Analysis of the new architectural dataset NeoFaçade and its potential in machine learning

Bianka Kowalska, Hubert Baran, Daniil Hardzetski, Halina Kwaśnicka, Aleksandra Marcinów, Małgorzata Biegańska

doi:10.37190/arc240408

Summary

The article presents an analysis of the NeoFaçade dataset and explores its applications in machine learning for architectural studies. The structure and content of the dataset and potential applications in automated architectural analysis are discussed. Special attention is paid to the possibilities of using machine learning in historical architecture research.

✨ Summary generated by Artificial Intelligence. The actual Abstract can be found in the article below.

Full article view is only available on bigger screens.

2024

4(80)

Bianka Kowalska*, Hubert Baran**, Daniil Hardzetski***,

Halina Kwaśnicka****, Aleksandra Marcinów*****,

Małgorzata Biegańska******

Analysis of the new architectural dataset NeoFaçade

and its potential in machine learning

DOI: 10.37190/arc240408

Published in open access. CC BY NC ND license

Abstract

The presence of articial intelligence (AI) in architecture has been growing rapidly in recent years. The collaboration between architects and

AI developers has led to signicant improvements in various design applications. Further development of machine learning techniques is highly

dependent on the availability of large, structured datasets. The aim of the article is to demonstrate the potential of a novel dataset, NeoFaçade, which

contains annotated pictures of historical tenements. A comparison of the dataset with existing benchmark datasets, the CMP Façade and the Paris Art-

Deco datasets, highlights its exceptional features. Its applications in three machine learning tasks are also presented: semantic segmentation, image

translation and image generation. In all three tasks, the models trained with NeoFaçade provide satisfactory results and indicate the great potential

of this collection. The planned further development of the dataset will allow the training of more precise models that will be able to distinguish more

elements and features of the façades and assist architects in designing tenements.

Key words: dataset, image processing, machine learning, architecture

Introduction

Architecture, as one of the oldest professions, has

evolved over centuries, and with it the tools of architec-

tural design. In recent years, articial intelligence (AI) has

emerged in the eld of architecture and has revolutionized

design optimization, empowering architects and designers.

Traditionally, architecture is associated with creativity,

aesthetics, and spatial design. However, in today’s world,

the scientic and technological aspects of design are also

playing an increasingly important role. This is where AI

comes in, oering new tools and opportunities for archi-

tects. For instance, AI was implemented in the design of

the Bo-DAA apartment project in Seoul, South Korea

(Rhee, Chung 2019).

Articial intelligence, especially Machine Learning (ML)

algorithms, neural networks, and vision systems, can be

used in various aspects of architecture. When applying AI

techniques, one should consider the ethical implications

of AI in architecture (Liang et al. 2024). The true power

of AI in architecture lies in its collaboration with human

architects (Bölek, Tutal, and Özbaşaran 2023). Combining

the knowledge of architects with AI techno logies opens the

way to innovative solutions and projects. It can improve

architecture’s eciency, functionality, and sustainability,

* ORCID: 0009-0008-6289-0762. Faculty of Information and

Communication Technology, Wrocław University of Science and Tech-

nology, Poland.

** ORCID: 0009-0001-2976-4923. Faculty of Information and

Communication Technology, Wrocław University of Science and Tech-

nology, Poland.

*** ORCID:0009-0001-2815-8948. Faculty of Information and

Communication Technology, Wrocław University of Science and Tech-

nology, Poland.

**** ORCID: 0000-0002-4368-4110. Faculty of Information and

Communication Technology, Wrocław University of Science and Tech-

nology, Poland, e-mail: halina.kwasnicka@pwr.edu.pl

***** ORCID: 0000-0003-4409-3563. Faculty of Architecture, Wro -

cław University of Science and Technology, Poland.

****** ORCID: 0009-0005-4302-2618. Faculty of Architecture, Wro-

cław University of Science and Technology, Poland.

76 Bianka Kowalska, Hubert Baran, Daniil Hardzetski, Halina Kwaśnicka, Aleksandra Marcinów, Małgorzata Biegańska

ensuring that architects feel valued and integral to the pro-

cess (Nabizadeh, Nabizadeh 2023).

Articial intelligence is a practical tool that can be used

in project management and construction, helping plan and

monitor work progress, optimize costs, and forecast pos-

sible problems. AI systems can analyze data related to the

operation of buildings, predict their energy consumption,

and respond to changes in environmental conditions in

real time, giving architects a sense of condence and reas-

surance in their work. Using AI techniques, designers can

generate and analyze thousands of design concepts, identi-

fy optimal solutions, or respond quickly to changing needs

and requirements (Sourek 2024).

