Who Loves Virtue as much as He Loves Beauty?: Deep Learning based
Estimator for Aesthetics of Portraits
Tobias Gerlach
1,∗
, Michael Danner
1,2,∗
, Le Ping Peng
1,4
, Aidas Kaminickas
2
, Wu Fei
3
and Matthias R
¨
atsch
1
1
ViSiR, Reutlingen University, Reutlingen, Germany
2
Centre for Vision, Speech & Signal Processing, University of Surrey, Guildford, U.K.
3
School of Computer Science, Xi’an Polytechnic University, Xi’an 710048, China
4
Philosophy, Hunan University of Science and Technology, Xiangtan, China
∗
T. Gerlach and M. Danner contributed equally to this work as first authors
Keywords:
Benchmark Testing, Facial Databases, Attractiveness of Faces, Social Ethics, ELO Rating, Predictive Models,
Deep Learning, Extreme-Gradient-Boosting Regressor, 3D Morphable Model.
Abstract:
”I have never seen one who loves virtue as much as he loves beauty,” Confucius once said.
If beauty is more important as goodness, it becomes clear why people invest so much effort in their first im-
pression. The aesthetic of faces has many aspects and there is a strong correlation to all characteristics of
humans, like age and gender. Often, research on aesthetics by social and ethic scientists lacks sufficient la-
belled data and the support of machine vision tools. In this position paper we propose the Aesthetic-Faces
dataset, containing training data which is labelled by Chinese and German annotators. As a combination of
three image subsets, the AF-dataset consists of European, Asian and African people. The research communi-
ties in machine learning, aesthetics and social ethics can benefit from our dataset and our toolbox. The toolbox
provides many functions for machine learning with state-of-the-art CNNs and an Extreme-Gradient-Boosting
regressor, but also 3D Morphable Model technologies for face shape evaluation and we discuss how to train
an aesthetic estimator considering culture and ethics.
1 INTRODUCTION
1.1 Motivation
”In the post-Cold War world, the most important dis-
tinctions among peoples are not ideological, political,
or economic. They are cultural.” (Salzborn and Stich,
2016). Culture is the wealth created by mankind.
Since human beings exist and develop in regions,
countries, and nations, culture must present diversity,
and each culture is different.
From a philosophical point of view: Chinese aes-
thetics and Western aesthetics are the aesthetics of
two different ideological and cultural systems. The
traditional Chinese aesthetics represented by Confu-
cius (Wang, 2016) and Laozi (Kelly, 1998) are ex-
perience, ethical, and social philosophies. In con-
trast, Western aesthetics represented by Plato (Hy-
land, 2008) and Aristotle (Porter, 2017) are rational,
religious, and psychological philosophies.
The standard of aesthetics depends on different
cultures. For example, Book of Poetry, the oldest col-
lection of folk poems in China, was written from the
11
th
century BC to the 6
th
century BC, reflecting peo-
ple’s thoughts and feelings, life style and understand-
ing of natural phenomena in that era. It describes the
beauty of an image as: ”Her fingers like soft blades
of reed, like larva white her neck is slender” (Yuan-
chong, 1993). Although time is changing, Asians still
prefer to have white skin and a, bright-eyed look with
open eyes (Sturm et al., 2010).
An Asian with an umbrella is a common sight in
summer to provide protection from being sun-tanned.
”White skin” has emerged as a central desideratum of
consumer culture in affluent Asia (Sahay and Piran,
1997; Li et al., 2008), but from the study of West-
ern cultural history and film and television works, it
is found that American skin beauty advertisements
tend to use models with a tanned or darker skin tone
rather than a pale complexion as beauty (Xie and
Gerlach, T., Danner, M., Peng, L., Kaminickas, A., Fei, W. and Rätsch, M.
Who Loves Virtue as much as He Loves Beauty?: Deep Learning based Estimator for Aesthetics of Portraits.
DOI: 10.5220/0009172905210528
In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages
521-528
ISBN: 978-989-758-402-2; ISSN: 2184-4321
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
521

Zhang, 2013). Jones et al. (Jones et al., 2000) ex-
amined United Kingdom college students and found
that the majority reported engaging in and enjoying
sunbathing.
