Efficient assessment of window views in high-rise, high-density urban areas using 3D color City
Information Models
Maosu Li 1*, Fan Xue 2, Anthony G.O. Yeh 3
This is the peer-reviewed post-print version of the paper:
Li, M., Xue, F., & Yeh, A.G.O. (2023). Efficient assessment of window views in high-rise, highdensity urban areas using 3D color City Information Models. In Proceedings of the 18th
International Conference on Computational Urban Planning and Urban Management (CUPUM
2023), 20-22 June 2023, Montreal, Canada. Doi: 10.17605/osf.io/6yr5v
The final version of this paper is available at: https://osf.io/2hsbu The use of this file must follow the
Creative Commons Attribution Non-Commercial No Derivatives License.

Abstract
5

Urban-scale quantification of window views can inform housing selection and valuation, landscape
management, and urban planning. However, window views are numerous in high-rise, high-density urban
areas and current automatic assessments of window views are inaccurate and time-consuming. Thus, both
accurate and efficient assessment of window views is significant in improving the automation for urbanscale window view applications. The paper presents an automatic, accurate, and efficient assessment of

10

window view indices (WVIs) of greenery, sky, waterbody, and construction using 3D color City Information
Models (CIMs). The workflow includes: i) 3D semantic segmentation of photorealistic CIM and Digital
Surface Model (DSM), and ii) batch computation of WVIs. Experimental results showed the estimated WVIs
were more accurate (RMSE < 0.01), and the proposed method was more efficient (3.68 times faster) than Li
et al.’s (2022) 2D semantic segmentation. Thus, the proposed method can facilitate large-scale WVI

15

assessment and update in healthy high-rise, high-density urban development.
Keywords
Window view, Efficient assessment; High-rise, high-density cities; City Information Model; 3D semantic
segmentation; Urban computing.
1 Introduction

20

A high-quality window view with more greenery, sky, waterbody, and fewer construction elements is
treasured by urban dwellers, especially in high-rise, high-density urban areas. The recognized benefits of a
high-quality window view include stress relief, life satisfaction, and productivity improvement [1, 2, 3]. In
the post-Covid-19 era, the benefits of window views are further amplified for urban dwellers because of
increasingly long-term indoor occupation and reduced social activities [4].

25

Assessment of window views can help provide quantified evidence for informed housing selection and
Department of Urban Planning and Design, The University of Hong Kong, Hong Kong SAR, China
* Corresponding author, Email: maosulee@connect.hku.hk
2
Department of Real Estate and Construction, The University of Hong Kong, Hong Kong SAR, China
3
Department of Urban Planning and Design, The University of Hong Kong, Hong Kong SAR, China
1

valuation, landscape management and urban planning, and new building design. For example, renters and
purchasers can holistically sort and select rooms by window view indices [5]. Property agencies can precisely
value the room using quantified window view indicators [6, 7]. Urban planners and architectural designers
30

can leverage the automatic assessment tool to compare effects of different plans and designs on window
views for living environment improvement [8, 9].
However, window views are numerous especially in high-rise, high-density urban areas and often changed
in large numbers with the vertical development of neighborhoods. Fine-scale quantification of large-scale

35

window views can consume plenty of manpower or computation resources [10, 5]. The high time cost often
hinders urban planners, designers, and other decision-makers to assess and update the window view indices
and examine different plans and designs for window view optimization [10, 11]. Thus, both efficient and
accurate quantification of window view features is significant in advancing the window view assessment for
urban-scale applications, such as housing selection and valuation, landscape management, and urban

40

planning.
Previously, window views are often collected through onsite photography and manual modeling in
psychology, built environment, and urban health fields [12, 13]. These methods are high-cost, laborious, and
thus unscalable to the urban scale. Recently, simulation methods i.e., visibility analysis and view

45

photography have been used to assess urban-scale floor-level and window views [14, 5]. Generally,
traditional visibility analysis on a large-scale landscape view assessment still shows a preference for
oversimplified models, such as Digital Surface Model (DSM) and polygon building models [15]. In
comparison, 3D photorealistic City Information Models (CIMs) are less used due to more intensive
intervisibility computations. By contrast, direct view photography on 3D photorealistic CIMs can capture