The presence of articial intelligence in architecture has

been increasing since 2012 and remains a topic of growing

interest. Publications cover various applications, from per-

formance-based studies to spatial programming and resto-

ration work (Bölek, Tutal, and Özbaşaran 2023). In recent

years, generative articial intelligence has made a notable

surge within the realm of architecture, benetting from the

rapid growth of deep models, such as generative adversar-

ial networks (GANs) (Li et al. 2024). GANs consist of two

networks: a generator, trained to produce outputs that can-

not be distinguished from real images, and a discriminator,

which is trained to detect the generator’s fakes. Advanced

generative models are capable of generating multiple de-

sign options in a short time, allowing architects to explore

dierent design concepts quickly. AI takes over repetitive

tasks, allowing architects to focus on the design’s creative

and strategic aspects.

One of the challenging design tasks is designing histor-

ical tenement façades, since the façade must be placed in

the tight proximity of other buildings as well as meet the

historical and cultural demands of the region. AI has also

been employed in this context, with applications including

the analysis of historical building data and the generation

of design recommendations based on popular architectur-

al styles from a specic era (Enjellina, Beyan and Rossy

2023).

This paper demonstrates the potential of the authors’

dataset, NeoFaçade, using some example applications.

The usefulness of this dataset is compared to two acces-

sible benchmark datasets, but from dierent cities and ar-

chitectural styles. NeoFaçade was created by a team from

Wrocław University of Science and Technology. It con-

tains façades of Wrocław’s buildings, mainly from the 19

century. The intention is to use this dataset and others to

renovate existing tenement façades or to construct inlls

between old buildings. In the next stage, this dataset will

be expanded with photos of tenement façades from Ber-

lin and Szczecin, since the tenements in these cities are of

a similar style. More information about the current version

of NeoFaçade can be found in (Marcinów et al. 2024). It is

worth mentioning that we plan to make this dataset avail-

able for research.

The paper is structured as follows: Section 2 briey

reviews the related work on the topic. The third section

describes the datasets used in experiments, and the fourth

section presents the results of various experiments. The

summary concludes the paper.

Related work

In recent years, there has been a notable rise in the quan-

tity of studies on architectural design that makes use of

generative AI. Our perception of visual culture becomes

more mechanic with time, which inuences architecture.

This problem is considered in (Yau et al. 2023), where the

authors debunk the three ways of thinking about architec-

tural form in order to “lay the foundational preconditions

for a machine-learnable architecture.” Most of the work

focuses on architectural plan design, which includes cre-

ating horizontal section views at specic site elevations

guided by objective conditions and subjective decisions.

Many authors point out that more research is needed in

the

elds of structural system design, architectural 3D form

re-

nement and optimization design, and architectural façade

design (Sourek 2024; Ploennigs, Berger 2023; Nabiza-

deh, Nabizadeh 2023; Li et al. 2024; Bölek, Tutal, and

Özbaşaran 2023).

The applications of generative AI in structural system

design primarily involve the prediction of structural lay-

out and structural dimensions (Pizarro et al. 2021). In the

Bo-DAA apartment project described above, the architects

utilized a 3D model and architectural plan to dene the

building’s spatial form and structural load distribution

(Rhee, Chung 2019).

Façade design aims to create the building’s external ap-

pearance and includes all structural design demands and

encapsulates the areas and positions of façade elements

while adopting a specic style. Applications of generative

AI in architectural façade design encompass two primary

categories: the generation of façade images, typically em-

ploying semantic segmentation maps of the façades, and

the generation of semantic segmentation maps based on

2D images (Xie et al. 2021). Segmentation is a popular

problem in computer vision. In semantic segmentation, we

want to train a model that will assign a label to each pix-

el for a given image; this is in contrast to classication,

where we assign only one label to a whole image (Peng

et al. 2020). The generation approach utilizing semantic

segmentation consequently involves generating images

under the geometrical constraints of the given area. It can

be posed as translating an input image into a corresponding

output image. In 2018, Pix2Pix was introduced; this im-

age-to-image translation method is capable of translating

one possible representation of a scene into another using

deep learning, given sucient training data (Isola et al.