China has a proverb: ”Loving beauty is part of
human nature”. No matter where you’re from, no
matter what aesthetic standards you have, the yearn-
ing for beauty is the same because aesthetics has af-
fected many aspects of peoples’ lives. Various evi-
dences have shown that facial attractiveness plays a
key role in a variety of judgements in the course of
social interaction (Little and Roberts, 2012; van der
Geld et al., 2007; Langlois et al., 2000; Anderson
et al., 2008), attractive people are even more likely
to be hired (Aharon et al., 2001).
It should be noted that this study is not based
on a male-centred culture and uses male positions,
concepts, and perspectives to measure and assess the
value of women. In fact, we are gradually enrich-
ing the proportion of male photos in the photo library
which will extend our dataset and further study will
be published later. Because we label pairwise, for
annotators it is too hard to rank attractiveness in one
pair with a male and female face. The purpose of this
work is to compare current East and West objective
aesthetic standards to better understand the aesthetic
needs of different cultures and to provide an aesthetic
tool for cross-cultural communication, international
business, and other fields and avoid cultural discrimi-
nation or conflict.
1.2 Literature Research
(Xie et al., 2015) published a database, containing
500 images with aesthetics ratings: ”SCUT-FBP: A
Benchmark Dataset for Facial Beauty Perception”.
They not only built the database but also trained dif-
ferent models for assigning aesthetic values automat-
ically. Main take away of this work is a dataset for
Asian subjects and the indication, that Deep Learning
approaches give the best results for that kind of task.
We will not only improve on the choice of the net-
work but also improve the data preparation. SCUT-
FBP only used 400 images for training without any
augmentations. The network they used was a 6 lay-
ered Convolutional Neural Network.
(Eisenthal et al., 2006) also describe a machine
learning approach for beauty ratings which shows an
objective concept of beauty and its origins very well.
They also used a machine learning approach, but only
had 184 images for training, which is a very small
number in machine learning. They used support vec-
tor machines and Neural Networks, which rely on
features, selected by the programmer. We will use
Convolutional Neural Networks to select features au-
tomatically.
(Redi et al., 2015) focuses on the quality of digital
portraits. They only used manually extracted features
for regression upon which we will improve.
1.3 Main Contribution
Research into aesthetic estimation using the latest ma-
chine learning tools is a technological challenge that
psychologists and computer scientists have recently
addressed, and there is still a lot of potential for im-
provement. The results of this work should help sci-
entists in the future to combine aesthetic aspects with
artificial intelligence.
Aesthetic-Faces-dataset (AF-dataset): First of all,
we provide to the best of our knowledge the biggest
dataset with currently 12.684.492 aesthetic score an-
notations for the machine learning, aesthetics and so-
cial ethics research communities. This database is
under continuous development and will be published
on GitLab. Currently, the database consists of three
subsets of images: Olympic sports athletes, celebri-
ties and a dataset containing half Asian, half Cau-
casian faces (together 5.484 subject faces). We pro-
vide the aesthetic scores, but also several other meta-
data, like age, gender, eyeglass wearer, sport, ethnic
group and other characteristics of the subject faces
and also meta-data of the over 1,000 annotators with
different gender, age and cultural origin.
Aesthetics Estimation Toolbox: Additionally, we
develop and release a toolbox for researchers in so-
cial ethics to analyse or synthesise faces and prove
their assumptions in ethical, legal and social implica-
tions (ELSI). This paper includes a baseline approach
for an aesthetics score estimator. But the toolbox
is also interesting for the machine learning and face
modelling community. 3D Morphable Model tech-
nology is provided for deeper research and for re-
gression and we will use standard and state-of-the-
art approaches. Additionally, a comparison of ma-
chine learning tools like Convolutional Neural Net-
works and Extreme Gradient Boosting is projected.
Social Ethics Correlations: Finally, we describe
and analyse first correlations and social ethic aspects
within the AF-dataset based on the integration of ar-
tificial intelligence technologies into the practice of
social research.