50

photorealistic window views more easily thanks to well-developed OpenGL rendering techniques [16].
Specifically, urban-scale photorealistic window views have been generated and assessed using 3D
photorealistic CIM and deep transfer learning [5]. However, the method is inaccurate for processing closerange window views and less efficient due to the repetitive segmentation of window views without reusing
the shared intermediates. Thus, the next generation of assessment method needs to improve the processing

55

efficiency for an accurate urban-scale window view assessment.
This study presents an automatic, accurate, and efficient assessment method for urban-scale window
views. Specifically, we extend an assessment of the window view index (WVI) defined in Li et al.’s [5]
method for color window view images generated from 3D color CIMs. A two-step workflow using

60

photorealistic CIM and 3D semantic segmentation aims to improve both the accuracy and efficiency of the
current window view assessment for urban-scale applications. The remainder is arranged as follows: research
methods, experiment, discussion, and conclusion.
2 Research methods
Fig. 1 shows the workflow of the proposed method for efficiently assessing the urban-scale window views.

65

The input of the method comprises a photorealistic CIM, Building Information Models (BIMs), a DSM, and
a Normalized difference vegetation index (NDVI) map, as shown in Fig. 1a. Fig. 1b and 1c show the method
including two steps: i) 3D semantic segmentation of photorealistic CIMs and DSM and ii) batch computation

of WVIs using generated color window views. The outputs are four WVIs as shown in Fig. 1d.

70

Fig. 1. The workflow of the proposed method. (a) Input, a proposed two-step method comprising (b) 3D semantic segmentation
of photorealistic CIM and DSM and (c) batch computation of WVIs, and (d) output.

2.1 Definition of Window View Index (WVI)
This study applies the WVIs defined in [5] for color window view images generated from a 3D color scene
as shown in Fig. 2b and 2c. Instead of using a photorealistic scene as shown in Fig. 2a, Fig. 2c and 2d show
75

the WVI is defined as a ratio on an 8-bit RGB color view image c with n pixels,
WVIl = |{p| p∈c, m(pcolor) = l}| ⁄ n, l ∈L,
L= {‘greenery’, ‘waterbody’, ‘sky’, ‘construction’},

(1)

where m is a mapping function to determine the semantic label l of a pixel p based on its color pcolor. For
example, Fig. 2c shows that the label of p in green is recognized as greenery and the label of p in red is
recognized as construction elements. The semantic label set L in this study includes greenery (green),
waterbody (water), sky, and construction (const.). And | · | is a function to calculate the total number of pixels
80

with the label l in the image c. Thus, the value of WVIl ranges from 0 to 1, as shown in Fig. 2d. And the
higher the value, the larger the proportion of l in the window view image.

Fig. 2. The WVIl [5] defined on a color window view image.

2.2 3D segmentation of photorealistic CIM and DSM for a 3D color scene
85

To set up a 3D scene colored by L as shown in Fig. 2b, we implement a 3D semantic segmentation on
photorealistic CIM and DSM, respectively. First, a deep learning model KPConv [17] is trained to predict l
for every vertex of the photorealistic CIM triangles. Specifically, we evenly sample the surface of 3D
photorealistic CIM into dense point clouds. Thereafter, we train the KPConv model on point clouds annotated
by L, finetune the parameters, and then apply the well-trained model to predict the rest of the dense point

90

clouds. Last, each point is assigned a semantic label l and the triangle vertex of CIM is assigned a label same
to that of the closest point.
Then, distant landscape elements represented by DSM are segmented into greenery, construction, and
waterbody using the NDVI map. Since the distant view elements are rendered at a lower resolution, we use

95

DSM to represent the distant landscape layer according to the observable saturation of measurement
accuracy [10]. Specifically, the NDVI map is first segmented into categories of greenery, construction, and
waterbody using Equation 2.
greenery, NDVIg > 0.1,
construction,
0 ≤ NDVIg ≤ 0.1,
lg = �
waterbody, NDVIg = no data,