2017). The framework employs GANs in a conditional

setting, meaning that the generator is tasked with not only

fooling the discriminator, but also being near the ground

truth output (Newton 2019; Gui et al. 2021). The authors

explored the generality of conditional GANs by testing the

method on various tasks, for example, map to aerial photo,

day to night, black-and-white to color, and architectural la-

bels to façade photo.

Following the successful deployment of the Transform-

ers architecture in natural language processing tasks, there

has been a growing interest in adapting Transformers for

computer vision. The rst model was a vision Transformer

(ViT) created for classication tasks. ViT architecture was

Analysis of the new architectural dataset NeoFaçade and its potential in machine learning 77

later used in segmentation transformer (SETR) architec-

ture to study the performance of Transformers in semantic

segmentation tasks. SETR achieved promising results, and

further work began to create a better architecture utilizing

Transformers without SETR’s limitations. One of those

architectures is SegFormer (Xie et al. 2021). Its main dif-

ference is that it utilizes Mix Transformers (MiT), which

return multi-scale features, while ViT returns only single-

scale features. SegFormer shows good accuracy and run

time on benchmark datasets.

Although models based on neural networks are undeni-

ably the most popular articial intelligence models these

days, other approaches also have applications in various

architecture-related tasks. In the eld of façade image

generation, stochastic split grammars are used to build fa-

çade structure-aware generative models (Riemenschneider

et al. 2012; Martinovic, Van Gool 2013a; Gadde, Marlet

and Paragios 2016). A split grammar generates a façade

image by splitting a given input shape into a set of smaller

shapes, where each shape represents some part of a façade

image. For instance, if the input shape is a rectangle repre-

senting a building, the split grammar might divide it into

smaller rectangles representing the stories of the building.

This process continues until the desired level of detail is-

reached and the nal shapes are obtained (e.g., doors or

windows), which are called terminals. In such a grammar,

one may “merge” two shapes to be used interchangeably

(with some probabilities) and obtain a stochastic gram-

mar that can generate new façade examples. The ap-

proach may be extended to generating semantic segmenta-

tion masks.

Data used in experiments

Machine learning models require large amounts of high-

quality training data to learn complex patterns and re la-

tionships within them eectively. The scarcity of struc tured

architectural training data presents a signicant challenge,

undermining the initial stages of model training. There-

fore, the NeoFaçade dataset obtained by ( Marcinów et al.

2024) is a crucial asset for developing com prehensive, pre-

cise models.

NeoFaçade represents a novel approach to addressing

the growing demand for structured, high-quality data in the

eld of architecture. The dataset contains 400 annotated

images of tenement façades in Wrocław. The images rep-

resent a variety of architectural styles from the 19

and

centuries. Each image is assigned a color-coded anno-

tation mask, where the colors represent the various façade

elements (e.g., windows, cornices, or details). A detailed

description of the dataset and its creation is available in

(Marcinów et al. 2024).

As the ecacy of AI models is highly dependent on

the quality of the data used for training, it is essential to

conduct a comparative analysis to compare NeoFaçade

with existing benchmark datasets. Two available datasets

of similar content and structure are the Center for Ma-

chine Perception (CMP) Façade Dataset (Tyleček, Šára

2013) and the Paris ArtDeco Dataset (Gadde, Marlet, and

Paragios 2016). These datasets were selected because of

their shared theme and labels, with the images depicting

residential buildings and sharing elements of interest, such

as balconies or details. The former consists of 378 recti-

ed and cropped façade images from various locations.

The pictures were manually annotated by the authors with

a set of overlapping rectangular shapes, each coded to

represent one of the 12 classes that were distinguished in

the collection. The latter contains 79 façade images from

Paris following the Art Deco style. Each picture is asso-

ciated with an annotation text le containing a matrix of

numbers in the image shape. Each number represents one

of the seven classes featured in the collection. Figure 1 il-

lustrates example pictures and their annotations from both

datasets.

Compared to the other two datasets, NeoFaçade is dis-

tinguished by its larger set of images and their higher res-

olution, as shown in Table 1. The number of basic façade

elements distinguished in the set is the same as in the CMP

dataset, while Paris ArtDeco is characterized by the lowest

number of classes.