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
522

2 METHODOLOGY
2.1 Machine Learning Approaches
We have been researching aesthetics estimation since
2014 and published some work about aesthetic eval-
uation of face images on CNN trained networks (Fei
et al., 2019). By changing the image background for
training, we could improve the overall performance
and show that the network benefits from different
backgrounds to prevent overfitting and increase accu-
racy. Further improvements have been achieved by
switching to state-of-the-art regression methods.
Further improvements on the architecture and
data preparation of the CNN approach are currently
in development and will be published. The LeNet
and F-Net approaches achieve first promising re-
sults but provide too few features and layers to
generate accurate results. Therefore, we will repeat
our experiments with state-of-the-art networks
from (Russakovsky et al., 2015), change the output
layers fitting for regression, and evaluate how they
perform.
We will also test extreme gradient boosting, since
many Kaggle competitions are won by applying the
XG-Boost regressor (Chen and Guestrin, 2016).
Kaggle is a commonly used homepage for providing
databases and data science competitions.
2.2 3D Morphable Models
Using 3D Morphable Models to analyse human face is
a popular approach to face analysis. Characteristic for
Morphable Models is to generate a 3D shape and its
colour separately. We will concatenate the AF-dataset
with 3D face shapes and analyse geometric propor-
tions of the mouth, nose, eyes and chin in 3D space
to aesthetic labels. The dataset will then also be clas-
sified with typical face types like oval, triangle, dia-
mond, rectangle, round, heart and oblong.
2.3 XG-Boost Regressor
We will also use ”Extreme Gradient Boosting”, an
MatLab implementation of Gradient Boosting (Fried-
man, 1999) algorithm. This algorithm was first pub-
lished by Friedmann, who might have the same influ-
ence on machine learning as Rosenblatt has with the
perceptron (Rosenblatt, 1958). Although this algo-
rithm is very old in the fast-evolving machine learn-
ing world, many competitive regression tasks are still
won with XG-Boost regressor.
3 AESTHETIC-FACES DATASET
(AF-DATASET)
Every AI’s solution to a problem begins with prepar-
ing a dataset. To investigate student surveys for rat-
ing the aesthetics of women’s faces it was necessary
to prepare datasets with several thousands of images.
Our research has already taken 6 years and in this
time, we have investigated 12 surveys with 4 differ-
ent datasets. In 2013 we started with the dataset that
we call the Olympics-subset. Since 2016 we have
been working on celebrity images and this year we
first started with SCUT-FBP pictures. None of the
images are normalized (except SCUT-FBP dataset),
which means we rate the beauty perception of the
image themselves, rather than the beauty of the fe-
males faces contained in the images. Below we will
briefly discuss all above-mentioned datasets, but we
will pay more attention to our recent surveys based
on celebrities and SCUT-FBP images. Dataset anal-
ysis was started exclusively for this proposal paper.
The resulting database will contain a mixture of all the
researched datasets with aesthetic correlated labels.
For the annotators, we used different datasets and
evaluation methods. Data labels were acquired by
annotating images from Celebrities- and Olympic-
subsets. The datasets were split into two halves which
had to compete against each other. The student has
to select either the left or the right image to find the
more aesthetic person (see figure 1). This has sev-
eral advantages over the direct rating of the images as
mostly described in (Xie et al., 2015).
Figure 1: One out of 1352 annotation screens to choose ei-
ther the left or right image on the Asian-Caucasian-subset.
The ratings of all students were saved to a text file
and then evaluated. Starting this year, we have been
using the ELO-rating (Lehmann and Wohlrabe, 2017)
to replace mean calculations.
Who Loves Virtue as much as He Loves Beauty?: Deep Learning based Estimator for Aesthetics of Portraits
523

3.1 Olympics-subset
The Olympics-subset (International Olympic Com-
mittee, 2013) consists of 1156 images of female
Olympic athletes. Every image is labelled with player
age, country of origin, height, weight, and sports
art. In 2014, 322 students participated in our survey.
Based on the survey output every image was rated
with a score value ranging from 0 to 1. In Figure 2
the average attractiveness is mapped on an Olympic
sport. According to this distribution, the most aes-
thetic female athletes are figure-skaters with an aver-
age score of 0.6 in opposite to speed-skating athletes
with a 0.45 ranking.
