Automatic building information model reconstruction in high-density urban areas:
Augmenting multi-source data with architectural knowledge
Ke Chen1, Weisheng Lu2, Fan Xue3, Pingbo Tang4, and Ling Hin Li5
1

Post-doctoral Fellow, Department of Real Estate and Construction, Faculty of Architecture, The
University of Hong Kong, Hong Kong, leochen@connect.hku.hk;
2
Associate Professor, Department of Real Estate and Construction, Faculty of Architecture, The
University of Hong Kong, Hong Kong, wilsonlu@hku.hk;
3
Research Assistant Professor, Department of Real Estate and Construction, Faculty of Architecture,
The University of Hong Kong, Hong Kong, xuef@hku.hk;
4
Assistant Professor, Del E. Webb School of Construction, Arizona State University, USA,
tangpingbo@asu.edu;
5
Associate Professor, Department of Real Estate and Construction, Faculty of Architecture, The
University of Hong Kong, Hong Kong, lhli@hku.hk;

This is the peer-reviewed post-print version of the paper:
Chen, K., Lu, W., Xue, F., Tang, P., & Li, L. H. (2018). Automatic building
information model reconstruction in high-density urban areas: Augmenting multisource data with architectural knowledge. Automation in Construction, 93, 22-34.
Doi: 10.1016/j.autcon.2018.05.009
The final version of this paper is available at: https://doi.org/10.1016/j.autcon.2018.05.009.
The use of this file must follow the Creative Commons Attribution Non-Commercial No
Derivatives License, as required by Elsevier’s policy.

Abstract: Many studies have been conducted to create building information models (BIMs) or
city information models (CIMs) as the digital infrastructure to support various smart city
programs. However, automatic generation of such models for high-density (HD) urban areas
remains a challenge owing to (a) complex topographic conditions and noisy data irrelevant to
the buildings, and (b) exponentially growing computational complexity when the task is
reconstructing hundreds of buildings at an urban scale. This paper develops a method — multiSource recTification of gEometric Primitives (mSTEP) — for automatic reconstruction of
BIMs in HD urban areas. By retrieving building base, height, and footprint geodata from
topographic maps, level of detail 1 (LoD1) BIMs representing buildings with flat roof
configuration were first constructed. Geometric primitives were then detected from LiDAR
point clouds and rectified using architectural knowledge about building geometries (e.g. a
rooftop object would normally be in parallel with the outer edge of the roof). Finally, the
rectified primitives were used to refine the LoD1 BIMs to LoD2, which show detailed
geometric features of roofs and rooftop objects. A total of 1,361 buildings located in a four
square kilometer area of Hong Kong Island were selected as the subjects for this study. The
1

evaluation results show that mSTEP is an efficient BIM reconstruction method that can
significantly improve the level of automation and decrease the computation time. mSTEP is
also well applicable to point clouds of various densities. The research is thus of profound
significance; other cities and districts around the world can easily adopt mSTEP to reconstruct
their own BIMs/CIMs to support their smart city programs.

Keywords: building information model; city information model; LiDAR point clouds;
topographic map; architecture; high-density city

1. Introduction
City reconstruction in 3D digital format emerges popularity in the era of information (Heo et
al. 2013). A city information model (CIM) contains spatial data and virtual representations of
all objects of interest in an urban area. A well-developed CIM can facilitate the work of city
planners and urban designers in addressing urban problems such as traffic congestion,
accessibility, connectivity, and the potential impact of natural disasters (AECbytes 2016). From
a city administrator’s perspective, a CIM with rich information can be useful for city
governance, while at the individual citizen level, a CIM enables applications such as
transportation navigation, emergency response, and many other location-based services. Cities
such as New York, London, Berlin, and Adelaide have all created their CIMs to support many
of the applications cited above (Over et al. 2010; Gröger and Plümer 2012; Sun and Salvaggio
2013).

Buildings are the most important manmade objects in the urban scene (Henricsson and
Baltsavias, 1997). Many studies, over the years, have focused on the reconstruction of building
information models (BIMs) (e.g. Tang et al. 2010; Xue et al. 2018) which can be stitched
together to form a CIM. Another approach is to create CIMs using Geographic Information
Systems (GIS) (Li et al. 2015) and remote sensing (Haala and Kada 2010; Lillesand et al. 2015).
In these CIMs, individual buildings could be roughly represented by prisms or “boxes” without
precise information on the “as-is” condition. With the advancements of data acquisition and
processing technologies, the trend is to reconstruct BIMs that contain detailed geometric
features of roofs and rooftop objects (so termed as Level of Detail 2 [LoD2] defined by the
Open Geospatial Consortium [OGC] [2012]) to extend CIM applications, e.g. green roof
development (Choi et al. 2017) and energy performance improvement (Yang and Zou 2016).

2

However, the reconstruction of BIMs, particularly those with greater details, is labor-intensive,
time-consuming, and error-prone (Volk et al. 2014). The process requires a considerable
amount of manual rectifications and computational power, and this becomes extreme
burdensome when the task is at the urban scale (Sun and Salvaggio 2013; Li et al. 2016; Wu
et al. 2017).

Researchers have attempted to improve the efficiency of BIM reconstruction by introducing
automatic or semi-automatic approaches. Images, 3D laser scanning point clouds, and total
station surveying data are commonly used for model reconstruction (e.g. Awrangjeb et al. 2013;
Li et al. 2016; Wu et al. 2017). Algorithms have been developed to process different types of
data and reconstruct BIMs (e.g. Heo et al. 2013; Xue et al. 2018). In addition, with data from
multiple sources become affordable, it is now possible to use multi-source data to overcome
some of the inherent problems (e.g. inaccurate/”noisy” data, incomplete information)
associated with single-source data (e.g. Habib et al. 2010; Cheng et al. 2011; Gilani et al.
2016). Acknowledging considerable achievements in the field of BIM/CIM reconstruction,
BIM/CIM reconstruction in high-density (HD) urban areas remains an open problem
(Musialski et al. 2013). Firstly, city features such as trees, roads, and terrain introduce a lot of
noise that undermines the quality of the measurement data. Secondly, densely-distributed
buildings make it difficult to segment data for generating individual BIMs. Lastly,
reconstructing thousands of buildings at an urban scale exponentially increases the
computational complexity, bringing many difficulties for methods which rely on general object
recognition approaches to derive geometric primitives to form the building models.

This study aims to improve automatic BIM reconstruction in HD urban areas by proposing a
reconstruction method called multi-Source recTification of gEometric Primitives (mSTEP).
mSTEP harnesses the data from multiple sources and makes use of architectural features (e.g.
parallels and symmetries) to reduce the noisy data and fine-tune the geometric primitives to
reconstruct LoD2 BIMs automatically. The data employed in this study comes from the Hong
Kong governmental agencies, comprising digital topographic map and light detection and
ranging (LiDAR) point cloud. Given the fact that such types of data are extensively available
in many cities and districts around the world, mSTEP can be applied to reconstruct their
BIMs/CIMs in an efficient manner.