Fig. 1. Example pictures and their annotations from the CMP dataset (left) and Paris ArtDeco (right)

(elaborated by B. Kowalska)

Il. 1. Przykładowe zdjęcia i ich adnotacje ze zbioru danych CMP (po lewej) i Paris ArtDeco (po prawej)

(oprac. B. Kowalska)

78 Bianka Kowalska, Hubert Baran, Daniil Hardzetski, Halina Kwaśnicka, Aleksandra Marcinów, Małgorzata Biegańska

The Paris ArtDeco dataset exhibits the most balanced

distribution of classes, with all distinguished elements oc-

curring in almost all images. This is due to its elements be-

ing more common and the dataset not focusing on specic

architectural details. In contrast, the other datasets display

less of a balance among classes. The CMP dataset and

NeoFaçade exhibit a comparable imbalance ratio (the ratio

of majority class to minority class).

The exceptionally high resolution of images within Neo-

Façade allows us to generate images of exceptional qual-

ity and to conduct a detailed analysis of tenement façades

from the 19

and 20

centuries. The dataset comparison

provides a comprehensive analysis of multiple datasets

within the eld of architecture, revealing dierences in

terms of size, scope, and quality. The analysis conrms the

value and potential impact of NeoFaçade for driving fu-

ture research advancements. The usefulness of the dataset

was briey studied; the experiments are presented in the

next section.

Verification of the dataset’s potential

in machine learning tasks

In the case of façade generation, the main tasks are seg-

mentation and image translation. In this section, we present

the results of three ML models used for these tasks, which

were trained and evaluated using the data described above.

Semantic segmentation

One of the popular tasks in computer vision systems is

semantic segmentation. This process assigns a label to ev-

ery pixel, providing a detailed understanding of the image’s

contents. Several techniques have been developed for this

task; they involve classifying each pixel in an image into

a predened category. Semantic segmentation is a power-

ful technique for detailed image analysis, enabling precise

understanding and categorization of every part of an image.

We trained the SegFormer model using a pre-existing

implementation (Yin et al. 2023) with the three datasets

described in Section 3 in order to evaluate how well-suited

the datasets are for semantic segmentation. The objective

of this experiment was to:

– compare the performance of the model trained on

NeoFaçade against those trained on existing benchmark

datasets,

– identify the common problems encountered in façade

semantic segmentation, and

– identify the specic issues associated with training the

model on NeoFaçade.

Figures 2 and 3 illustrate the segmentation results of

models trained and evaluated with the three aforemen-

tioned

atasets. Upon initial observation, the model de mon -

strates an ability to identify the majority of classes. The an-

notations

produced by the model are placed correctly but

do not maintain the original rectangular shapes.

Table 2 presents the metrics of the SegFormer trained

on the three datasets. Notably, the results of the model

trained on NeoFaçade fall between the other two models.

This can be attributed to the characteristics of the data. The

Paris ArtDeco dataset, for instance, features fewer classes

and clear façade images. In contrast, the CMP dataset of-

ten includes foreign objects in the pictures, such as trees

that heavily cover the façades. We can use some metrics to

evaluate the performance of the trained model. The popu-

lar ones are as follows:

– Precision – the ratio of correctly predicted positive

observations to the total predicted positive observations.

– Recall – the ratio of correctly predicted positive ob-

servations to all observations in a given class.

– F1 measure – a harmonic mean of recall and precision

which evenly takes into account precision and recall and is

more suitable for comparing models.

Four measures obtained on three datasets are shown

in Table 2. MacroAvg (macro-average F1 measure) is an

F1 measure that is calculated separately for each class,

then averaged using the arithmetic mean. It is unsuitable

for imbalanced data because it prefers dominant classes.

Weighted Avg F1 (weighted average F1 measure) is an F1

measure calculated for each class separately and averaged

with weights depending on the number of true labels of

each class. The weight of class x depends on the propor-

tion of data belonging to that class within the dataset.

This measure considers the minor classes and is better

for imbalanced data. The situation is similar with preci-

sion measurement: Macro Avg Precision and Weighted Avg

Precision.

The number of classes in NeoFaçade is comparable

to that of the CMP dataset. However, unlike in CMP, the

data is less obscured by foreign objects. Consequently, the

weighted average of metrics is very similar for the Neo-

Façade and Paris ArtDeco datasets.

To gain further insight into the model’s performance,

it is necessary to analyze the confusion matrix (Fig. 4).

A confusion matrix is a matrix that provides an overview

of the performance of a machine learning model on a set of

test data. It is a tool for visualizing and analyzing the accu-

racy of a model’s predictions and is frequently employed

to assess the eectiveness of models designed to assign

a categorical label to each data input. The matrix compares

predicted labels to true labels, which indicates the exact

mistakes the model made. Instances where the prediction

was accurate are represented on the diagonal of the matrix.

