban Sustainability 6 (1) (2026) 6. https://doi.org/10.1038/
s42949-025-00306-9
[112] E. Cuce, S.B. Riﬀat, C.-H. Young, Thermal insulation, power generation, lighting
and energy saving performance of heat insulation solar glass as a curtain wall application in Taiwan: a comparative experimental study, Energy Convers. Manage.
96 (2015) 31–38. https://doi.org/10.1016/j.enconman.2015.02.062
[113] J. Shi, Z. Pan, L. Jiang, X. Zhai, An ontology-based methodology to establish city
information model of digital twin city by merging BIM, GIS and IoT, Adv. Eng. Inf.
57 (2023) 102114. https://doi.org/10.1016/j.aei.2023.102114
[114] B.G. Pantoja-Rosero, A. Rusnak, F. Kaplan, K. Beyer, Generation of LOD4 models for
buildings towards the automated 3D modeling of BIMs and digital twins, Autom.
Constr. 168 (2024) 105822. https://doi.org/10.1016/j.autcon.2024.105822
[115] S.-H. Chen, F. Xue, Automatic BIM detailing using deep features of 3D views, Autom. Constr. 148 (2023) 104780. https://doi.org/10.1016/j.autcon.2023.104780
[116] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S.
Whitehead, A.C. Berg, W.-Y. Lo, et al., Segment anything, in: Proceedings of the
IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026.
https://doi.org/10.1109/ICCV51070.2023.00371
[117] C. Lee, S. Park, H. Song, J. Ryu, S. Kim, H. Kim, S. Pereira, D. Yoo, Interactive
multi-class tiny-object detection, in: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2022, pp. 14136–14145. https://doi.
org/10.1109/CVPR52688.2022.01374
[118] S. Meng, X. Su, G. Sun, M. Li, F. Xue, From 3D pedestrian networks to wheelable
networks: an automatic wheelability assessment method for high-density urban areas using contrastive deep learning of smartphone point clouds, Comput. Environ.
Urban Syst. 117 (2025) 102255. https://doi.org/10.1016/j.compenvurbsys.2025.
102255
[119] P. Meixner, F. Leberl, 3-Dimensional building details from aerial photography
for internet maps, Remote Sens. 3 (4) (2011) 721–751. https://doi.org/10.3390/
rs3040721
[120] L. Truong-Hong, D.F. Laefer, T. Hinks, H. Carr, Flying voxel method with delaunay
triangulation criterion for façade/feature detection for computation, J. Comput.
Civ. Eng. 26 (6) (2012) 691–707. https://doi.org/10.1061/(ASCE)CP.1943-5487.
0000188
[121] G.G. Demisse, D. Borrmann, A. Nüchter, Interpreting thermal 3D models of indoor environments for energy eﬃciency, in: 2013 16th International Conference
on Advanced Robotics (ICAR), 2013, pp. 1–8. https://doi.org/10.1109/ICAR.2013.
6766550
[122] R.-C. Dumitru, D. Borrmann, A. Nüchter, Interior reconstruction using the 3D
Hough transform, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XL-5-W1
(2013) 65–72. https://doi.org/10.5194/isprsarchives-XL-5-W1-65-2013
[123] L. Truong-Hong, D.F. Laefer, T. Hinks, H. Carr, Combining an angle criterion with
voxelization and the ﬂying voxel method in reconstructing building models from
LiDAR data, Comput.-Aided Civ. Infrastruct. Eng. 28 (2) (2013) 112–129. https:
//doi.org/10.1111/j.1467-8667.2012.00761.x
[124] C. Wang, Y.K. Cho, M. Gai, As-is 3D thermal modeling for existing building envelopes using a hybrid LIDAR system, J. Comput. Civ. Eng. 27 (6) (2013) 645–656.