3

The organization of the paper is as follows: Section 2 reviews the state-of-the-art studies on
BIM reconstruction. Section 3 describes the overall research progress, the subject area and
characteristics of the corresponding topographic map and LiDAR point clouds. Section 4
details the BIM reconstruction method – mSTEP. Section 5 provides a comprehensive
discussion of the evaluation results, discusses the parameter configuration of mSTEP, and
shows the compatibility of mSTEP with a denser point cloud. Section 6 concludes with a
summary and highlights future research directions.

2. Literature Review
Previous studies on the generation of BIMs in the urban environment can be viewed from two
different perspectives: (1) The raw data used for BIM reconstruction; and (2) the methods
employed in processing the data. While the two perspectives are related, the discussion deals
with each perspective separately for the sake of clarity.

2.1 Original datasets
Aerial and satellite images are typical data sources for large-scale BIM reconstruction.
Spaceborne sensors like IKONOS, QuickBird, and GeoEye-1 have provided 1m-resolution
satellite images for 3D building reconstruction (Lafarge et al. 2008; Poli et al. 2015). Aerial
image resolution can be even higher than that of satellite images, in some cases, reaching
decimeter accuracy. Owing to their high image resolution as well as the widespread use of
unmanned aerial vehicles (UAVs), the uses of aerial images to create 3D models of individual
buildings or even to reconstruct the entire urban scenes are increasing (Li et al. 2016).

LiDAR point clouds have also been widely used for BIM reconstruction. It typically utilizes
laser light which is projected on surfaces and its reflected backscattering is captured for
generating 3D point clouds. Heo et al. (2013) used LiDAR point clouds to develop the models
of 29 buildings. Other studies including Sun and Salvaggio (2013), Xiong et al. (2014), and
Yan et al. (2016) used airborne LiDAR point clouds to model a limited number of buildings
with roofs of various shapes. With the availability of city-scale LiDAR point clouds, Poullis
and You (2009) created simplified BIMs within a large city area.

Topographic maps, which describe urban objects in terms of geometry, land use, and other
attributes, are another important data source for BIM reconstruction at city- or district-scale. In
addition to the topographic maps produced by government agencies, recent years have seen the
4

emergence of open-access geographic datasets. For example, OpenStreetMap (OSM) – a
prominent volunteered geographic information service – has been used for BIM reconstruction
(Over et al. 2010). However, without official verification, the common problem of open-access
geographic datasets is their completeness and accuracy.

The use of single-source data for BIM reconstruction, be it aerial images or LiDAR point
clouds, is prevailing but still poses problems such as “noise” data caused by complex urban
features and incomplete information (Cheng et al. 2011). These drawbacks have given rise to
increased use of multi-source data for BIM reconstruction. A number of studies have confirmed
that using multi-source data can overcome some of the problems associated with the use of
single-source data (Rottensteiner and Jansa 2002; Alexandar et al. 2009; You and Lin 2011;
Henn et al. 2013; Zhang et al. 2014; Zhu et al. 2015).

2.2 Data processing methods
Various data processing methods have been proposed for BIM/CIM reconstruction (Haala and
Kada 2010; Musialski et al. 2013). Aerial or satellite images can be processed into digital
elevation model (DEM) (Lafarge et al. 2008; Poli et al. 2015) from which building models can
be extracted by applying height thresholds. With the further development of image matching,
an alternative way is to generate colored point clouds from a number of aligned aerial images
and then processed the generated point clouds into textured BIMs, which are formed by a large
number of small geometric primitives (Singh et al. 2014). This method, however, requires
careful selection of images and manual interpretation to adjust the building models is often
needed. Li et al. (2016) also generated point clouds from images, and proposed an object-level
point cloud segmentation and roof extraction. However, their method was only tested on
buildings with flat roofs.

Processing LiDAR data starts with segmentation of the point clouds of individual buildings.
This can generally be achieved using semantic segmentation approach (Lin et al. 2013),
classification or clustering algorithms (Zhu et al. 2015; Cao et al. 2017) or the reflectance value
captured by the LiDAR sensors (Sun and Salvaggio 2013). LiDAR point clouds can also be
integrated with other datasets. For example, building footprints retrieved from a topographic
map can provide a reference for segmenting the point clouds of buildings (Alexander et al.
2009; Ledoux and Meijers 2011). The segmentation of LiDAR point clouds can also be
improved with aerial images that provide regions of homogeneous gray level or color
5

distribution (Rottensteiner and Jansa 2002). Once segmented, the point clouds are used to
model the buildings with roofs and rooftop objects by various methods. A typical method is to
decompose the roof shapes into simple pre-defined ones by 2D plans (Henn et al. 2013) or
graph matching technique (Xiong et al. 2014), but the reconstruction may fail if the roof shape
is not pre-defined in the model library. Connecting the extracted primitives to form the roof
features is also widely-used due to its flexibility (Poullis and You 2009; Zhang et al. 2014; Yan
et al. 2016). However, such kind of method is sensitive to noise (Goebbelsa and PohleFröhlicha 2016) and so far have only been used for specific roof forms.

Although it is difficult to directly compare all these reconstruction methods since they are
developed under different context with their own emphasis, our review has revealed that
existing data processing methods usually require much time for noise filtering and assume
buildings with flat or other simplified roof structures. Those methods that can generate more
differentiated building and roof structures require considerable manual interpretation for preor post-processing. Actually, architectural designs commonly exhibit some conventional
features such as parallels, symmetries or other structural regularities, which are not accidental,
but often the result of economical, manufacturing, functional, or aesthetic considerations (Mitra
and Pauly 2008). Parallel and perpendicular features have been used as the constraints to
segment the point cloud to extract planar segments that constitute approximate building roof
structure (Dorninger and Pfeifer 2008, Sampath and Shan 2010). However, few studies have
systematically applied architectural rules to reconstructing BIMs of densely-distributed
buildings in large-scale urban areas. As will be demonstrated in this study, rules derived from
architectural conventions can help reduce the noise in the collected data and improve the
efficiency of the modeling method and the informativeness of the reconstructed models.

3. Research approach
3.1 Overview of the research process
Given the aim of this study, a design science research approach (Peffers et al. 2008) is adopted
to develop an automatic and efficient BIM reconstruction method. Design science focuses on
not only understanding problems, but also developing methods or artifacts with the explicit
intention of improving human performance (Van Aken 2005). A diagrammatic illustration of
the research process is presented in Figure 1.