Dataset NeoFaçade CMP

Paris

ArtDeco

Number of

pictures

400 378 79

Number

of classes

12 12 7

City Wrocław various Paris

Mean resolution

[MPx]

11.71 0.65 0.26

Imbalance ratio 2.62 2.61 1.11

Table 1. Comparison of façade datasets (elaborated by B. Kowalska)

Tabela 1. Porównanie zawartości zbiorów danych (oprac. B. Kowalska)

Analysis of the new architectural dataset NeoFaçade and its potential in machine learning 79

Fig. 2. Ground truth annotations and semantic segmentation results on the CMP dataset (left) and Paris ArtDeco (right)

(elaborated by D. Hardzetski)

Il. 2. Oryginalna anotacja i wyniki segmentacji semantycznej na zbiorze CMP (po lewej) i zbiorze Paris ArtDeco (po prawej)

(oprac. D. Hardzetski)

Fig. 3. Ground truth annotation and semantic segmentation results on the NeoFaçade dataset (elaborated by D. Hardzetski)

Il. 3. Oryginalna anotacja i wyniki segmentacji semantycznej na zbiorze NeoFaçade (oprac. D. Hardzetski)

Table 2. SegFormer – performance measures for three datasets (elaborated by D. Hardzetski)

Tabela 2. SegFormer – miary skuteczności dla trzech zbiorów danych (oprac. D. Hardzetski)

Model Macro Avg F1 Weighted Avg F1 Macro Avg Precision Weighted Avg Precision

CMP 0.42 0.60 0.51 0.62

Paris ArtDeco 0.61 0.75 0.68 0.76

NeoFaçade 0.50 0.73 0.54 0.75

80 Bianka Kowalska, Hubert Baran, Daniil Hardzetski, Halina Kwaśnicka, Aleksandra Marcinów, Małgorzata Biegańska

The most prevalent error in the model trained on Neo-

Façade was various classes being mislabeled as a wall, al-

though this issue was also present in the two other models.

This is attributed to the fact that the real labels are not pre-

cise at pixel level and the model has an error in determin-

ing the shape of an object. The model does not correctly

determine the contour of an element, but it still nds it in

the correct spot.

A further limitation of the model is its tendency to mis-

label other classes as the detail class. This occurs when the

model mistakenly identies other wall elements, such as

window frames and cornices, as details due to their visual

similarities. Balconies and roofs may be the most surpris-

ing of these mislabeled classes. This can be attributed to the

model’s inability to accurately comprehend these objects’

volume. In images, these elements appear at and have a dif-

ferent texture than the wall, which may lead to mislabeling.

The most challenging objects for our model are ver-

tical elements. The model exhibited a complete inability

to identify this label. Based on the examples containing

this class, we may assume the reason for such poor per-

formance is that the elements in question closely resemble

walls. The main distinct feature of vertical elements is their

pillar shape, which our model failed to learn.

Image translation

Image-to-image translation is the process-s of trans-

forming an input image into an output image while pre-

serving some semantic properties. It encompasses a range

of tasks, including image synthesis, resolution enhance-

ment or image colorization. In the case of the NeoFaçade

dataset, image translation can be utilized to transform se-

mantic segmentation masks into real façade images.

Fig. 4. Confusion matrix – performance of the model on NeoFaçade (elaborated by D. Hardzetski)

Il. 4. Macierz pomyłek – działanie modelu na zbiorze NeoFaçade (oprac. D. Hardzetski)

Analysis of the new architectural dataset NeoFaçade and its potential in machine learning 81

In the experiment, the ready-made implementation of

the Pix2Pix model was used (Isola et al. 2017). The model

consists of two neural networks: a generator, whose task is

to generate new images, and a discriminator, which classi-

es whether the generated image is real or fake. The gen-

erator learns the underlying connections between images

from the two domains (segmentation masks and façade

images) and the discriminator examines each fragment of

the generated picture and attempts to classify the patch as

authentic or articial. The discriminator’s examination of

individual small parts of the image facilitates sharp and

detailed results with the Pix2Pix model.

The model was trained for 250 epochs on two datasets

separately: NeoFaçade and CMP. Figures 5 and 6 illu strate

the example results of the Pix2Pix model on the translation

task (segmentation mask to façade image).

In both cases, the model was able to correctly generate

desired façade elements in the relevant locations. However,

more detailed objects, such as transoms and balconies, ap-

pear blurred in the generated images. This is because they

occur less frequently in the datasets. The model encounters

diculties lling in the blank gaps left in the images due

to image warping.