https://doi.org/10.1061/(ASCE)CP.1943-5487.0000273

[125] C. Riedinger, M. Jordan, H. Tabia, 3D models over the centuries: From old ﬂoor
plans to 3D representation, in: 2014 International Conference on 3D Imaging
(IC3D), 2014, pp. 1–8. https://doi.org/10.1109/IC3D.2014.7032583
[126] D. Frommholz, M. Linkiewicz, H. Meissner, D. Dahlke, A. Poznanska, Extracting semantically annotated 3D building models with textures from oblique aerial
imagery, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XL-3-W2 (2015)
53–58. https://doi.org/10.5194/isprsarchives-XL-3-W2-53-2015
[127] S. Hong, J. Jung, S. Kim, H. Cho, J. Lee, J. Heo, Semi-automated approach to indoor
mapping for 3D as-built building information modeling, Comput. Environ. Urban
Syst. 51 (2015) 34–46. https://doi.org/10.1016/j.compenvurbsys.2015.01.005
[128] Y. Liu, S. Qin, Facade model reﬁnement by fusing terrestrial laser data and image,
in: MIPPR 2015: Pattern Recognition and Computer Vision, 9813, SPIE, 2015, pp.
291–296. https://doi.org/10.1117/12.2214179
[129] F. Sadeghi, H. Areﬁ, A. Fallah, M. Hahn, 3D building façade reconstruction using handheld laser scanning data, Int. Arch. Photogramm. Remote
Sens. Spatial Inf. Sci. XL-1-W5 (2015) 625–630. https://doi.org/10.5194/
isprsarchives-XL-1-W5-625-2015
[130] A. Arnaud, J. Christophe, M. Gouiﬀes, M. Ammi, 3D reconstruction of indoor
building environments with new generation of tablets, in: Proceedings of the
22nd ACM Conference on Virtual Reality Software and Technology, VRST ’16,
Association for Computing Machinery, New York, NY, USA, 2016, pp. 187–190.
https://doi.org/10.1145/2993369.2993403
[131] T. Colleu, S. Benitez, Ray-based detection of openings in urban areas using mobile
LiDAR data, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XLI-B3 (2016)
591–598. https://doi.org/10.5194/isprs-archives-XLI-B3-591-2016
[132] L. Díaz-Vilariño, P. Boguslawski, K. Khoshelham, H. Lorenzo, L. Mahdjoubi, Indoor navigation from point clouds: 3D modelling and obstacle detection, Int.
Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XLI-B4 (2016) 275–281. https:
//doi.org/10.5194/isprs-archives-XLI-B4-275-2016
[133] L. Gimenez, S. Robert, F. Suard, K. Zreik, Automatic reconstruction of 3D building
models from scanned 2D ﬂoor plans, Autom. Constr. 63 (2016) 48–56. https://doi.
org/10.1016/j.autcon.2015.12.008
[134] J. Jung, S. Hong, S. Yoon, J. Kim, J. Heo, Automated 3D wireframe modeling of
indoor structures from point clouds using constrained least-squares adjustment for
as-built BIM, J. Comput. Civ. Eng. 30 (4) (2016) 04015074. https://doi.org/10.
1061/(ASCE)CP.1943-5487.0000556
[135] E. Lachat, T. Landes, P. Grussenmeyer, Combination of TLS point clouds and 3D
data from kinect V2 sensor to complete indoor models, Int. Arch. Photogramm.
Remote Sens. Spatial Inf. Sci. XLI-B5 (2016) 659–666. https://doi.org/10.5194/
isprs-archives-XLI-B5-659-2016
[136] W. Nguatem, M. Drauschke, H. Mayer, Automatic generation of building models
with levels of detail 1-3, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci.
XLI-B3 (2016) 649–654. https://doi.org/10.5194/isprsarchives-XLI-B3-649-2016
[137] V. Paravolidakis, K. Bacharidis, F. Sarri, L. Ragia, M. Zervakis, Reduction of
building façade model complexity using computer vision, in: 2016 IEEE International Conference on Imaging Systems and Techniques (IST), 2016, pp. 454–459.