6

Figure 1 Research process

By identifying the challenges of BIM reconstruction in high-density urban areas, multi-Source
recTification of gEometric Primitives (mSTEP) was proposed, which takes the topographic
map, point clouds, and architectural conventions as inputs and generates BIMs with detailed
rooftop structures and objects identified. A 2km×2km area in northwestern Hong Kong Island,
containing 1,361 blocks of densely-distributed buildings of varying heights and shapes, was
selected as the subject area since all identified challenges of BIM reconstruction exist in this
area: The buildings are surrounded by urban features like trees on hills and slopes, commercial
signs and power lines, which create noise in the data captured for BIM reconstruction; The
narrow gaps between buildings also cause severe occlusions in the collected data. In
combination, these factors present difficulties for segmentation, refinement and other BIM
reconstruction processes. After applying mSTEP on test datasets, the generated BIMs were
compared with results from another two reconstruction methods (i.e. methods described in
Javanmardi et al. [2015] and Wu et al. [2017]) and manual modeling for evaluation.

3.2 Test datasets
The topographic map used in this study was purchased from the Lands Department (LD) of
HKSAR. It is in Geodatabase (GDB) format with a scale of 1:1000. The Hong Kong 1980 Grid
Coordinates provides the latitude and longitude of the map, and heights are in meters above
the Hong Kong Principal Datum. The topographic map contains feature datasets including
buildings, land cover, transportation, etc. As shown in Figure 2, the buildings in the selected
region are shown in green with their boundaries in black. The map also contains data on
building ID, shape, shape area, type of building block, base level, roof level, and data source.
A preliminary analysis of the base and roof levels found that the datasets in the topographic
map are collected from various sources such as building plans, photogrammetry, and
7

topographic survey. Data for the buildings was last updated in the period 1 July 2014 to 27
April 2016.

Figure 2 Data of the topographic map of the subject area

The LiDAR point clouds used in this study was provided by the Civil Engineering and
Development Department (CEDD) of HKSAR. The data, comprising buildings, roads, and
many other urban features, was collected between 1 December 2010 and 8 January 2011 by
the CEDD by hiring an airborne LiDAR surveying company. In the original dataset, the point
density is about 4 points/m2. Points classified according to American Society for
Photogrammetry and Remote Sensing (ASPRS) laser (LAS) format specification 1.1 (ASPRS
2005) are shown in different colors (see Figure 3). The classification codes represent the type
of object that has reflected the laser pulse.

8

Figure 3 An illustration of the LiDAR point clouds

4. BIM reconstruction method
The main purpose of the proposed method — multi-Source recTification of gEometric
Primitives (mSTEP) — is to enable automatic and efficient BIM reconstruction. The overall
procedure of mSTEP is shown in Figure 4, which consists of four main phases. In this section,
each of the four phases of mSTEP is introduced in details.

Figure 4 Overall procedure of mSTEP

4.1 Reconstructing LoD1 BIMs from topographic map
In the first phase, data in the topographic map is filtered to retrieve the data of building footprint,
base level, and top level. By vertically stretching the footprint of each building according to
the base level and top level of that building, the LoD1 BIM with flat roof configuration can be
created (see Figure 5).

9

Figure 5 An illustration of created LoD1 BIMs

4.2 Filtering the LiDAR point clouds
The LiDAR point clouds is first filtered in ESRI ArcScene by using the classification codes to
separate building point clouds from those of trees, terrain, and other features (ESRI 2016).
However, it was found that many points of building structures were unclassified (see the circled
area in Figure 6). Therefore, points classified as either “building” or “unclassified” were thus
kept for further rectifications and other points were removed. Then, the point cloud of
individual buildings was obtained by segmentation, a process accelerated through the use of
building footprints contained in the topographic map. After converting the LiDAR point clouds
and topographic map in the same coordinate system, points with XY-coordinates falling into
the area zoned by a building footprint was segmented to that building.

Figure 6 An example of unclassified points in the raw LiDAR point clouds

4.3 Developing object rectification rules from conventional architectural features
Architectural conventions refer to the domain knowledge applied in the architectural design,
such as geometrical or physical features (e.g. flatness, parallels, and symmetries) of building
objects and their spatial relationships (Cantzler 2003). Some of the conventions can be obtained
10

from ordinances or standards, while others can be originated from general observation in realworld situations. In this study, two object rectification rules are developed based on very simple
architectural conventions to ensure “structural regularity”: (1) the top surface of rooftop objects
are either in parallel with the horizontal plane or with a considerable angle of dip; and (2) the
rooftop objects are normally in parallel with the outer edge of roof.

Both rules should be satisfied when rectifying the geometric primitives detected from the
segmented LiDAR point clouds to form roofs and rooftop objects. An illustration of
rectification by the first rule is presented in Figure 7(a). The rectangle in blue is the top surface
of a rooftop object, which has a very small angle β of dip to the horizontal plane (i.e. the roof).
Therefore, this detected primitive is against the first rule, and should be rotated to be the
rectangle in gray, which is in parallel with the horizontal plane. An illustration of rectification
by the second architectural rule is shown in Figure 7(b). The rectangle in dashed line is not in
parallel with the outer edge of the roof, but has a deviation of angle α. This is rare in
architectural conventions, therefore, it should be rotated to be the rectangle in gray.

(a) First object rectification rule

(b) Second object rectification rule

Figure 7 An illustration of the two rules of rectification

In order to enable automatic application of these two rules, especially the second one, guiding
directions were are derived from the building footprints since the outer edge of the roof and
building footprint usually have the same shape, unless it is an irregular-shaped building. As
can be seen in Figures 8(a) and (b), for simple, regular-shaped building footprint, its centerline
provide the guiding direction for the rooftop object to be in parallel with, while for irregularshaped building footprint, the centerlines of its top 25% longest lines will provide guiding
directions for the computer to check whether the rooftop object is in parallel with at least one
of the directions.

11

(a) Regular-shaped footprint

(b) Irregular-shaped footprint

Figure 8 An illustration of deriving guiding direction(s) from building footprints

4.4 Reconstructing LoD2 BIM by using object rectification rules
In the last phase of mSTEP, LoD2 BIM is developed by using the two object rectification rules.
The pseudo code of this process is shown in Figure 9. Given a target building, its LoD1 model
is inherited from the topographic map (Line 6). A variant of RANSAC (Fischler and Bolles
1981; Schnabel et al. 2007) is adopted for detecting geometric primitives from LiDAR point
cloud of the building rooftop (Lines 7-8). The application of RANSAC requires a set of
parameters to be determined (see Table 1). Different values of these parameters have been
tested in order to find the optimal parameter set regarding the LiDAR point clouds to be
processed. It is decided to set minimum number of support points to 10, maximum distance to
primitive to 0.02m, and sampling resolution to 1.0. Maximum normal deviation and
overlooking probability, are set to 25o and 1.0×10-6 respectively. More details regarding the
parameters configuration are provided in Section 5.2.