Façade generation with grammars

We applied generative grammar for façade generation, in-

spired by the approach from (Martinovic, Van Gool 2013a).

Several models were trained on small subsets of the data-

set. First, a rectangular lattice was generated for each

façade example, using pixel labels from the segmentation

mask. Lattices were built similarly to the one proposed

by (Riemenschneider et al. 2012), with the additional re-

moval of redundant split lines. Next, each façade lattice

was converted into a hierarchical structure called a parse

tree, which broke down a façade into oors and then into

smaller parts.

Then, parse trees were converted into grammars and

merged into one grammar, containing shapes from all fa -

çades in the training set. The merged grammar underwent

symbol merging. The best symbols to merge were chosen

by parsing candidate grammars from merging with 2D

Segmentation mask Original image Predicted image

Fig. 5. Results of the image2image translation task on NeoFaçade images (elaborated by B. Kowalska)

Il. 5. Wynik zadania translacji image2image na obrazach ze zbioru NeoFaçade (oprac. B. Kowalska)

Fig. 6. Results of the image2image translation task on CMP dataset (elaborated by B. Kowalska)

Il. 6. Wynik zadania translacji image2image na obrazach ze zbioru CMP (oprac. B. Kowalska)

82 Bianka Kowalska, Hubert Baran, Daniil Hardzetski, Halina Kwaśnicka, Aleksandra Marcinów, Małgorzata Biegańska

Earley Parser (Martinovic, Van Gool 2013b) and calculat-

ing their log-likelihood.

In Figure 7, the left column presents examples of gen-

erated façade images. The right column shows examples

from the training set that are the most similar to the gener-

ated ones. The model basically generates input examples

with some small modications (e.g., a window is replaced

with another window, possibly from another input façade

image). More training steps could make induced grammars

introduce more modications to the input images.

Summary

The primary aim of this study was to present and ana-

lyze the innovative architectural dataset NeoFaçade, high-

lighting its potential applications in machine learning. We

demonstrated its quality and versatility by comparing this

collection with two benchmark datasets of similar structure.

Our analysis included the evaluation of three dierent

machine learning models: semantic segmentation, image

translation, and image generation. Each model was tested

using the dataset and the results underscored the dataset’s

ability to produce satisfactory outcomes across these var-

ied tasks. These ndings suggest signicant potential for

further renement and optimization in future studies.

The continuous expansion of the dataset, incorporating

additional photographic material from diverse urban loca-

tions, is anticipated to enhance its applicability and per-

formance. This ongoing enrichment is expected to yield

increasingly favorable results, paving the way for devel-

Fig. 7. Façade generation

with grammars using NeoFaçade:

generated façades on the left

and similar real photos

in the data set on the right

(elaborated by H. Baran)

Il. 7. Generowanie fasad

z zastosowaniem gramatyki

dla zbioru NeoFaçade:

po lewej wygenerowane fasady,

po prawej – najbardziej podobne

zdjęcia ze zbioru danych

(oprac. H. Baran)

Analysis of the new architectural dataset NeoFaçade and its potential in machine learning 83

oping an architect-friendly generative model capable of

producing highly detailed and contextually accurate archi-

tectural designs.

Future research will focus on leveraging the detailed

metadata included in the dataset. This metadata encom-

passes basic elements of façades and distinguishes between

various architectural styles and elements. Such compre-

hensive annotations oer a rich source of information that

can be employed to train more precise and more sophisti-

cated models. These models will be capable of generating

façades that adhere to rigorous architectural and spatial

specications, thereby meeting the high standards required

in professional architectural design.

Moreover, the planned inclusion of tenements from

Berlin and Szczecin will further enhance the dataset’s di-

versity and robustness. This expansion is expected to sig-

nicantly improve the performance of the presented mod-

els, leading to better and more accurate results. The dataset

will provide a more comprehensive foundation for training

advanced machine learning models by incorporating these

additional urban landscapes.

The NeoFaçade dataset stands out as a high-quality re-

source for machine learning applications in architecture.

Its rich and detailed annotations, combined with contin-

uous updates and expansions, position it as a valuable

tool for developing innovative solutions in architectural

design. The ndings of this study highlight the dataset’s

potential to support the creation of advanced, generative

models that align with the precise demands of architec-

tural practice.

