
==== Front
Data Brief
Data Brief
Data in Brief
2352-3409
Elsevier

S2352-3409(24)00848-5
10.1016/j.dib.2024.110885
110885
Data Article
Building façade datasets for analyzing building characteristics using deep learning
Wang Seunghyeon
Park Sangkyun psk930706@hanyang.ac.kr
⁎
Park Sungman psm533@hanyang.ac.kr
⁎
Kim Jaejun
Department of Architectural Engineering, Hanyang University, Seungdong-Gu, Seoul 133791, Korea
⁎ Corresponding authors. psk930706@hanyang.ac.krpsm533@hanyang.ac.kr
28 8 2024
12 2024
28 8 2024
57 11088511 7 2024
1 8 2024
23 8 2024
© 2024 The Author(s)
2024
https://creativecommons.org/licenses/by-nc/4.0/ This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
Building characteristics are vital across various domains such as construction management and architectural design. Static Street View Images (SSVIs) can be utilized with deep learning techniques to interpret building characteristics without the need for a physical visit. Deep learning approaches have demonstrated a high capability for generalization, enabling the automation of manual tasks related to image analysis. However, there is no publicly available labeled dataset of building characteristics from building facade images for training deep learning models. In this article, we focus on constructing a dataset for four different tasks: classification of the number of stories, classification of building typologies, classification of exterior cladding materials, and classification of usable SSVIs. To develop deep learning models, this article constructed a dataset sourced from London and Scotland in the UK. The dataset was labeled by annotation experts. While the focus of this research is on specific tasks, the raw dataset can be used for other purposes (e.g., ascertaining the age of buildings or identifying window types) by annotating the data for the corresponding tasks.

Keywords

Building characteristics
Street view images
Deep learning
Convolutional neuralnnetwork
==== Body
pmcSpecifications TableSubject	Environmental Engineering, Artificial Intelligence	
Specific subject area	Deep Learning based detection from building façade images for building characteristics	
Type of data	Image	
Data collection	This dataset was constructed for four tasks: classification of the number of stories, classification of building typologies, classification of exterior cladding materials, and classification of usable SSVIs. SSVIs were collected automatically based on geographical information (i.e., latitude and longitude) from an official government website. Each image was annotated according to its corresponding task. All images are in PNG format with a size of 640 × 640 pixels. The dataset was divided into subsets for training, validation, and testing purposes, organized into five folders suitable for training deep learning models. To increase the amount of training data, image augmentation techniques were applied only to the training set. The total size of the dataset is 5.05 GB, and a ZIP file is provided.	
Data source location	Country: the UK, Region: London and Scotland	
Data accessibility	Repository name: FigShare
Data identification number: https://doi.org/10.6084/m9.figshare.24979947.v1
Direct URL to data: https://figshare.com/articles/dataset/Dataset_of_building_characteristics_from_building_fa_ade_images/25931941	
Related research article	Effectiveness of Image Augmentation Techniques on Detection of Building Characteristics from Street View Images Using Deep Learning [1]. Automated detection of exterior cladding material in urban area from street view images using deep learning [2]	

1 Value of the Data

• This dataset was constructed for four tasks: classification of the number of stories, classification of building typologies, classification of exterior cladding materials, and classification of usable SSVIs.

• As the entire dataset is labeled, supervised learning based on a data-driven approach is recommended to develop models for each task.

• An artificially generated dataset based on real data was also constructed, which can be used as a benchmark dataset for developing new augmentation techniques.

• In addition to the annotated dataset, the raw data can be used for new tasks such as the classification of window types and the segmentation of window-to-wall ratios.

2 Background

Static Street View Images (SSVIs) offer a human-eye view of building facades, capturing details such as materials and window types. Utilizing SSVIs is cost-effective and quicker for identifying building characteristics compared to physical site visits. In particular, Google Street View (GSV) covers over 16 million kilometers across 83 countries, providing easy access to images via geographical coordinates and addresses [3]. Deep learning approaches have shown high generalization capability, automating many tasks using SSVIs [4]. Existing research provides façade datasets for tasks such as window state detection [5], tile defect segmentation, window detection [6], and building segmentation [7]. Even with similar images, different annotations are required for different tasks, making the process time-consuming and labor-intensive. However, no publicly available raw and labeled data exists for four specific tasks using SSVIs: classifying the number of stories, building typologies, exterior cladding materials [2], and usable SSVIs. This article aims to provide labeled datasets for these four tasks to encourage the development of robust deep learning techniques. The dataset can support future research and help many real-world applications since these tasks are ongoing research areas, as the extracted information can be applied to urban investigation, construction management, and building performance simulation.

3 Data Description

The following chapters describe in detail how four different tasks are constructed: classification of the number of stories, classification of building typologies, classification of exterior cladding materials, and classification of usable SSVIs.