https://doi.org/10.1109/IST.2016.7738269
[138] J. Chen, Y.K. Cho, K. Kim, Region proposal mechanism for building element recognition for advanced scan-to-BIM process, in: Construction Research Congress 2018,
2018, pp. 221–231. https://doi.org/10.1061/9780784481264.022
[139] S. Franz, R. Irmler, U. Rüppel, Real-time collaborative reconstruction of digital
building models with mobile devices, Adv. Eng. Inf. 38 (2018) 569–580. https:
//doi.org/10.1016/j.aei.2018.08.012
[140] J. Jung, C. Stachniss, S. Ju, J. Heo, Automated 3D volumetric reconstruction
of multiple-room building interiors for as-built BIM, Adv. Eng. Inf. 38 (2018)
811–825. https://doi.org/10.1016/j.aei.2018.10.007
[141] M. Previtali, L. Díaz-Vilariño, M. Scaioni, Indoor building reconstruction from occluded point clouds using graph-cut and ray-tracing, Appl. Sci. 8 (9) (2018) 1529.
https://doi.org/10.3390/app8091529
[142] Q. Wu, P. Meng, G. Liu, Reconstruction of 3D building model based on the information in ﬂoor plan, Traitem. Signal 35 (3-4) (2018) 303–316. https://doi.org/10.
3166/TS.35.303-316
[143] Y. Cui, Q. Li, B. Yang, W. Xiao, C. Chen, Z. Dong, Automatic 3-D reconstruction
of indoor environment with mobile laser scanning point clouds, IEEE J. Sel. Top.
Appl. Earth Obs. Remote Sens. 12 (8) (2019) 3117–3130. https://doi.org/10.1109/
JSTARS.2019.2918937
[144] S. Tang, Y. Zhang, Y. Li, Z. Yuan, Y. Wang, X. Zhang, X. Li, Y. Zhang, R. Guo,
W. Wang, Fast and automatic reconstruction of semantically rich 3D indoor maps
from low-quality RGB-D sequences, Sensors 19 (3) (2019) 533. https://doi.org/10.
3390/s19030533
[145] H. Fan, Y. Wang, J. Gong, Layout graph model for semantic façade reconstruction
using laser point clouds, Geo-spatial Inf. Sci. 24 (3) (2021) 403–421. https://doi.
org/10.1080/10095020.2021.1922316
[146] J. Vor̆ís̆ek, B. Patzák, E. Dvor̆áková, D. Rypl, Automated BIM entity reconstruction
from unstructured 3d pointclouds, Acta Polytech. CTU Proc. 30 (2021) 126–130.
https://doi.org/10.14311/APP.2021.30.0126
[147] C. Wang, S. Wei, S. Du, D. Zhuang, Y. Li, X. Shi, X. Jin, X. Zhou, A systematic method to develop three dimensional geometry models of buildings for urban building energy modeling, Sustainable Cities Soc. 71 (2021) 102998. https:
//doi.org/10.1016/j.scs.2021.102998
˛
[148] M. Jarzabek-Rychard,
H.G. Maas, Automatic enrichment of indoor 3D models using
a deep learning approach based on single images with unknown camera poses,
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci. VIII-5-W1-2022 (2022)
1–7. https://doi.org/10.5194/isprs-annals-VIII-5-W1-2022-1-2022
[149] Y. Jiang, S. Han, Y. Bai, Scan4Façade: automated as-is façade modeling of historic
high-rise buildings using drones and AI, J. Archit. Eng. 28 (4) (2022) 04022031.
18

Energy & Buildings 360 (2026) 117384

S. Meng et al.

https://doi.org/10.1061/(ASCE)AE.1943-5568.0000564
[150] D. Li, T. Yu, Y. Yang, Y. Zhao, Z. Chen, J. Liu, Intelligent heating monitoring method based on infrared image segmentation and target detection: poster
abstract, in: Proceedings of the 9th ACM International Conference on Systems
for Energy-Eﬃcient Buildings, Cities, and Transportation, BuildSys ’22, Association for Computing Machinery, New York, NY, USA, 2022, pp. 301–302. https:
//doi.org/10.1145/3563357.3567755
[151] B.G. Pantoja-Rosero, R. Achanta, M. Kozinski, P. Fua, F. Perez-Cruz, K. Beyer, Generating LOD3 building models from structure-from-motion and semantic segmentation, Autom. Constr. 141 (2022) 104430. https://doi.org/10.1016/j.autcon.2022.