Then, the detected geometric primitives are rectified to form the building roofs and rooftop
objects. A tolerance level (ε), which is set to 0.05π, is used to determine whether one primitive
needs rectification or not. The detected geometric primitives of roofs and rooftop objects will
be rectified by the horizontal plane (Lines 9-13) and the guiding directions (Lines 14-21). After
rectification, the volumetric models of roofs and rooftop objects can be created by projection
to the top level of individual buildings, and will be integrated with the LoD1 BIMs created in
the first phase to enrich them into LoD2 ones that have differentiated roof structures with higher
completeness and accuracy (Lines 22-24).
1

procedure multi-source primitives rectification

12

2

input: a set of building exterior data sources (S) of a building, including data in topographic map and
LiDAR point clouds,

3

five parameters (param) of RANSAC,

4

a tolerance level (ε) for orientation rectification (default value of 0.05π)

5

output: an LoD2 building model (M) with detailed roof and rooftop objects

6

M  a LoD1 model from topographic map (S)

//referring to subsection 4.1

7

r  filter point cloud of rooftops (S);

//referring to subsection 4.2

8

P  RANSAC (r, param);

//detecting geometric primitives

9

loop for each p in P

//for each detected geometric primitive

10

α  heading of normal (p); β  dip of plane (p);

// for α and β referring to Figure 7

11

if |β| < ε or | π - β| < ε

//if β is close to the horizontal plane

β0

12

//rectify w.r.t. the horizontal plane (as shown in Figure 7[a])

13

end if

14

A  derived guiding directions;

15

loop for each α* in A

16

loop for q = -2 to 2

//for each direction α*

17

if | α - α + q × π / 2| < ε

18

α  α* - q × π / 2

//for each quadrant
//if α is close to α*

*

19

end if

20

end loop

//rectify w.r.t. the direction α* (as shown in Figure 7[b])

21

end loop

22

if α, β meets the two object rectification rules

23

M  M∪update and 3D projection(p, α, β)

24

end if

25

end loop

26

return M

//enrich LoD1 BIM with roof and rooftop objects

Figure 9 Pseudo code of the multi-source rectification

Table 1 Descriptions of the RANSAC parameters
Parameter

Description

minimum number of support

The minimum number of points required to identify a geometry. A larger

points

minimum support points means a more rigorous geometry detection process.

maximum distance to primitive

The maximum distance of an inlier point to a geometry.

sampling resolution

The parameter determines the sample rate of neighboring points

maximum normal deviation

The difference between the normal at a point and the normal of a geometry
at the closest project of that point onto that geometry.

overlooking probability

The parameter controls the population size of the primitive candidates.

13

A demonstration of multi-source rectification is shown in Figure 10. At the beginning,
geometric primitives are detected from LiDAR point clouds (see Figure 10[a]) by RANSAC,
but the detected geometric primitives may not be corrected from an architectural point of view.
For instance, in Figure 10(b), the rectangle marked in red circle conjuncts to the roof surface
at an angle. Then, the detected primitives are rectified according to the horizontal plane and
the guiding direction(s). After that, the rectified primitives will be projected to create
volumetric models of roofs and rooftop objects (see Figure 10[c]).

(a) LiDAR point clouds of one

(b) Geometric primitives detected

building roof of the subject area

by using RANSAC

(c) Rectified geometric primitives

Figure 10 A demonstration of the multi-source rectification

5. Evaluation
5.1 Overall results
In this study, the first phase of mSTEP is performed in ESRI ArcScene, and the remaining three
phases are automatically performed by using a plug-in of CloudComapre which is tailorprogrammed in C++ by the authors. A screenshot of the plug-in is shown in Figure 11. The
computation environment was a personal notebook with 2.6GHz Quad-core CPU, 16 GB RAM,
and a 64-bit Microsoft Windows 10 operating system. Applying mSTEP, the automatic
generation of BIMs in the subject area took 319.7 seconds. 81.9% of the 1,361 buildings were
successfully modeled. Among the remainder, 113 buildings do not have either the base level
or the roof level in the topographic map, and 134 buildings lacked points in the LiDAR dataset
for modeling of the roofs and the rooftop objects.

14

Figure 11 The screenshot of the developed CloudCompare plug-in

Some examples of the reconstructed models are shown in Figure 12. Among the buildings
shown include a building on a university campus (ID: 11**374), Edwardian Baroque in style
with a central clock tower and several turrets; a residential building with a pyramid-shaped
roof and many rooftop objects (ID: 11**486); a wholesale food market with a strip-shaped
roof and several box-shaped air conditioners (ID: 11**535); and a residential building (ID:
11**845) with flat roof and cuboid-shaped rooftop objects. Other building models shown in
Figure 12 also have their roof and rooftop objects in the correct shapes constituted without the
need for manual post-modifications. By manually checking the 3D building models and the
actual buildings, it can be seen that the models reach the LoD2 that possess more detailed
information than simple “boxes”. It can also be seen that the models are in an acceptable
standard of accuracy.
Building ID in the
topographic map

Reconstructed model (colors
are for distinguishing only)

11**374

11**486

15

Photo (collected from the
Internet or taken by the authors)

11**535

11**845

11**386

11**200

11**500

11**345

Figure 12 An illustration of the reconstructed models and actual buildings

5.1.1 Efficiency assessment
To assess the efficiency of reconstruction, the time spent on reconstructing buildings by using
mSTEP was compared with the time required for manual modeling. In SketchUp, two human
modelers, who have over four years’ experience in architectural drawing, created LoD2 BIMs
based on base and roof levels from the topographic map and the segmented LiDAR point clouds.
They began modeling at the same time and alerted the researchers once they finished each of
the nine models so that the time they spent on modeling each building was recorded. In addition,
mSTEP was compared with: (1) a concave hull and Hough transform based reconstruction
method (take LiDAR point cloud and topographic map as inputs) introduced in Javanmardi et
16

al. (2015); and (2) a bipartite graphic matching-based reconstruction method introduced in Wu
et al. (2017). The two reconstruction methods were selected for comparison due to two reasons.
First, all the three methods share the same vision of enhancing 3D model reconstruction by
taking advantage of architectural conventions and graph theory. The other reason was that their
input point clouds were all airborne LiDAR data of HD urban areas. The comparison was based
on three commonly-adopted metrics, including (1) reconstruction time; (2) percentage of
points that are segmented to support the detection of geometric primitives; and (3) root mean
square error (RMSE) of the distances of points to their corresponding primitives. The
assessment results on nine randomly-selected buildings are presented in Table 2.
Table 2 Assessment results
Javanmardi et al. (2015)†