In Figure 3 we show how attractiveness depends
on the age of Olympics players. As expected, the
most attractive are females from 18 to 30 years old.
Figure 2: Aesthetic ranking distribution on Olympic sports.
Every dot equates to a subject’s image.
Figure 3: Attractiveness vs. age. Red line shows average
values, every blue dot equates to a subject’s image.
Various evidences have shown that faces that ap-
pear older are less attractive (Deffenbacher et al.,
1998; Henss, 2006). Cigarette ads with young per-
sons were found to appear more often in magazines
with younger audiences and for menthol brands. Re-
gardless of viewer age, younger models were judged
as more attractive than older models (Mazis et al.,
1992). From this study on large datasets we demon-
strate that if the age increases, the attractiveness curve
gradually declines; the correlation between age and
beauty is significant even in all subsets across very
different cultures of the subjects but also across dif-
ferent cultures of annotators (see Figure 10, Figure 11
and Figure 6). It should be noted that the attractive-
ness curve is relatively linear from 20 to 35 and on
some subsets even till 60 years old subjects. For low
and high ages, the data are not trustworthy, because
there are only a few examples.
3.2 Celebrities-subset
In 2016 we have started to create a new dataset by
creating a subset of Labelled Faces in the Wild, ab-
breviated LFW (Huang et al., 2007) that is a public
benchmark for face verification also known as pair
matching. The LFW dataset consists of 13,233 pic-
tures of 5,749 popular actors, singers or politicians.
These images were sorted out from male faces by our
researchers. Afterwards, image information like age,
ethnicity, and glasses (yes/no) was generated and the
new image set with 1,578 pictures is now called the
Celebrities-subset.
Figure 4 shows the distribution of age in this
dataset. We can see a Gaussian distribution with a
mean value of 43 years. The majority of the celebri-
ties didn’t wear glasses as seen in Figure 5. The same
figure also shows that around 1300 are white, 200
black and the rest Asian females faces. Our work
concatenates aesthetic labels with characteristic at-
tributes; for example Figure 6 shows the evaluation
of aesthetic labels and subject’s ages.
Figure 4: Distribution of age.
Currently, 171 students from different cul-
tures participated in annotating the Asian-Caucasian-
subset. The annotation outcome was calculated by
using both the direct ranking that counts the up- and
down-votes and the ELO-rating system for calculat-
ing relative numbers.
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
524

Figure 5: Glasses and ethnicity distribution on the
Celebrities-subset.
Figure 6: Correlation of age and attractiveness on the
Celebrities-subset.
Figure 7 shows a histogram about ethnicity and
attractiveness relation. The result here depends a lot
on survey participants that will be discussed in more
detail later on. It can also be concluded that people
without glasses are more attractive.
Figure 7: Celebrities-subset glasses and ethnicity vs. attrac-
tiveness.
3.3 Asian-Caucasian-subset
This year we started investigating surveys on a diverse
benchmark database for multi-paradigm facial beauty
prediction as released by the Human-Computer In-
telligent Interaction Lab of South China Univer-
sity of Technology and called SCUT-FBP5500-
database (Liang et al., 2018). The SCUT-FBP dataset
has 5500 frontal faces which are categorized in Asian
male, Asian female, Caucasian male and Caucasian
female. The dataset is annotated with facial land-
marks, beauty scores and a beauty score distribu-
tion that allow different computational models with
different facial beauty prediction paradigms, such as
appearance-based and shape-based facial beauty clas-
sification, regression and ranking.
This year we provided our beauty perception sur-
vey both in a European university as well as in a uni-
versity in Asia. Currently, we are collecting both the
annotator’s age as shown in Figure 8 and the gender
which is presented in Figure 9.
Figure 8: Annotators age distribution.
Figure 9: Annotators gender distribution.
To compare the results of surveys in two differ-
ent cultures we created rankings for the images sep-
arately. Our hypothesis was that the average Asian
attractiveness should be higher than Caucasian ac-
cording to the Chinese annotators’ results, which is
confirmed by Figure 10. Asian annotators think that
Asian females are more attractive than Caucasian.