104430
[152] O. Wysocki, E. Grilli, L. Hoegner, U. Stilla, Combining visibility analysis and deep
learning for reﬁnement of semantic 3D building models by conﬂict classiﬁcation, ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci. X-4-W2-2022 (2022)
289–296. https://doi.org/10.5194/isprs-annals-X-4-W2-2022-289-2022
[153] J. Yoon, Y. Kim, S. Lee, M. Shin, UAV-based automated 3D modeling framework using deep learning for building energy modeling, Sustainable Cities Soc. 101 (2024)
105169. https://doi.org/10.1016/j.scs.2023.105169
[154] Y. Hu, J. Wei, S. Zhang, S. Liu, FDIE: a graph-based framework for extracting design
information from annotated building facade images, J. Asian Archit. Build. Eng. 24
(4) (2025) 2530–2553. https://doi.org/10.1080/13467581.2024.2373835
[155] D. Wang, J. Liu, H. Jiang, P. Liu, Q. Jiang, Existing buildings recognition and BIM
generation based on multi-plane segmentation and deep learning, Buildings 15 (5)
(2025). https://doi.org/10.3390/buildings15050691
[156] Y. Chen, T. Wang, G. Chen, K. Zhu, X. Tan, J. Wang, W. Guo, Q. Wang, X. Luo,
X. Zhang, BFA-YOLO: A balanced multiscale object detection network for building
façade elements detection, Adv. Eng. Inf. 65 (2025) 103289. https://doi.org/10.
1016/j.aei.2025.103289
[157] J. Chang, Y. Lv, J. Wang, H. Pang, Y. Liu, Raster image-based house-type recognition and three-dimensional reconstruction technology, Buildings 15 (7) (2025).
https://doi.org/10.3390/buildings15071178
[158] A.J. Su, A. Ren, K.C. Xu, T. Dogan, Scalable building reconstruction and window detection for urban building energy modelling applications, J. Build. Perform. Simul.
0 (0) (2025) 1–19. https://doi.org/10.1080/19401493.2025.2501151
[159] J. Yoon, S. Hwang, K. Kim, S. Lee, Construction and recording method of a threedimensional model to automatically manage thermal abnormalities in building exteriors, Buildings 15 (9) (2025). https://doi.org/10.3390/buildings15091558
[160] J. Xie, M. Li, J. Wu, X. Zhang, J. Zhang, Semantic segmentation of building façade
materials and colors for urban conservation, npj Heritage Sci. 13 (1) (2025) 378.
https://doi.org/10.1038/s40494-025-01888-4
[161] S. Mirzabeigi, R. Razkenari, P. Crovella, Building envelope thermal anomaly detection using an integrated vision-based technique and semantic segmentation, Buildings 15 (15) (2025). https://doi.org/10.3390/buildings15152672
[162] A. Yavariabdi, M.F. Asik, G. Dogan, M.H. Arslan, Vision-based analysis of soft
story, short columns, and vertical geometry in RC structures, Structures 80 (2025)
109951. https://doi.org/10.1016/j.istruc.2025.109951
[163] Y. Hany, W. Ahmed, A. Elshazly, A.M. Senousi, W. Darwish, Automated 3D reconstruction of interior structures from unstructured point clouds, ISPRS Int. J.
Geo-Inf. 14 (11) (2025). https://doi.org/10.3390/ijgi14110428
[164] Y. Fan, J. Mai, F. Xue, S.S.Y. Lau, S. Jiang, Y. Tao, X. Zhang, W.C. Tsang, UAV
and deep learning for automated detection and visualization of façade defects in
existing residential buildings, Sensors 25 (23) (2025). https://doi.org/10.3390/
s25237118

[165] M.H. Mehraban, S. Mirzabeigi, M. Wang, R. Liu, S.M.E. Sepasgozar, Automated
image-to-BIM using neural radiance ﬁelds and vision-language semantic modeling,
Buildings 15 (24) (2025). https://doi.org/10.3390/buildings15244549
[166] P. Zhu, W.R. Para, A. Frühstück, J. Femiani, P. Wonka, Large-scale architectural
asset extraction from panoramic imagery, IEEE Trans. Visual Comput. Graphics 28
(2) (2022) 1301–1316. https://doi.org/10.1109/TVCG.2020.3010694
[167] Z. Lu, T. Wang, J. Guo, W. Meng, J. Xiao, W. Zhang, X. Zhang, Data-driven ﬂoor
plan understanding in rural residential buildings via deep recognition, Inf. Sci. 567
(2021) 58–74. https://doi.org/10.1016/j.ins.2021.03.032
[168] Y. Pan, A. Braun, A. Borrmann, I. Brilakis, 3D deep-learning-enhanced voidgrowing approach in creating geometric digital twins of buildings, Proc. Inst. Civ.