(2)

where g is the pixel of the NDVI map, and lg and NDVIg are the semantic label and NDVI value of the pixel
g, respectively. Threshold values are set for distinguishing pixels of greenery, construction, and waterbody
100

following guidance of local geospatial contexts. For example, we set 0.1, 0, and no data for experimental
tests of Hong Kong in Section 3, following the NDVI-related report [18] of Planning Department, Hong

Kong SAR. Then, each pixel of the DSM is assigned a label l by segmented NDVI map through resampling
and geo-registration.
105

Last, we color the 3D CIM, DSM, and the sky layer for a fully 3D color scene using Equation 3,
RGB (0, 255, 0),
RGB (0, 0, 255),
color(vl ) = �
RGB (255, 255, 255),
RGB (255, 0, 0),

l = greenery,
l = waterbody,
l = sky,
l = construction,

(3)

where color is the function to map the semantic label l of a vertex v of the 3D scene into an 8-bit RGB color.
For example, the vertex v with the semantic label l = greenery is colored in green, RGB (0, 255, 0), whereas
the vertex with l = waterbody is in blue, RGB (0, 0, 255).
2.3 Batch computation of WVIs using color view images
110

This step batch computes WVIs using color window view images generated from the 3D color scene as
shown in Fig. 1c. The virtual camera with a field of view at 60 degrees is first placed on each window site
of the 3D color scene according to the window location information provided by the BIMs. Specifically, the
virtual camera is set on the window center with tilt = 0, pitch = 0, and heading = (heading of the window) to
capture the outside view. Then, the color view images are captured and saved in batch. Last, we set the

115

function m in Equation 1 as below regarding Equation 3,
greenery,
waterbody,
m(pcolor) = �
sky,
construction,

pcolor = RGB (0, 255, 0),
pcolor = RGB (0, 0, 255),
pcolor = RGB (255, 255, 255),
pcolor = RGB (255, 0, 0),

(4)

The total number of pixels of each l in every window view image is summarized for computing four WVIs
using Equations 4 and 1.
3 Experiment
3.1 Experimental settings
120

We chose 207 buildings in To Kwa Wan, Kowloon Peninsula of Hong Kong to test the feasibility of the
proposed method, as shown in Fig. 3a and 3b. The study area is located in the top density zone according to
the Hong Kong Planning Standards and Guidelines [19]. And the maximum building height gap is 103.15
m. There are 44,909 windows on the 207 buildings, where each building owns 217 windows on average. The
3D photorealistic CIM and DSM were collected from the Lands Department as shown in Fig. 3c and 3e [20,

125

21], while the window location information was extracted from BIM data shared by the Urban Renewal
Authority as shown in Fig. 3d [22]. Fig. 3f shows the 30-m NDVI map collected from [23].

Fig. 3. Study area of To Kwa Wan, Hong Kong. (a) Location, (b) 207 buildings, (c) 3D photorealistic CIM, (d) BIMs, (e) DSM,
and (f) NDVI map.
130

The computational environment was set up as follows. This study used a workstation with an Intel i911900K CPU (3.50 GHz, 16 cores), 64G memory, one 24G Nvidia GeForce RTX 3090 graphic card, and
Windows 10 operating system (64-bit). 3D semantic segmentation was implemented in a Docker (ver.
20.10.12) container with the environment of Pytorch (ver. 1.10.0) and Python (ver. 3.7.11). We used Cesium
(ver. 1.99) to set up the 3D color scene and generate the color window view images. The WVIs were

135

quantified through a developed Python program. To compare the accuracy and efficiency, we implemented
Li et al.’s [5] 2D image segmentation-based window view assessment method in the same workstation. The
same version of Cesium was used to generate photorealistic window view images. And we followed the
settings of [5] to quantify four WVIs using Deeplab V3+ [24] within the same Docker container. The oneoff annotation of 20 tiles of sampled point cloud for our 3D semantic segmentation and 110 window view

140

images for Li et al.’s 2D semantic segmentation consumed about 10 person-hours each. In addition, we
manually annotated another 100 window view images selected in Section 3.2 to test the accuracy of both
assessment results. Distant landscape elements beyond 2,000 m were colored using DSM and NDVI
referring to [10]. Same to Li et al.’s window view settings, window view images with 900 × 900 pixels were
generated in batch.