Wu et al. (2017)‡

Manual

Time
(s)

Segmented
#
(%)

RMSE
(m)*

Time (In
average; s)

0.119

0.01

88.3

0.341

266.65

58.7

0.157

0.01

60.4

0.223

171.23

0.02

84.7

0.098

0.01

78.8

0.370

94.33

0.196

0.05

31.6

0.199

0.01

99.8

0.315

2,486.75

90.0

0.057

0.03

60.2

0.049

0.01

89.6

0.251

515.65

0.23

91.1

0.067

0.04

26.4

0.067

0.01

90.4

0.280

597.92

8,987

0.26

85.1

0.065

0.09

18.8

0.097

0.01

100.0

0.257

1,386.16

24,878

2.23

86.3

0.068

0.13

66.2

0.163

0.02

99.1

0.340

2,498.84

mSTEP

Number
of
points

Time
(s)

Segmented
#
(%)

RMSE
(m)*

Time
(s)

Segmented
#
(%)

RMSE
(m)*

1

270

0.03

75.2

0.114

0.02

63.1

2

372

0.03

77.7

0.048

0.01

3

620

0.40

87.6

0.065

4

2,491

0.16

48.6

5

2,682

0.09

6

7,212

7
8

No.

9
29,506 0.47
90.1 0.088 0.21
71.1 0.230 0.02
99.9 0.291
701.43
†: The paprameter β of Javanmardi et al. (2015)’s method was set to 0.5m as the average point distance;

‡: The two parameters, i.e., variable n and contour interval di, were set as 250 and 0.5m based on Wu et al.
(2017)’s experimental results.
#

: Percentage of points that are segmented to support the detection of geometric primitives.

*

: Root mean square error of the distances of segmented points to their corresponding primitives.

The results in Table 2 showed that mSTEP achieved a competitive performance. Specifically,
both mSTEP and Javanmardi et al. (2015)’s method reconstructed accurate geometric
primitives for segmenting the point clouds. The proposed mSTEP segmented more points than
Javanmardi et al. (2015)’s to support the detection of geometric primitives and reconstruct
BIMs with less RMSE. In contrast, Wu et al. (2017)’s method achieved a fast reconstruction
with a high-level of segmentation; yet at a significant cost of error (see the level of RMSE).
The results confirmed that mSTEP could overcome the three identified challenges in BIM
reconstruction in HD urban areas, i.e., much noise in measurement data, difficulties in
segmenting data of densely-distributed buildings, and high computational complexity. With
17

the help of mSTEP, the efficiency of BIM reconstruction in HD urban areas can be
considerably improved.

Nevertheless, there seems no convincing way to measure the exact accuracy of the
reconstructed BIMs scientifically due to the lack of ground truth (Poullis and You 2009; Sun
and Salvaggio 2013). It is far from an effective measurement to reflect the resemblance
between the generated BIMs and the “as-is” condition. Neither can the model accuracy be
measured using a single index, such as the physical volume. A BIM professional, with expertise
and insights, can also tell whether the accuracy of a reconstructed 3D model is “acceptable” or
not. The issue of measurement standard is left for further studies.

5.1.2 Identified problems
During application of mSTEP, it was found that the modeling outcomes can be affected by the
accuracy of data used, including the building footprints and heights in the topographic map,
and the LiDAR point clouds. Regarding the topographic map used in this study, some building
footprints were inaccurate and some heights were generated by estimation. This affected the
quality and accuracy of the retrieved guiding directions and, in turn, the accuracy of geometric
primitives detected from the point clouds.

Regarding the accuracy of the LiDAR point clouds, this is adversely affected by missing points
and interference caused by densely-distributed buildings, city features, and the like. mSTEP
can address interference, but not the problem of missing data. In this study, three factors
resulted in missing data. Firstly, the LiDAR scan pulse typically does not provide a detectable
return from transparent materials such as glass. Therefore, the points for roofs and rooftop
objects made of such materials were missing and thus could not be used for BIM reconstruction.
Secondly, the density of the point cloud was 4 points/m2. This may not be sufficient to capture
data on certain rooftop objects, such as parapet walls that are relatively thin in the horizontal
dimension. Thirdly, the LiDAR point clouds, collected from 2010 and 2011, was not up-todate.

5.2 Parameter configuration of mSTEP
A set of tests was performed on the parameter configuration of mSTEP. The purpose is to
identify the parameters that can significantly impact the modeling outcomes measured by the
average number of detected geometric primitives. This indicator is chosen because it measures
18

the primitives generated from the LiDAR point clouds to form the roof and rooftop objects.
Generally, the larger value the indicator, the more details the roof and rooftop objects have
been reconstructed by mSTEP. However, excessive primitives detected are also undesirable
since they are often caused by a single rooftop object being separated into several small pieces
(e.g. a pipe separated into different trunks), which actually decrease the accuracy of
reconstruction.

The parameters to be tested include the tolerance level and the five RANSAC parameters in
terms of (a) minimum number of support points, (b) maximum normal deviation, (c) maximum
distance to primitive, (d) sampling resolution, and (e)overlook probability. The test subject is
the LiDAR point clouds of a randomly-selected building, which has an ID 11**200 in Figure
12. The tests aim at evaluating the sensitivity of each parameter, therefore when performing
the testing one parameter, other five parameters are controlled, i.e. remaining unchanged. In
the single factor sensitivity analysis, for each parameter, the tests were replicated for 100 times
to get the statistic values at 5th and the 95th percentiles. The results of all tests are shown in
Figure 13. In each of the six sub-figures, the Y-axis on the left denote the number of geometric
primitives detected by mSTEP, the X-axis denotes the value of the tested parameter. The box
chart shows the average values, and the 5th and the 95th percentiles of the indicator. The curve
depicted in each sub-figure illustrates the trend of value changes of the indicator as reflection
to the value change of the parameter under evaluation.