3.1 Dataset of number of stories

This dataset, designed for classifying the number of stories in buildings, consists of images categorized by their respective floors (1F, 2F, 3F, 4F, 5F, or Others). Since the dataset may contain images not related to building characteristics captured, an 'Others' category was incorporated. Fig. 1 illustrates how images are organized with each label. Starting with 700 raw images, the dataset was split into two subsets: 418 for training (80 %) and 282 for testing (20 %). Augmentation techniques were applied exclusively to the training subset, resulting in nine different datasets. Each augmentation technique (brightness, contrast, perspective, rotation, scale, shear, translation), along with a combination of all these techniques, effectively doubled the number of images. Table 1 below provides a detailed distribution of the datasets based on the number of stories.Fig. 1 Labelling structure of dataset for number of stories.

Fig 1

Table 1 Detailed distribution of datasets for number of stories.

Table 1Purpose	Number of images	Class	
1F	2F	3F	4F	5F	Others	
Original training	418	36	100	97	55	70	60	
Original + brightness	836	72	200	194	110	140	120	
Original + contrast	836	72	200	194	110	140	120	
Original + perspective	836	72	200	194	110	140	120	
Original + rotation	836	72	200	194	110	140	120	
Original + scale	836	72	200	194	110	140	120	
Original + shear	836	72	200	194	110	140	120	
Original + translation	836	72	200	194	110	140	120	
Original + a sum of techniques	3344	288	800	776	440	560	480	
Test	282	25	67	65	38	47	40	
Total	9896	853	2367	2296	1303	1657	1420	

3.2 Dataset of building typologies

This dataset, designed for classifying building typologies, includes images categorized by type (residential and non-residential). Since the dataset may contain images not related to building characteristics, an 'Others' category was incorporated. Fig. 2 illustrates how images are organized with each label. Initially, 450 raw images were collected and divided into two subsets: 270 images for training (80 %) and 180 images for testing (20 %). Augmentation techniques were applied only to the training subset, creating nine distinct datasets. These techniques (brightness, contrast, perspective, rotation, scale, shear, translation) and their combinations effectively doubled the number of images. Table 2 below details the distribution of datasets based on building typologies.Fig. 2 Labelling structure of dataset for building typologies.

Fig 2

Table 2 Detailed distribution of datasets for building typologies.

Table 2Purpose	Number of images	Class	
Residential	Non-residential	Others	
Original training	270	90	90	90	
Original + brightness	540	180	180	180	
Original + contrast	540	180	180	180	
Original + perspective	540	180	180	180	
Original + rotation	540	180	180	180	
Original + scale	540	180	180	180	
Original + shearing	540	180	180	180	
Original + translation	540	180	180	180	
Original + a sum of techniques	2160	720	720	720	
Test	180	60	60	60	
Total	6390	2130	2130	2130	

3.3 Dataset of SSVIs

3.3.1 Whole building façade

The dataset for classifying images into three categories for analyzing whole building facades includes: usable, potential, and non-usable. Fig. 3 shows the organization of images with each label. Initially, 1000 raw images were collected and divided into three subsets: 600 for training (60 %), 200 for validation (20 %), and 200 for testing (20 %). Six augmentation techniques (brightness, contrast, perspective, rotation, scale, and shear) were applied exclusively to the training subset, resulting in 3600 augmented images. Table 3 provides a detailed distribution of the datasets based on the building facade classifications.Fig. 3 Labelling structure of dataset for whole building facade.

Fig 3

Table 3 Detailed distribution of datasets for the whole building façade.

Table 3Purpose	Number of images	Class	
Usable	Potential	Non-usable	Total	
Training	600	230	202	168	600	
Augmentation	3600	1380	1212	1008	3600	
Validation	200	77	67	56	200	
Test	200	75	68	57	200	
Total	4600	1762	1549	1289	4600	

3.3.2 First story building

The dataset for analyzing the number of stories in buildings comprises images categorized by type: usable, potential, or non-usable. Fig. 4 depicts the organization of images for each category. Initially, 700 raw images were collected and divided into three subsets: 418 images for training (60 %), 140 images for validation (20 %), and 142 images for testing (20 %). Six augmentation techniques (brightness, contrast, perspective, rotation, scale, and shear) were applied solely to the training subset, generating 3600 augmented images. Table 4 provides a detailed distribution of the datasets based on the number of stories.Fig. 4 Labelling structure of dataset for first story building.

Fig 4

Table 4 Detailed distribution of datasets focusing on the first story of a building.

