THEME ARTICLE: Visualization for Smart City Applications
Mapping and Visualizing
Deep-Learning Urban
Beautification
Information visualization has great potential to make
sense of the increasing amount of data generated by
complex machine-learning algorithms. We design a
set of visualizations for a new deep-learning algorithm
called FaceLift (goodcitylife.org/facelift). This
algorithm is able to generate a beautified version of a
given urban image (such as from Google Street
View), and our visualizations compare pairs of
original and beautified images. With those visualizations, we aim at helping practitioners
understand what happened during the algorithmic beautification without requiring them
to be machine-learning experts. We evaluate the effectiveness of our visualizations to
do just that with a survey among practitioners. From the survey results, we derive
general design guidelines on how information visualization makes complex machine-
learning algorithms more understandable to a general audience.
Beautiful places make us feel better. Stendhal’s motto “beauty is the promise of happiness”
speaks to this and has been made use of in various studies to show that specific visual cues affect
our well-being.
1,2
But what are these visual cues of beauty? Our public realm is filled with exam-
ples people perceive as beautiful. Regardless of whether beauty emerges from planning or seren-
dipitously, we can identify its cues, and that is useful for supporting evidence-based design of
the urban spaces we intuitively love.
Based on that premise, here we present a design study that visualizes data-intensive and complex
results stemming from a range of deep neural networks. These generative adversarial networks
(GANs) beautify a Google Street View scene according to a trained concept of beauty.
3
To vali-
date the beautification process, we compare the original image and the beautified one in terms of
the elements that have been added or removed. These elements are then mapped into urban de-
sign metrics that the urban design literature has identified to characterize great urban spaces.
4
In
Tobias Kauer
University of Applied
Science Potsdam
Sagar Joglekar
King’s College London
Miriam Redi, Luca Maria
Aiello, and Daniele Quercia
Nokia Bell Labs Cambridge
70
IEEE Computer Graphics and Applications Published by the IEEE Computer Society
0272-1716/18/$33.00 ©2018 IEEE September/October 2018