19

Figure 13 Testing results of parameter configuration of mSTEP

Based on the results shown in Figure 13, it can be confirmed that minimum number of support
points and maximum normal deviation are the two parameters that have the most significant
impact on the indicator of average number of detected geometric primitives since the curves
shown in Figure 13(a) and (b) cover a much wider range in Y-axis than the curves in the other
four figures. Particularly in Figure 13(a), when minimum number of support points increase
from 3 to 10, the indicator decreases steeply from 95 to 33.6. Once minimum number of support
points exceed 10, the indicator will continuously decrease, at a slower pace, to 9. From the
three outputs of mSTEP when minimum number of support points equal to 3, 10 and 50
respectively, it can be seen that some of the rooftop objects such as the parapet wall are divided
into small pieces (see the comparison between the two parts marked in red rectangle of Figure

20

13[a]). When minimum number of support points is set to 3, too many primitives are detected.
By contrast, when minimum number of support points is set to 50, many details of rooftop
objects are missing. Setting minimum number of support points to 10 can generate an
appropriate output. In Figure 13(b), when maximum normal deviation increases from 0o to 90o,
the average number of detected geometric primitives increases from 0 to 55.4. Similar to the
evaluation on minimum support points, three outputs when maximum normal deviation equals
to 5o, 25o, and 90o are presented, from which, it is found that setting maximum normal deviation
to 5o and 90o lead to too few and too many detected primitives respectively. When maximum
normal deviation equals 25o, an appropriate amount of primitives can be detected.

For the remaining four parameters, the impact of maximum distance to primitive on the
indicator is relatively complex. As it increases from 1.0×10-4m to 1.0m, the average number
of detected geometric primitives fluctuates (see Figure 13[c]). It seems that setting maximum
distance to primitive to either 1.0×10-4m or 0.02m can deliver similar outputs in terms of
quantity, but it is found that when maximum distance to primitive equals to 1.0×10-4m, most of
the detected primitives are too small, which shows a lack of accuracy. Therefore, 0.02m is the
more appropriate value for maximum distance to primitive. Additionally, when tolerance level
increases from 0.001π to 0.5π, the average number of detected geometric primitives increases
from 3.58 to 33.91. Such impact, however, becomes negligible after tolerance level reaches
0.05π (see Figure 13[f]). For sampling resolution, when it is either less than 1.0 or large than
1.25, the average number of detected geometric primitives will become less than 30, which
decreases the accuracy of the reconstructed roof and rooftop objects (see Figure 13[d]). Finally,
for overlook probability, its overall influence is less significant compared with the other five
parameters (see Figure 13[e]). Its values between 1.0×10-6 and 5×10-5 can detect an appropriate
amount of primitives, but the best output is identified when setting overlook probability to
1.0×10-6.

5.3 Implementing mSTEP on dense point cloud
The density of LiDAR point clouds used to reconstruct buildings in the subject areas is 4
points/m2. With the improvement in sensing devices, the cost of data acquisition is expected to
continuously decrease and denser point clouds will become available. Therefore, it is necessary
to test the applicability of mSTEP to dense point clouds. Since few buildings in the subject area
have publicly available dense point cloud, a building at the main campus of the University of
Hong Kong in Hong Kong Island is selected for the test.
21

The dense point cloud of the selected building contains 482,404 points in total, which is
obtained by using SfM to process 200 photos taken by a UAV (see Figure 14[a]). Details about
SfM for processing images into point cloud can be found in Jancosek and Pajdla (2011). When
applying mSTEP to this dense point cloud, one parameter, i.e. minimum number of support
points, is changed from 10 to 30 in order to cope with the point density, while other five
parameters remain unchanged. The rooftop model is developed in 1.41s with 475 geometric
primitives detected (see Figure 14[b]). In doing so, mSTEP is proved to be capable of
reconstructing BIMs from a denser point cloud. A conclusion thus can be drawn that mSTEP
is applicable to both sparse point clouds (e.g. 4 points/m2) and dense point clouds for BIM
reconstruction. This suggests mSTEP can be applied to 3D building reconstruction in many
other cities or districts, which possess point clouds of various densities.

(a) Dense point cloud (482,404 points)

(b) Rooftop model generated from (a)

Figure 14 An illustration of applying mSTEP to a dense point cloud

6. Conclusions
The challenges of automatic reconstruction of building information models (BIMs)/city
information models (CIMs) in high-density (HD) urban areas are predominately twofold: (a)
the complex topographic conditions and noisy data, and (b) the heavy computational
complexity when the task is at an urban scale. This paper took the challenges by developing an
improved method — multi-Source recTification of gEometric Primitives (mSTEP) — for
automatic BIM reconstruction in complex urban areas. mSTEP comprises of several
interconnected steps to make good use of multi-source data including light detection and
ranging (LiDAR) point clouds, and topographic maps. The method has been validated through
a series of rigorous tests, and the results show that mSTEP is an efficient method to reconstruct
informative BIMs by significantly improving the level of automation and decreasing the
computation time.

22

Of particular originality of this research is augmenting the multi-source data with some simple
architectural conventions, which can effectively tackle the challenges of automatic BIM/CIM
reconstruction. Another original contribution made by this study is the optimal parameter
configuration of mSTEP, which is derived from a series of sensitivity analyses of the
parameters’ impacts on the accuracy of the reconstructed models. As the paper shows, mSTEP
can be applied to both sparse and dense point clouds. The research is thus of profound
significance; It can help other cities or districts, which have possessed such “common” datasets
as topographic map and LiDAR point clouds, to produce their own BIMs/CIMs to support their
smart city programs.

Future research will be conducted in mainly three aspects. Firstly, the two object rectification
rules used in the current version of mSTEP might be too restrictive when dealing with atypical
architectures. Improvements thus are desired to allow reconstruction of buildings with curved
and irregular-shaped roofs and rooftop objects. Secondly, data from other sources, such as
aerial images, will be integrated with the topographic map and LiDAR point clouds to increase
the level of detail (e.g. texture) of the generated models. In this connection, other architectural
conventions such as symmetries or repetitive patterns should be exploited to facilitate the BIM
reconstruction processes.

Acknowledgement
This study is financially supported by HKU Seed Fund (201702159013) and RGC General
Research Fund (17201717). The assistance of the Civil Engineering and Development
Department (CEDD) of Hong Kong SAR in providing LiDAR point clouds for research
purposes is much appreciated. The authors are particularly in debt to Prof. Kincho Law at
Stanford University, who has spent his precious time to provide several rounds of comments
on the paper as it was developed. The authors would like to thank all the editors and anonymous
reviewers for their constructive comments.

References
AECbytes

(2016).

City

Information

Modeling.

Available

at:

http://www.aecbytes.com/feature/2016/CityInformationModeling.html, Accessed date:
6 January 2017.
Alexander, C., Smith-Voysey, S., Jarvis, C., and Tansey, K. (2009). Integrating building
footprints and LiDAR elevation data to classify roof structures and visualise buildings.
23

Computers,

Environment

and

Urban

Systems,

33(4),

285-292.

doi:10.1016/j.compenvurbsys.2009.01.009.
ASPRS (American Society for Photogrammetry & Remote Sensing). (2005). LAS
specification

version

1.1.