Table 4Purpose	Number of images	Class	
Usable	Potential	Non-usable	Total	
Training	600	221	205	174	600	
Augmentation	3600	1326	1230	1044	3600	
Validation	200	74	68	58	200	
Test	200	72	69	59	200	
Total	4600	1693	1572	1335	4600	

3.4 Dataset of exterior cladding metarials

The dataset used for analyzing exterior cladding materials categorizes images by type: brick, concrete, glass, stone, mixed, and others. The “others” category includes uncommon textures and appearances, such as wood, metal, and synthetic sidings. Figs. 4 and 5 illustrate how images are organized with each label. In London, starting with 1550 raw images, the dataset was split into three subsets: 928 for training (60 %), 311 for validation (20 %), and 311 for testing (20 %). In Scotland, starting with 1017 raw images, the dataset was divided into 608 for training (60 %), 204 for validation (20 %), and 205 for testing (20 %).Fig. 5 Labelling structure of dataset for exterior cladding material (London).

Fig 5

Six augmentation techniques (brightness, contrast, perspective, rotation, scale, and shear) were applied exclusively to the training subsets. This process generated 5568 augmented images for the London dataset and 3648 augmented images for the Scotland dataset. Tables 5 and 6 below provide a detailed distribution of the datasets based on London and Scotland, respectively (Fig. 6).Table 5 Detailed distribution of datasets in London.

Table 5Purpose	Number of images	Class	
Brick	Concrete	Glass	Stone	Mixed	Others	Total	
Training	928	265	137	43	24	357	102	928	
Augmentation	5568	1590	822	258	144	2142	612	5568	
Validation	311	91	45	14	8	119	34	311	
Test	311	86	47	15	9	120	34	311	
Total	7118	2032	1051	330	185	2738	782	7118	

Table 6 Detailed distribution of datasets in Scotland.

Table 6Purpose	Number of images	Class	
Brick	Concrete	Glass	Stone	Mixed	Others	Total	
Training	608	46	80	9	174	241	58	608	
Augmentation	3648	276	480	54	1044	1446	348	3648	
Validation	204	15	26	3	61	80	19	204	
Test	205	16	28	4	56	81	20	205	
Total	4665	353	614	70	1335	1848	445	4665	

Fig. 6 Labelling structure of dataset for exterior cladding material (Scotland).

Fig 6

4 Experimental Design, Materials and Methods

4.1 Original dataset preparation

For collecting SSVIs, building addresses in London and Scotland from the OS Data Hub, the national mapping agency of the UK (https://osdatahub.os.uk/) were obtained. A sample of buildings was randomly selected from the North-West (NW) and Eastern (E) regions, as depicted in Fig. 7. Using these addresses, the GSV Static API (https://developers.google.com/maps/documentation/streetview/intro) was used to retrieve and download images of the buildings.Fig. 7 The area for data collection in London, and Scotland.

Fig 7

As detailed in Table 7, specific API parameters were manually adjusted to optimize visual cues for identifying building characteristics across four different tasks. The “Field of View (FOV)” parameter, which determines the zoom level or scope of the scene, was set between 10 and 50°. The “pitch” parameter, indicating the cameraʼs vertical angle relative to the street view vehicle, was adjusted to between 25 and 30°. The “heading” parameter, specifying the camera's horizontal direction, ranged from 33 to 55°. Images were obtained at a resolution of 640 × 640 pixels, the maximum available on the GSV platform.Table 7 The used camera parameters.

Table 7Parameters	
	Minimum	Maximum	Minimum	Maximum	
FOV	10	35	30	55	
Pitch	20	50	20	50	
Heading	30	60	30	60	

4.2 Augmenting the training datasets

The parameters of image augmentation techniques significantly impact the quality of images and improve accuracy [8]. These parameters were chosen through a trial-and-error process to generate images that are as realistic as possible. Examples of artificially augmented images are shown in Fig. 8.Fig. 8 Examples of augmented images.

Fig 8

For brightness adjustment, pixel intensity was randomly altered by adding a value between −30 and +30 to the intensity of the original images. The contrast level for the augmented images was modified by varying the alpha value (α) randomly within a range of 0.5 to 2.0. In scale augmentation, images were independently scaled along the x and y axes to random values ranging from 80 % to 120 % of their original size, accommodating the diverse dimensions of building features such as buildings and windows. Perspective changes were simulated through transformation by adjusting the 'scale' parameter within a range of 0.01 to 0.15, applied randomly and independently to each image.

Rotation augmentation involved rotating the original images by a degree between −25 and +25, with the rotation parameter denoted by θ. For translation, the extent and direction of image shifting were controlled by adjusting tx and ty, allowing for horizontal and vertical movement within a range of −0.2 to +0.2 times the image's dimensions. Shearing augmentation was executed within a −15 to +15° range, introducing a skew to the image features by modifying the shearing parameters shx and shy.