145

3.2 Experimental results
Overall, the performance of the trained KPConv in our method reached a mIoU of 0.91 for the segmentation
of greenery, waterbody, and construction from the 3D photorealistic CIM. The values of R2 for estimating
four WVIs using Li et al.’s 2D method were above 0.95. Specifically, we randomly generated 100 window

view images from the 207 buildings for the accuracy and efficiency test. Table 1 shows that the assessment
150

results of our method reached a higher accuracy (RMSE < 0.01) than those of Li et al.’s 2D method.
Observable reason was more accurate segmentation of close-range window views using our method as
shown in Figure 4. Figure 4b shows there existed misclassification of close-range construction elements into
other landscape elements such as vegetation using Li et al.’s 2D method. By contrast, the results of our 3D
method were more satisfactory due to the holistic 3D segmentation of the photorealistic CIM and DSM. Fig.

155

4c shows close-range view elements were correctly recognized as construction.

Fig. 4. Comparison of the segmentation accuracy against close-range view elements using two methods. (a) Example
photorealistic window view generated from CIMs, (b) segmented window view using Li et al.’s [5] 2D method, and (c) our
color window view generated on segmented 3D CIMs.
160

Table 1. Comparison of assessment accuracy of two methods on 100 test windows
RMSE
Li et al.’s 2D method

Our 3D method

Improvement

WVIgreen

0.0283

0.0059

79.15%

WVIwater

0.0243

0.0048

80.25%

WVIsky

0.0098

0.0044

55.10%

WVIconst.

0.0405

0.0092

77.28%

Average

0.0257

0.0061

76.26%

Meanwhile, Table 2 shows the processing of 100 window view images using our method saved 73% of the
total time cost compared to Li et al.’s 2D method. The window view generation of our method was efficient
(3.59 times faster) without the preparation and rendering of the 3D CIM texture images. The quantification
165

of WVIs improved 81% of the efficiency by avoiding the repetitive 2D segmentation of window view images.
In summary, the results in Tables 1 and 2 confirmed that our method was both accurate and efficient for
urban-scale window view assessment.

Table 2. Comparison of computational time of the two methods (average of 100 windows)
Step

Processing

Li et al.’s 2D
method

170

Our 3D method

Improvement

1

Window view generation

1.94 s

0.54 s

72%

2

Quantification of WVIs

0.16 s

0.03 s

81%

Total

2.10 s

0.57 s

73%

4 Discussion
Both accurate and efficient window view assessment method is important for examining and updating urbanscale window view indices, which can benefit precise housing valuation and selection, landscape
management and urban planning, and new building design. For example, an efficient assessment method can
encourage urban planners and property agencies to quantify and timely update the latest window view indices

175

of greenery, waterbody, sky, and construction at the urban scale. The updated accurate window view indices
of greenery, waterbody, sky, and construction can help urban planners to identify the urban areas with less
nature exposure for prioritized planning and design practices such as planning more blue-green spaces [11].
Property agencies can use large-scale accurate window view indicators for precise valuation. Instead of the
previous qualitative judgment, housing purchasers and renters can select rooms with a satisfactory view by

180

comparing and sorting quantified view indicators of rooms holistically. In addition, leveraging an accurate
and efficient assessment tool, architectural designers can compare the effects of different plans and design
drafts on window views with a low time cost.
Recently, the urban-scale window view assessment has been implemented at the urban scale using 3D