Available

at:

http://www.asprs.org/wp-

content/uploads/2010/12/asprs_las_format_v11.pdf, Accessed date: 6 January 2017.
Awrangjeb, M., Zhang, C., and Fraser, C. S. (2013). Automatic extraction of building roofs
using LIDAR data and multispectral imagery. ISPRS Journal of Photogrammetry and
Remote Sensing, 83, 1-18. doi:10.1016/j.isprsjprs.2013.05.006.
Cantzler, H. (2003). Improving architectural 3D reconstruction by constrained modelling.
Doctoral

thesis,

University

of

Edinburgh.

Available

at:

https://www.inf.ed.ac.uk/publications/thesis/online/IP030017.pdf, Accessed date: 6
January 2017.
Cao, R., Zhang, Y., Liu, X., and Zhao, Z. (2017). 3D building roof reconstruction from airborne
LiDAR point clouds: a framework based on a spatial database. International Journal of
Geographical

Information

Science, 31(7),

1359-1380.

doi:10.1080/13658816.2017.1301456.
Cheng, L., Gong, J., Li, M., and Liu, Y. (2011). 3D building model reconstruction from multiview aerial imagery and lidar data. Photogrammetric Engineering & Remote Sensing,
77(2), 125-139. doi:10.14358/PERS.77.2.125.
Choi, J. O., Chen, X. B., and Kim, T. W. (2017). Opportunities and challenges of modular
methods in dense urban environment. International Journal of Construction
Management, 1-13. doi:10.1080/15623599.2017.1382093.
Dorninger, P., and Pfeifer, N. (2008). A comprehensive automated 3D approach for building
extraction, reconstruction, and regularization from airborne laser scanning point
clouds. Sensors, 8(11), 7323-7343. doi:10.3390/s8117323.
ESRI

(2016).

Lidar

point

classification,

Available

at:

http://desktop.arcgis.com/en/arcmap/10.4/manage-data/las-dataset/lidar-pointclassification.htm, Accessed date: 15 January 2017.
Fischler, M. A., and Bolles, R. C. (1981). Random sample consensus: a paradigm for model
fitting

with

applications

to

image

analysis

and

automated

cartography.

Communications of the ACM, 24(6), 381-395. doi:10.1145/358669.358692.
Gilani, S. A. N., Awrangjeb, M., and Lu, G. (2016). An automatic building extraction and
regularisation technique using LiDAR point cloud data and Orthoimage. Remote
Sensing, 8(3), 258. doi:10.3390/rs8030258.
24

Goebbelsa, S., and Pohle-Fröhlicha, R. (2016). Roof reconstruction from airborne laser
scanning data based on image processing methods. ISPRS Annals of Photogrammetry,
Remote Sensing and Spatial Information Sciences, 407-414. doi:10.5194/isprs-annalsIII-3-407-2016.
Gröger, G., and Plümer, L. (2012). CityGML–Interoperable semantic 3D city models. ISPRS
Journal

of

Photogrammetry

and

Remote

Sensing,

71,

12-33.

doi:10.1016/j.isprsjprs.2012.04.004.
Haala, N., and Kada, M. (2010). An update on automatic 3D building reconstruction. ISPRS
Journal

of

Photogrammetry

and

Remote

Sensing,

65(6),

570-580.

doi:10.1016/j.isprsjprs.2010.09.006.
Habib, A. F., Zhai, R., and Kim, C. (2010). Generation of complex polyhedral building models
by integrating stereo-aerial imagery and lidar data. Photogrammetric Engineering &
Remote Sensing, 76(5), 609-623. doi:10.14358/PERS.76.5.609.
Henn, A., Gröger, G., Stroh, V., and Plümer, L. (2013). Model driven reconstruction of roofs
from sparse LIDAR point clouds. ISPRS Journal of Photogrammetry and Remote
Sensing, 76, 17-29. doi:10.1016/j.isprsjprs.2012.11.004.
Henricsson, O., and Baltsavias, E. (1997) 3-D Building Reconstruction with ARUBA: A
Qualitative and Quantitative Evaluation. In: Gruen, A., Baltsavias, E.P., Henricsson, O.
(eds) Automatic Extraction of Man-Made Objects from Aerial and Space Images (II).
Monte Verità (Proceedings of the Centro Stefano Franscini Ascona). Birkhäuser, Basel.
doi:10.1007/978-3-0348-8906-3_7.
Heo, J., Jeong, S., Park, H. K., Jung, J., Han, S., Hong, S., and Sohn, H. G. (2013). Productive
high-complexity 3D city modeling with point clouds collected from terrestrial LiDAR.
Computers,

Environment

and

Urban

Systems,

41,

26-38.

doi:10.1016/j.compenvurbsys.2013.04.002.
Jancosek, M., and Pajdla, T. (2011). Multi-view reconstruction preserving weakly-supported
surfaces. In , 2011 IEEE Conference on Computer Vision and Pattern Recognition, pp.
3121-3128. IEEE. doi:10.1109/CVPR.2011.5995693.
Javanmardi, M., Gu, Y., Javanmardi, E., Hsu, L. T., and Kamijo, S. (2015). 3D building map
reconstruction in dense urban areas by integrating airborne laser point cloud with 2D
boundary map. In 2015 IEEE International Conference on Vehicular Electronics and
Safety, pp. 126-131. IEEE. doi:10.1109/ICVES.2015.7396906.
Lafarge, F., Descombes, X., Zerubia, J., and Pierrot-Deseilligny, M. (2008). Automatic
building extraction from DEMs using an object approach and application to the 3D-city
25

modeling. ISPRS Journal of Photogrammetry and Remote Sensing, 63(3), 365-381.
doi:10.1016/j.isprsjprs.2007.09.003.
Ledoux, H., and Meijers, M. (2011). Topologically consistent 3D city models obtained by
extrusion. International Journal of Geographical Information Science, 25(4), 557-574.
doi:10.1080/13658811003623277.
Li, X., Lv, Z., Hu, J., Zhang, B., Shi, L., and Feng, S. (2015). XEarth: A 3D GIS platform for
managing massive city information. In 2015 IEEE International Conference on
Computational Intelligence and Virtual Environments for Measurement Systems and
Applications, pp. 1-6. IEEE. doi:10.1109/CIVEMSA.2015.7158625.
Li, M., Nan, L., Smith, N., and Wonka, P. (2016). Reconstructing building mass models from
UAV images. Computers & Graphics, 54, 84-93. doi:10.1016/j.cag.2015.07.004.
Lillesand, T., Kiefer, R. W., and Chipman, J. (2015). Remote sensing and image interpretation.
John Wiley & Sons (ISBN-13: 978-1118343289).
Lin, H., Gao, J., Zhou, Y., Lu, G., Ye, M., Zhang, C., Liu, L., and Yang, R. (2013). Semantic
decomposition and reconstruction of residential scenes from LiDAR data. ACM
Transactions on Graphics, 32(4), 66. doi:10.1145/2461912.2461969.
Mitra, N. J., and Pauly, M. (2008). Symmetry for architectural design. In Advances in
Architectural

Geometry,

13-16.