References

Bölek, Buse, Osman Tutal, and Hakan Özbaşaran. “A systematic review

on articial intelligence applications in architecture.” Journal of

Design for Resilience in Architecture and Planning 4, no 1 (2023):

91–104. https://doi.org/10.47818/DRArch.2023.v4i1085.

Enjellina, Eleonora Vilgia Putri Beyan, and Anastasya Gisela Cinintya

Rossy. “Review of AI Image Generator: Inuences, challenges, and

future prospects for architectural eld.” Journal of Articial Intelli-

gence in Architecture (JARINA) 2, no. 1 (2023): 53–65. https://doi.

org/10.24002/jarina.v2i1.6662.

Gadde, Raghudeep, Renaud Marlet, and Nikos Paragios. “Learning gram -

mars

for architecture-specic facade parsing.” International Jour-

nal of Computer Vision 117, no. 3 (2016): 290–316. https://doi.

org/10.1007/s11263-016-0887-4.

Gui, Yingbin, Biao Zhou, Xiongyao Xie, Wensheng Li, and Xifang Zhou.

“GAN-Based Method for generative design of visual comfort in un-

derground space.” IOP Conference Series: Earth and Environmen-

tal Science 861, no. 7 (2021): 072015. https://doi.org/10.1088/1755-

1315/861/7/072015.

Isola, Phillip, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. “Image-

to-Image Translation with conditional adversarial networks.” 2017

IEEE Conference on Computer Vision and Pattern Recognition (CVPR),

(2017): 5967–76. https://doi.org/10.1109/CVPR.2017.632.

Li, Chengyuan, Tianyu Zhang, Xusheng Du, Ye Zhang, and Haoran Xie.

“Generative AI models for dierent steps in architectural design:

A literature review.” Frontiers of Architectural Research (2024).

https://doi.org/10.1016/j.foar.2024.10.001.

Liang, Ci-Jyun, Thai-Hoa Le, Youngjib Ham, Bharadwaj R.K. Mantha,

Marvin H. Cheng, and Jacob J. Lin. “Ethics of articial intelligence

and robotics in the architecture, engineering, and construction in-

dustry.” Automation in Construction 162 (June 2024): 105369.

https://doi.org/10.1016/j.autcon.2024.105369.

Marcinów, Aleksandra, Małgorzata Biegańska, Bianka Kowalska, Hu-

bert Baran, Daniil Hardzetski, and Halina Kwaśnicka. “Building

a dataset of Wrocław’s historic tenements: Image annotation for ma-

chine learning applications.” Architectus 79, no. 3 (2024): 55–64.

https://doi.org/10.37190/arc240306.

Martinovic, Andelo, and Luc Van Gool. “Bayesian grammar learning

for inverse procedural modeling.” 2013 IEEE Conference on Com-

puter Vision and Pattern Recognition (2013a): 201–8. https://doi.

org/10.1109/CVPR.2013.33.

Martinovic, Andelo, and Luc Van Gool. “Early Parsing for 2D Stochastic

Context Free Grammars.” Technical Report KUL/ESAT/PSI/1301,

KU Leuven, 2013b.

Nabizadeh Rafsanjani, Hamed, and Amir Hossein Nabizadeh. “Towards

human-centered articial intelligence (AI) in architecture, engineer-

ing, and construction (AEC) industry.” Computers in Human Be-

havior Reports 11 (August 2023): 100319. https://doi.org/10.1016/j.

chbr.2023.100319.

Newton, David. “Generative deep learning in architectural design.”

Tech nology|Architecture + Design 3, no. 2 (2019): 176–89. https://

doi.org/10.1080/24751448.2019.1640536.

Pizarro, Pablo N., Leonardo M. Massone, Fabián R. Rojas, and Rafael

O. Ruiz. “Use of convolutional networks in the conceptual structural

design of shear wall buildings layout.” Engineering Structures 239

(2021): 112311. https://doi.org/10.1016/j.engstruct.2021.112311.

Peng, Jizong, Guillermo Estrada, Marco Pedersoli, and Christian Des-

rosiers. “Deep co-training for semi-supervised image segmenta-

tion.” Pattern Recognition 107 (November 2020): 107269. https://

doi.org/10.1016/j.patcog.2020.107269.

Ploennigs, Joern, and Markus Berger. “AI art in architecture.” AI in

Civil Engineering 2, 8 (2023). https://doi.org/10.1007/s43503-023-

00018-y.

Rhee, Jinmo, and Jae-Won Chung. “Applicability of Articial Intelligence

in Apartment Complex Design.” In Annual Conference in Architec-

tural Institute of Korea, 2019.