To follow this logic, the European annotators result
should be the opposite. As we see in Figure 11, the
average attractiveness of Caucasian females is now
higher. And this is again proof that attractiveness de-
pends a lot on cultures.
Who Loves Virtue as much as He Loves Beauty?: Deep Learning based Estimator for Aesthetics of Portraits
525

Figure 10: Chinese annotations - attractiveness per age.
Figure 11: German annotations - attractiveness per age.
4 EXPERIMENTS AND SOCIAL
ETHICS ASPECTS
4.1 Female-Face-network (F-net)
To achieve initial results on aesthetic estimation, we
trained the labelled aesthetic dataset using our own
F-net network model for face image aesthetic clas-
sification. The network model structure is shown in
Figure 12.
The F-net network model consists of four convo-
lutional layers, four pooling layers, and two fully con-
nected layers. The model uses the ReLU activation
function to add local response normalised LRN op-
erations. The maximum pooling operation is used in
each pooling layer. At the same time, F-net uses the
same filling method to ensure that image edge infor-
mation is not lost. The model has the highest recog-
nition accuracy in the Female-Face-dataset, reaching
73%. F-net has a 7% improvement in recognition ac-
curacy compared to the LeNet-5 model, a 15% im-
provement in AlexNet recognition accuracy, and a
9% improvement in VGG-16 recognition accuracy.
The F-net network model is compared with the clas-
sical Convolutional Neural Network models LeNet-5,
AlexNet, and VGG-16 in the Female-Face-dataset.
Figure 12: Structure of Female-Face-network model.
4.2 Network Model Analysis
We trained the labelled aesthetic dataset on LeNet-
5, AlexNet and VGG-16. LeNet-5 consists of two
convolutional layers, two pooling layers and two fully
connected layers. We achieved a relatively high clas-
sification recognition accuracy of 66%. AlexNet’s
classification accuracy in the Female-Face-dataset is
only 58%. The accuracy of classification and recog-
nition of test sets is 62% for VGG-16.
4.3 Different Cultures Social Impact on
Aesthetic
How can a beautiful face be defined? Different cul-
tures have different answers (Larglois et al., 2000;
Rhodes et al., 2001). Machine learning is to a cer-
tain extent the embodiment of human psychological
behaviour and cultural differences lead to a differ-
ent standard of aesthetics, just as stated in Darwin’s
(1871) observation that ”It is certainly not true that
there is in the mind of man any universal standard of
beauty with respect to the human body” (Cunningham
et al., 1995)
This study reflects the different aesthetic standards
for female faces in Asia and Europe. It should be
noted that human-like biases observed in ML algo-
rithms have numerous harmful effects, and there ex-
ists a growing need to regulate and correct these bi-
ases (Fuchs, 2018). From racist Twitter bots (Munger,
2017) to unfortunate Google search results (Stephens-
Davidowitz, 2014), deep-learning software easily
picks up on biases. AI-driven decision-making can
lead to discrimination in several ways (Borgesius,
2018). Joanna Bryson, a computer scientist at the
University of Bath and a co-author of a human-liked
bias researching, said: ”A lot of people are saying
this is showing that AI is prejudiced. No. This
is showing we’re prejudiced and that AI is learning
it.” (Aylin Caliskan, 2017) In this study, we attach
great importance to avoid machine learning bias on
aesthetic judgment.
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
526

Some researches show that glasses cause a nega-
tive effect on ratings of social appearance. The stud-
ies from (Terry and Kroger, 1976) and (Thornton,
1944) showed that people with glasses were judged
to be less attractive than those without glasses, and a
statistically significant effect was found (p < 0.05).
Figure 7 demonstrates experiments on the AF-dateset
across different cultures, subjects and annotators. We
can show that humans with glasses are older on av-
erage and therefore the glass-wearer-ratings become
reasonable based on the demonstrated correlation of
aesthetics and age as shown in in Figure 10 and Fig-
ure 11.