185

photorealistic CIM and deep transfer learning. However, the method is still inaccurate and inefficient for
processing urban-scale window views. The root reasons are inaccurate 2D segmentation of close-range
window view elements, the high workload of texture rendering, and 2D segmentation with repeated
computations. For example, the close-range window view elements can be segmented into incorrect
categories due to the limited resolution of the CIM texture images. The generation of window view images

190

is less efficient due to the high workload of CIM texture rendering. Landscape elements shared by numerous
windows are repetitively segmented for every window view without reusing the intermediates.
The contribution of this paper is to present a novel method to improve the accuracy and efficiency of
window view assessment. A 3D scene colored by four view types (e.g., greenery, waterbody, sky, and

195

construction) using 3D CIMs and 3D semantic segmentation is set up to avoid the close-range view
segmentation and repetitive computation, which effectively improves the accuracy and efficiency for an
urban-scale window view assessment. Experimental results show that the assessment results of our method
were more accurate (RMSE < 0.01) and the proposed method was 3.68 times faster than Li et al.’s [5] 2D
semantic segmentation.

200

There still exist some limitations in the study. First, a full 3D color scene needs to be prepared for assessing
four WVIs. Small-scale quantification, e.g., for one or two window views may not afford the large-scale but
one-off preprocessing cost. In other words, the proposed method is born for a large-scale window view

assessment. In addition, since neighborhood windows especially on the upper stories of high-rise buildings
205

often share similar views, the repetitive quantification regardless of the spatial context may hinder the
efficiency of the urban-scale assessment. Thus, future directions include simplification of preprocessing of
3D CIMs and mining of window view patterns of different types, locations, and storeys of buildings and
blocks for a more efficient urban-scale window view assessment.
5 Conclusion

210

Window views are numerous and changed in large numbers with the vertical development of the urban
landscapes especially in high-rise, high-density urban areas. The current methods are inaccurate and
inefficient to process window views at the urban scale due to inaccurate segmentation and high workloads
of repetitive computation. Thus, automatic, accurate, and efficient window view assessment is significant
for assessing and updating thousands of window view indices for urban-scale applications such as housing

215

selection and valuation, landscape management, and urban planning.
This study extends the assessment of four defined Window View Indices (WVIs) of greenery, waterbody,
sky, and construction for 3D color window view images generated from a 3D color scene. The 3D scene set
up by photorealistic CIM and DSM is first colored by the four types of landscape elements using 3D semantic

220

segmentation. Then, 3D color window view images are generated from the 3D color scene by placing the
virtual camera on the window sites. Last, we programmatically summarize the pixels in different colors for
computing four WVIs. Experimental results in To Kwa Wan, Hong Kong showed that our assessment results
were more accurate (RMSE < 0.01) and the proposed method was 3.68 times faster than the state-of-the-art
method. The both accurate and efficient assessment of window views can be useful for encouraging property

225

agencies, urban planners, and other decision-makers to timely assess and update urban-scale window view
indices for housing valuation and selection, landscape urban management, and urban planning. Future work
includes simplifying the preprocessing of 3D CIMs and pattern mining of window views for a more efficient
urban-scale window view assessment.
References
1.

Ulrich, R. S.: View through a window may influence recovery from surgery. Science, 224(4647), 420421 (1984).

2.

Kaplan, R.: The role of nature in the context of the workplace. Landscape and Urban Planning, 26(14), 193-201 (1993).

3.

Kaplan, R.: The nature of the view from home: Psychological benefits. Environment and Behavior,
33(4), 507-542 (2001).

4.

Labib, S. M., Browning, M. H., Rigolon, A., Helbich, M., James, P.: Nature's contributions in coping
with a pandemic in the 21st century: A narrative review of evidence during COVID-19. Science of
The Total Environment, 833, 155095 (2022).

5.

Li, M., Xue, F., Wu, Y., Yeh, A. G.: A room with a view: Automated assessment of window views for
high-rise high-density areas using City Information Models and transfer deep learning. Landscape
and Urban Planning, 226, 104505 (2022).

6.