Riemenschneider, Hayko, Urlich Krispel, Wolfgang Thaller, Michael Do-

noser, Sven Havemann, Dieter Fellner, and Horst Bischof. “Irre gu-

lar lattices for complex shape grammar facade parsing.” 2012 IEEE

Conference on Computer Vision and Pattern Recognition (2012):

1640–47. https://doi.org/10.1109/CVPR.2012.6247857.

Sourek, Michal. “AI in architecture and engineering from misconcep tions

to game-changing prospects.” Architectural Intelligence 3, 4 (2024).

https://doi.org/10.1007/s44223-023-00046-9.

Tyleček, Radim, and Radim Šára. “Spatial pattern templates for recog-

nition of objects with regular structure.” In Pattern Recognition.

GCPR 2013. Lecture Notes in Computer Science, edited by Joachim

Acknowledgements

This research was funded by the NCN Miniatura 7 Grant, number 2023/

07/X/ST8/01424.

We would like to express our gratitude to all those who have contribu-

ted to this project: scholars from the chair of the History of Architecture,

Art and Technology of the Faculty of Architecture of Wrocław University

of Science and Technology (Aleksandra Brzozowska-Jawornicka, PhD

Arch, Bartłomiej Ćmielewski, PhD, Maria Legut-Pintal, PhD and Ro-

land Mruczek, PhD) and the students of the Faculty of Architecture, who

took the photographs of the tenements and carried out the annotation

(with particular thanks to Katarzyna Blicharz and Agnieszka Pałka).

84 Bianka Kowalska, Hubert Baran, Daniil Hardzetski, Halina Kwaśnicka, Aleksandra Marcinów, Małgorzata Biegańska

Streszczenie

Analiza nowego zbioru danych architektonicznych NeoFaçade oraz jego potencjału w uczeniu maszynowym

Rola sztucznej inteligencji (AI) w architekturze gwałtownie wzrosła w ciągu ostatnich lat. Współpraca między architektami i programistami

AI doprowadziła do usprawnień w wielu dziedzinach projektowych. Dalszy rozwój technik maszynowego uczenia w znacznym stopniu zależy od

dostępności dużych i ustrukturyzowanych zbiorów danych. Celem autorów artykułu jest pokazanie potencjału nowego zbioru danych, nazwanego

NeoFaçade, zawierającego opisane (anotowane) obrazy kamienic historycznych. Porównując zbiór z innymi ogólnodostępnymi zbiorami – CMP

Facade oraz Paris ArtDeco – podkreślono jego potencjalną użyteczność. Zaprezentowane również zostało wykorzystanie zbioru w trzech zadaniach

uczenia maszynowego: segmentacji semantycznej, translacji obrazów z generacji obrazów. We wszystkich trzech zadaniach modele wytrenowane na

zbiorze NeoFaçade dają satysfakcjonujące wyniki i wskazują na wysoki potencjał zbioru. Planowany dalszy rozwój zbioru umożliwi wytrenowanie

dokładniejszych modeli, które będą w stanie rozróżniać więcej elementów i cech fasad oraz wspomagać architektów w projektowaniu kamienic.

Słowa kluczowe: zbiór danych, przetwarzanie obrazów, uczenie maszynowe, architektura

Weickert, Matthias Hein, Bernt Schiele, Springer, 2013. https://doi.

org/10.1007/978-3-642-40602-7_39.

Xie, Enze, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Al-

varez, and Ping Luo. “SegFormer: Simple and ecient design for

semantic segmentation with transformers.” (2021). https://doi.org/

10.48550/arXiv.2105.15203.

Yau, Ho Man, Theodoros Dounas, Wassim Jabi, and Davide Lombardi.

“Timber joints analysis and design using shape and graph grammar-

based machine learning approach.” In

Digital Design Reconsidered

– Pro ceedings of the 41

Conference on Education and Research in

Com puter Aided Architectural Design in Europe (eCAADe 2023),

20–23 September 2023, Graz, Austria, edited by Wolfgang Dokonal,

Urs Hirschberg, Gabriel Wurzer. Vol. 1. eCAADe, 2023. https://doi.

org/ 10.52842/conf.ecaade.2023.1.569.

Yin, Wei, Yifan Liu, Chunhua Shen, and Baichuan Sun. “SSIW.” Ac-

cessed 2023. https://github.com/Xpitre/SSIW/.