Available

at:

http://www0.cs.ucl.ac.uk/staff/n.mitra/research/archi_symm/paper_docs/archi_symm_
agg_08.pdf, Accessed date: 6 January 2017.
Musialski, P., Wonka, P., Aliaga, D. G., Wimmer, M., Gool, L. V., and Purgathofer, W. (2013).
A survey of urban reconstruction. Computer Graphics Forum, 32(6), 146-177.
doi:10.1111/cgf.12077.
Open Geospatial Consortium (2012). OGC City Geography Markup Language (CityGML)
Encoding Standard. Available at: http://www.opengeospatial.org/standards/citygml,
Accessed date: 6 January 2017.
Over, M., Schilling, A., Neubauer, S., and Zipf, A. (2010). Generating web-based 3D City
Models from OpenStreetMap: The current situation in Germany. Computers,
Environment

and

Urban

Systems,

34(6),

496-507.

doi:10.1016/j.compenvurbsys.2010.05.001.
Peffers, K., Tuunanen, T., Rothenberger, M. A., and Chatterjee, S. (2007). A design science
research methodology for information systems research. Journal of Management
Information Systems, 24(3), 45-77. doi:10.2753/MIS0742-1222240302.

26

Poli, D., Remondino, F., Angiuli, E., and Agugiaro, G. (2015). Radiometric and geometric
evaluation of GeoEye-1, WorldView-2 and Pléiades-1A stereo images for 3D
information extraction. ISPRS Journal of Photogrammetry and Remote Sensing, 100,
35-47. doi:10.1016/j.isprsjprs.2014.04.007.
Poullis, C., and You, S. (2009). Automatic reconstruction of cities from remote sensor data. In
2009. IEEE Conference on Computer Vision and Pattern Recognition, pp. 2775-2782.
IEEE. doi:10.1109/CVPR.2009.5206562.
Rottensteiner, F., and Jansa, J. (2002). Automatic extraction of buildings from LIDAR data
and aerial images. International Archives of Photogrammetry Remote Sensing and
Spatial

Information

Sciences.

Available

at:

http://www.isprs.org/proceedings/XXXIV/part4/pdfpapers/204.pdf, Accessed date: 6
January 2017.
Sampath, A., and Shan, J. (2010). Segmentation and reconstruction of polyhedral building
roofs from aerial lidar point clouds. IEEE Transactions on Geoscience and Remote
Sensing, 48(3), 1554-1567. doi:10.1109/TGRS.2009.2030180.
Schnabel, R., Wahl, R., and Klein, R. (2007). Efficient RANSAC for point‐cloud shape
detection.

Computer

Graphics

Forum,

26(2),

214-226.

doi:10.1111/j.1467-

8659.2007.01016.x.
Singh, S. P., Jain, K., and Mandla, V. R. (2014). A new approach towards image based virtual
3D city modeling by using close range photogrammetry. ISPRS Annals of
Photogrammetry, Remote Sensing and Spatial Information Sciences, 2(5), 329.
doi:10.5194/isprsannals-II-5-329-2014.
Sun, S., and Salvaggio, C. (2013). Aerial 3D building detection and modeling from airborne
LiDAR point clouds. IEEE Journal of Selected Topics in Applied Earth Observations
and Remote Sensing, 6(3), 1440-1449. doi:10.1109/JSTARS.2013.2251457.
Tang, P., Huber, D., Akinci, B., Lipman, R., and Lytle, A. (2010). Automatic reconstruction
of as-built building information models from laser-scanned point clouds: A review of
related

techniques.

Automation

in

Construction,

19(7),

829-843.

doi:10.1016/j.autcon.2010.06.007.
Van Aken, J. E. (2005). Management research as a design science: Articulating the research
products of mode 2 knowledge production in management. British Journal of
Management, 16(1), 19-36. doi:10.1111/j.1467-8551.2005.00437.x.

27

Volk, R., Stengel, J., and Schultmann, F. (2014). Building Information Modeling (BIM) for
existing buildings—Literature review and future needs. Automation in Construction,
38, 109-127. doi:10.1016/j.autcon.2013.10.023.
Xiong, B., Elberink, S. O., and Vosselman, G. (2014). A graph edit dictionary for correcting
errors in roof topology graphs reconstructed from point clouds. ISPRS Journal of
Photogrammetry

and

Remote

Sensing,

93,

227-242.

doi:10.1016/j.isprsjprs.2014.01.007.
Wu, B., Yu, B., Wu, Q., Yao, S., Zhao, F., Mao, W., and Wu, J. (2017). A graph-based
approach for 3D building model reconstruction from airborne LIDAR point clouds.
Remote Sensing, 9(1), 92. doi:10.3390/rs9010092.
Xue, F., Chen, K., Liu, D., Niu, Y., and Lu, W. (2018). An optimization-based semantic
building model generation method with a pilot case of a demolished construction. In:
Chau K., Chan I., Lu W., and Webster C. (eds) Proceedings of the 21st International
Symposium on Advancement of Construction Management and Real Estate. Springer,
Singapore. doi:10.1007/978-981-10-6190-5_22.
Yan, J., Jiang, W., and Shan, J. (2016). A global solution to topological reconstruction of
building roof models from airborne LiDAR point clouds. ISPRS Annals of
Photogrammetry, Remote Sensing and Spatial Information Sciences, 379-386.
doi:10.5194/isprs-annals-III-3-379-2016.
Yang, R. J., and Zou, P. X. (2016). Building integrated photovoltaics (BIPV): costs, benefits,
risks, barriers and improvement strategy. International Journal of Construction
Management, 16(1), 39-53. doi:10.1080/15623599.2015.1117709.
You, R. J., and Lin, B. C. (2011). A quality prediction method for building model
reconstruction using LiDAR data and topographic maps. IEEE Transactions on
Geoscience and Remote Sensing, 49(9), 3471-3480. doi:10.1109/TGRS.2011.2128326.
Zhang, W., Wang, H., Chen, Y., Yan, K., and Chen, M. (2014). 3D building roof modeling by
optimizing primitive’s parameters using constraints from LiDAR data and aerial
imagery. Remote Sensing, 6(9), 8107-8133. doi:10.3390/rs6098107.
Zhu, L., Lehtomäki, M., Hyyppä, J., Puttonen, E., Krooks, A., and Hyyppä, H. (2015).
Automated 3D scene reconstruction from open geospatial data sources: Airborne laser
scanning and a 2D topographic database. Remote Sensing, 7(6), 6710-6740.
doi:10.3390/rs70606710.

28