Eng. Smart Infrastruct. Constr. 176 (1) (2022) 24–40. https://doi.org/10.1680/
jsmic.21.00035
[169] B. Yu, J. Hu, X. Dong, K. Dai, D. Xiao, B. Zhang, T. Wu, Y. Hu, B. Wang, A robust automatic method to extract building facade maps from 3D point cloud data,
Remote Sens. 14 (16) (2022). https://doi.org/10.3390/rs14163848
[170] B. Haznedar, R. Bayraktar, A.E. Ozturk, Y. Arayici, Implementing PointNet for point
cloud segmentation in the heritage context, Heritage Sci. 11 (1) (2023) 2. https:
//doi.org/10.1186/s40494-022-00844-w
[171] B.G. Pantoja-Rosero, R. Achanta, K. Beyer, Automated image-based generation of
ﬁnite element models for masonry buildings, Bull. Earthquake Eng. 22 (7) (2024)
3441–3469. https://doi.org/10.1007/s10518-023-01726-7
[172] M. Dai, J. Jurczyk, H. Arbabi, R. Mao, W. Ward, M. Mayﬁeld, G. Liu, D.D. Tingley, Component-level residential building material stock characterization using
computer vision techniques, Environ. Sci. Technol. 58 (7) (2024) 3224–3234.
https://doi.org/10.1021/acs.est.3c09207
[173] M. Escalada, Leveraging deep learning segmentation techniques and connected
component analysis to automate high-level cost estimates of facade retroﬁts using 2D images, VITRUVIO Int. J. Archit. Technol. Sustainability 9 (2) (2024).
https://doi.org/10.4995/vitruvio-ijats.2024.22421
[174] C. Fotsing, W.C. Tchuitcheu, L.I. Besong, D.W. Cunningham, C. Bobda, A specialized pipeline for eﬃcient and reliable 3D semantic model reconstruction of buildings from indoor point clouds, J. Imaging 10 (10) (2024). https://doi.org/10.3390/
jimaging10100261
[175] S. Li, F. Yan, K. Ma, Q. Hu, F. Wang, W. Liu, Optimal feature-guided positionshape dual optimization for building point cloud facade detail enhancement, Remote Sens. 16 (22) (2024). https://doi.org/10.3390/rs16224324
[176] M. Mehranfar, M.A. Vega-Torres, A. Braun, A. Borrmann, Automated data-driven
method for creating digital building models from dense point clouds and images
through semantic segmentation and parametric model ﬁtting, Adv. Eng. Inf. 62
(2024) 102643. https://doi.org/10.1016/j.aei.2024.102643
[177] Z. Duan, K. Lu, M. Li, S. Zhang, B. Lin, Z. Wang, Automating exterior BIM for
existing buildings using GIS data and UAV-based 3D modeling, Autom. Constr.
178 (2025) 106422. https://doi.org/10.1016/j.autcon.2025.106422
[178] Y. Zhong, J. Jiang, W. Quan, M. Zhao, D.-m. Yan, Distinctive learning of latent space feature for occlusion-aware facade parsing, Build. Environ. 279 (2025)
112955. https://doi.org/10.1016/j.buildenv.2025.112955
[179] N. Forouzandeh, J. Huang, L. Nan, E. Brembilla, J. Stoter, Semi-automated indoor geometry reconstruction for daylight simulation, Build. Environ. 289 (2026)
114045. https://doi.org/10.1016/j.buildenv.2025.114045

19