Geometric transformations such as scaling, rotation, perspective shifts, translation, and shearing can cause parts of an image to extend beyond its original boundaries, resulting in voids where there is no image data. Conversely, some areas may become compressed, leading to pixel overlap and gaps. Deep learning methods typically require input data in the form of fixed-length vectors, necessitating that augmented images retain the same size. In this study, voids in the augmented images were filled with 255-value pixels, effectively padding these areas with white.

4.3 Annotation

In image classification tasks, the annotation process involves assigning a specific class to each image within the ground truth dataset. This article details the annotation process by analyzing visual cues to classify images for each task. Consequently, images were sorted and stored in designated folders corresponding to their classifications.

Limitations

In the realm of deep learning, the success of any model hinges on the quality and diversity of the data it is trained on [5,9,10]. One of the primary limitations of our dataset is its geographic specificity. Collected from buildings in London and Scotland, the dataset encapsulates the architectural styles, materials, and environmental conditions unique to these regions. While this provides a rich and detailed view of these locales, it may not capture the architectural diversity found in other parts of the world. Buildings in London and Scotland may have different facade characteristics compared to those in tropical climates, earthquake-prone areas, or regions with different historical and cultural influences. As a result, models trained on this dataset might not perform as well when applied to images from vastly different environments.

In addition, the raw dataset of constructed images can be used for other tasks, such as detecting window types (hinged and sliding) and identifying obstacles on building facades (e.g., cars and pedestrians) by newly annotating the images for different tasks. While the image resolution is sufficient for the four targeted tasks in this article, it may not be high enough for detailed tasks like crack detection, crack segmentation, and tile segmentation.

Despite these limitations, the dataset remains a valuable resource for developing deep learning models aimed at analyzing building facades. It provides a foundational basis for training and validating algorithms, paving the way for advancements in this field. However, to enhance the robustness and applicability of the models, it is essential to complement this dataset with additional data from diverse regions, building types, and temporal settings.

Ethics Statement

The authors have read and followed the ethical guidelines for publishing in Data in Brief, and they affirm that the present study does not involve human subjects, animal research, or data gathered from social media sites.

CRediT authorship contribution statement

Seunghyeon Wang: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. Sangkyun Park: Methodology, Writing – review & editing. Sungman Park: Conceptualization, Investigation, Formal analysis, Writing – review & editing. Jaejun Kim: Conceptualization, Methodology.

Data Availability

Building façade datasets for analyzing building characteristics using deep learning (Original data) (FigShare).

Acknowledgments

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of Competing Interest

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.
==== Refs
References

1 Han J. Kim J. Kim S. Wang S. Effectiveness of image augmentation techniques on detection of building characteristics from street view images using deep learning J. Constr. Eng. Manag. 150 2024 1 18 10.1061/JCEMD4.COENG-15075
2 Wang S. Han J. Automated detection of exterior cladding material in urban area from street view images using deep learning J. Build. Eng. 96 2024 1 18 10.1016/j.jobe.2024.110466
3 Zou S. Wang L. Detecting individual abandoned houses from google street view: a hierarchical deep learning approach ISPRS J. Photogramm. Remote Sens. 183 2021 234 250 10.1016/j.isprsjprs.2021.03.020
4 Shehadeh A. Alshboul O. Almasabha G. Slope displacement detection in construction: an automated management algorithm for disaster prevention Expert Syst. Appl. 237 2024 121505 10.1016/j.eswa.2023.121505
5 Wang S. Korolija I. Rovas D. Impact of traditional augmentation methods on window state detection CLIMA 2022 Conference 2022 1 8 10.34641/clima.2022.375
6 Cao M.T. Drone-assisted segmentation of tile peeling on building façades using a deep learning model J. Build. Eng. 2023 50 67 10.1016/j.jobe.2023.108063
7 Bai Y. Cao M. Wang R. Liu Y. Wang S. How street greenery facilitates active travel for university students J. Transp. Heal. 2022 10.1016/j.jth.2022.101393
8 Wang S. Kim M. Hae H. Cao M. Kim J. The development of a rebar-counting model for reinforced concrete columns: using an unmanned aerial vehicle and deep-learning approach J. Constr. Eng. Manag. 149 2023 1 13 10.1061/JCEMD4.COENG-13686
9 Fang Z. Roy K. Chen B. Sham C.W. Hajirasouliha I. Lim J.B.P. Deep learning-based procedure for structural design of cold-formed steel channel sections with edge-stiffened and un-stiffened holes under axial compression Thin-Walled Struct. 164 2021 1 19 10.1016/j.tws.2021.108076
10 Roy K. Lau H.H. Lim J.B.P. Finite element modelling of back back-to-back built built-up cold cold-formed stainless-steel lipped channels under axial compression Steel Compos. Struct. 33 2019 37 66 10.12989/scs.2019.33.1.869