Jim, C.-Y., Chen, W. Y.: Value of scenic views: Hedonic assessment of private housing in Hong Kong.
Landscape and Urban Planning, 91(4), 226-234 (2009).

7.

Turan, I., Chegut, A., Fink, D., Reinhart, C.: Development of view potential metrics and the financial
impact of views on office rents. Landscape and Urban Planning, 215, 104193 (2021).

8.

Li, W., Samuelson, H., A new method for visualizing and evaluating views in architectural design.
Developments in the Built Environment, 1, 100005 (2020).

9.

Laovisutthichai, V., Li, M., Xue, F., Lu, W., Tam, K., Yeh, A. G.: CIM-enabled quantitative view
assessment in architectural design and space planning. In: Proceedings of 38th International
Symposium on Automation and Robotics in Construction (ISARC 2021), pp. 65-72. IAARC, Dubai
(2021).

10.

Cimburova, Z., Blumentrath, S.: Viewshed-based modelling of visual exposure to urban greenery–
An efficient GIS tool for practical planning applications. Landscape and Urban Planning, 222,
104395 (2022).

11.

Li, M., Xue, F., Yeh, A. G.: Bi-objective analytics of 3D visual-physical nature exposures in high-rise
high-density cities for landscape and urban planning. Landscape and Urban Planning, 233, 104714
(2023).

12.

Lottrup, L., Stigsdotter, U. K., Meilby, H., Claudi, A. G.: The workplace window view: a determinant
of office workers’ work ability and job satisfaction. Landscape Research, 40(1), 57-75 (2015).

13.

Masoudinejad, S., Hartig, T.: Window view to the sky as a restorative resource for residents in densely
populated cities. Environment and Behavior, 52(4), 401-436. (2020).

14.

Yu, S., Yu, B., Song, W., Wu, B., Zhou, J., Huang, Y., Wu, J., Zhao, F., Mao, W.: View-based greenery:
A three-dimensional assessment of city buildings’ green visibility using Floor Green View Index.
Landscape and Urban Planning, 152, 13-26 (2016).

15.

Inglis, N. C., Vukomanovic, J., Costanza, J., Singh, K. K.: From viewsheds to viewscapes: Trends in
landscape visibility and visual quality research. Landscape and Urban Planning, 224, 104424 (2022).

16.

Li, M., Xue, F., Yeh, A. G., Lu, W.: Classification of photo-realistic 3D window views in a highdensity city: The case of Hong Kong. In: Proceedings of the 25th International Symposium on
Advancement of Construction Management and Real Estate, pp. 1339-1350. Springer, Wuhan (2021).

17.

Thomas, H., Qi, C. R., Deschaud, J.-E., Marcotegui, B., Goulette, F., Guibas, L. J.: Kpconv: Flexible
and deformable convolution for point clouds. In: 2019 IEEE/CVF International Conference on
Computer Vision (ICCV), pp. 6411-6420. IEEE, Seoul, Korea (South) (2019).

18.

HKPlanD: Urban Climatic Map and Standards for Wind Environment. Planning Department,
Government of Hong Kong SAR, Hong Kong (2009).

19.

HKPlanD: Hong Kong Planning Standards and Guidelines. Planning Department, Hong Kong SAR,
Hong Kong (2018).

20.

HKLandsD: 3D Visualization Map. Lands Department, Goverment of Hong Kong SAR, Hong Kong
(2017).

21.

HKLandsD: Digital Surface Model from 2020 LiDAR Survey. Lands Department, Government of
Hong Kong SAR, Hong Kong (2020).

22.

URA: 3D Intelligent Map of an area in To Kwa Wan. Urban Renewal Authority, Hong Kong (2022).

23.

Morgan, B., Guénard, B.: New 30 m resolution Hong Kong climate, vegetation, and topography
rasters indicate greater spatial variation than global grids within an urban mosaic. Earth System
Science Data, 11(3), 1083-1098 (2019).

24.

Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable
convolution for semantic image segmentation. In: Proceedings of the European conference on
computer vision (ECCV), pp. 801-818. Springer, Munich (2018).

230