5 CONCLUSION
Confucius has never seen one who loves virtue as
much as he loves beauty (Confucius, 1999). The
reader of the paper has seen that aesthetics is complex
correlated to culture. The Aesthetic-Faces-dataset can
be used to quantify this effect by classifying face im-
ages and prove hypothesis on a new level. We pro-
vide a huge database with many labels and a tool-
box with useful AI regression approaches for further
work and the research community in machine learn-
ing, aesthetics and even social ethics. We want to pro-
pose and use 3D Morphable Model face shape analyse
and synthesise technologies combined with research
with conventional and innovative Convolutional Re-
gression Networks and Extreme Gradient Boosting.
The next step is to publish our database on GitLab.
After that, the machine learning and face modelling
algorithms will be tested, evaluated and published as
tool box on GitLab too.
ACKNOWLEDGEMENTS
This work is partially supported by a grant of the
BMBF FHprofUnt program, no. 03FH049PX5
REFERENCES
Aharon, I., Etcoff, N., Ariely, D., Chabris, C., O’Connor, E.,
and Breiter, H. C. (2001). Beautiful Faces Have Vari-
able Reward Value: fMRI and Behavioral Evidence.
Neuron.
Anderson, S., Adams, G., and Plaut, V. (2008). The cultural
grounding of personal relationship: The importance of
attractiveness in everyday life. Journal of personality
and social psychology, 95:352–68.
Aylin Caliskan, Joanna J. Bryson, A. N. (2017). Semantics
derived automatically from language corpora contain
human-like biases. Science.
Borgesius, F. Z. (2018). Discrimination, artificial intelli-
gence, and algorithmic decision-making. Directorate
General of Democracy.
Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree
boosting system. CoRR, abs/1603.02754.
Confucius (1999). The analects of confucius: a philosoph-
ical translation. New York :Ballantine Books, page
9:18.
Cunningham, M., Roberts, A., Barbee, A., Perri, D., and
Wu, C.-H. (1995). Their Ideas of Beauty Are, On the
Whole, the Same as Ours. Journal of Personality and
Social Psychology.
Deffenbacher, K., Vetter, T., Johanson, J., and O’Toole, A.
(1998). Facial aging, attractiveness, and distinctive-
ness. Perception, 27:1233–43.
Eisenthal, Y., Dror, G., and Ruppin, E. (2006). Facial at-
tractiveness: Beauty and the machine. Neural compu-
tation, 18:119–42.
Fei, W., Xinjuan, Z., Xiaojun, W., and R
¨
atsch, M. (2019).
Aesthetic evaluation of face image based on convolu-
tional neural network.
Friedman, J. (1999). Greedy function approximation: A
gradient boosting machine. The Annals of Statistics,
29.
Fuchs, D. J. (2018). The Dangers of Human-Like Bias in
Machine-Learning Algorithms. Missouri S&T’s Peer
to Peer2.
Henss, R. (2006). Perceiving age and attractiveness in facial
photographs. Journal of Applied Social Psychology,
21:933 – 946.
Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller,
E. (2007). Labeled faces in the wild: A database for
studying face recognition in unconstrained environ-
ments. Technical Report 07-49, University of Mas-
sachusetts, Amherst.
Hyland, D. (2008). Plato and the Question of
Beauty. Studies in continental thought.
https://books.google.de/books?id=C1LaszjLNawC.
International Olympic Committee (Retrieved 03/08/2013).
Olympic Photos, Galleries and Slideshows,.
https://www.olympic.org/photos.
Jones, F., Harris, P., and Chrispin, C. (2000). Catching the
sun: An investigation of sun-exposure and skin pro-
tective behaviour. Psychology Health & Medicine -
PSYCHOL HEALTH MED, 5:131–141.
Kelly, M. (1998). Encyclopedia of Aesthetics. Oxford Uni-
versity Press.
Langlois, J., Kalakanis, L., Rubenstein, A., Larson, A., Hal-
lam, M., and Smoot, M. (2000). Maxims or myths of
beauty? a meta-analytic and theoretical review. Psy-
chological bulletin, 126:390–423.
Larglois, J., Kalakanis, L., Rubenstein, A., Larson, A., Hal-
lam, M., and Smoot, M. (2000). Maxims or myths of
beauty? A meta-analytic and theoretical review. Psy-
chol. Bull.
Who Loves Virtue as much as He Loves Beauty?: Deep Learning based Estimator for Aesthetics of Portraits
527

Lehmann, R. and Wohlrabe, K. (2017). Who is the ‘journal
grand master’? a new ranking based on the elo rating
system. Journal of Informetrics, 11(3):800 – 809.
Li, E. P., Min, H. J., Belk, R. W., Kimura, J., and Bahl,
S. (2008). Skin lightening and beauty in four asian
cultures. Advances in Consumer Research, 35:444–
449.
Liang, L., Lin, L., Jin, L., Xie, D., and Li, M. (2018).
Scut-fbp5500: A diverse benchmark dataset for multi-
paradigm facial beauty prediction.
Little, A. and Roberts, S. (2012). Evolution, appearance,
and occupational success. Evolutionary psychology :
an international journal of evolutionary approaches
to psychology and behavior, 10:782–801.
Mazis, M., Ringold, D., Perry, E., and Denman, D. (1992).
Perceived age and attractiveness of models in cigarette
advertisements. Journal of Marketing, 56.
Munger, K. P. B. (2017). Tweetment Effects on the
Tweeted: Experimentally Reducing Racist Harass-
ment. Springer.
Porter, J. I. (2017). Beauty, Value, and the Aesthetics of Life
in Kant and Aristotle. Republics of Letters, Volume 5,
Issue 1. University of California, Berkeley.
Redi, M., Rasiwasia, N., Aggarwal, G., and Jaimes, A.
(2015). The beauty of capturing faces: Rating the
quality of digital portraits.
Rhodes, G., Yoshikawa, S., Clark, A., Lee, K., McKay, R.,
and Akamatsu, S. (2001). Attractiveness of facial av-
erageness and symmetry in non-Western cultures: in
search of biologically based standards of beauty. Per-
ception.
Rosenblatt, F. (1958). The perceptron: A probabilistic
model for information storage and organization in the
brain. Psychological Review, pages 65–386.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,
Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-
stein, M. S., Berg, A. C., and Li, F. (2015). Imagenet
large scale visual recognition challenge. International
Journal of Computer Vision, 115(3):211–252.
Sahay, S. and Piran, N. (1997). Skin-color preferences
and body satisfaction among south asian-canadian
and european-canadian female university students.
The Journal of Social Psychology, 137(2):161–171.
PMID: 9140216.
Salzborn, S. and Stich, T. (2016). Samuel P. Huntington:
The Clash of Civilizations and the Remaking of World
Order, New York: Simon & Schuster 1996, 368 S. (dt.
Kampf der Kulturen. Die Neugestaltung der Weltpoli-
tik im 21. Jahrhundert, M
¨
unchen/Wien: Europa Ver-
lag 1996, 584 S.), pages 401–403. Springer Fachme-
dien Wiesbaden, Wiesbaden.
Stephens-Davidowitz, S. (2014). The cost of racial animus
on a black candidate: Evidence using Google search
data. Journal of Public Economics.
Sturm, A., Brissett, A., and Brissett, A. (2010). Ethnic
trends in facial plastic surgery. Facial plastic surgery
: FPS, 26:69–74.
Terry, R. L. and Kroger, D. L. (1976). Effects of eye cor-
rectives on ratings of attractiveness. Perceptual and
Motor Skills.
Thornton, G. R. (1944). The effect of wearing glasses upon
judgments of personality traits of persons seen briefly.
Journal of Applied Psychology.
van der Geld, P., Oosterveld, P., van Heck, G., and
Kuijpers-Jagtman, A. (2007). Smile attractiveness.
self-perception and influence on personality. Angle
Orthodontist, 77(5):759–765.
Wang, L. (2016). On confucius’s ideology of aesthetic or-
der.
Xie, D., Liang, L., Jin, L., Xu, J., and Li, M. (2015). SCUT-
FBP: A benchmark dataset for facial beauty percep-
tion. In SMC, pages 1821–1826. IEEE.
Xie, Q. V. and Zhang, M. (2013). White or tan? a cross-
cultural analysis of skin beauty advertisements be-
tween china and the united states. Asian Journal of
Communication, 23(5):538–554.
Yuanchong, X. (1993). Book of Poetry. Changsha: Hunan
Publishing House.
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
528