17K-Grafﬁti: Spatial and Crime Data Assessments in S

ao Paulo City

Bahram Lavi

, Eric K. Tokuda

, Felipe Moreno-Vera

, Luis Gustavo Nonato

, Claudio T. Silva

and Jorge Poco

School of Applied Mathematics, Getulio Vargas Foundation, Rio de Janeiro, Brazil

Institute of Mathematics and Statistics, University of S

ao Paulo, Brazil

Computer Science Department, San Pablo Catholic University, Arequipa, Peru

Institute of Mathematics and Computer Sciences, S

ao Carlos, Brazil

Computer Science and Engineering, New York University, New York, U.S.A.

Keywords:

Grafﬁti, Object Detection, Spatial Infrastructure, Crime Analysis, Data Correlation.

Abstract:

Grafﬁti is an inseparable element of most large cities. It is of critical value to recognize whether it is an artistry

product or a distortion sign. This study develops a larger grafﬁti dataset containing a variety of grafﬁti types

and annotated boundary boxes. We use this data to obtain a robust grafﬁti detection model. Compared with

existing methods on the task, the proposed model achieves superior results. As a case study, the created model

is evaluated on a vast number of street view images to localize grafﬁti incidence in the city of S

ao Paulo,

Brazil. We also validated our model using the case study data, and, again, the method achieved outstanding

performance. The robustness of the technique enabled further analysis of the geographical distribution of

grafﬁti. Considering grafﬁti as a spatial element of the city, we investigated its relation with crime occurrences.

Relatively high correlation values were obtained between grafﬁti and crimes against pedestrians. Finally, this

work raises many questions, such as the understanding of how these relationships change across the city

according to the types of grafﬁti.

1 INTRODUCTION

A man got impressed when scratches on a cave called

the attention for the ﬁrst time — known as “Grafﬁti”

these days. It is an inﬂuential social element that is

applied to manifest or express the culture of a com-

munity. Grafﬁti can be written and/or drawn with

spray paint, stickers, wheat paste and can be repre-

sented in different forms and types (e.g., tags, gang,

mural, etc.). Although grafﬁti can express culture,

manifesting the vision of a group of people, its im-

pact on the urban environment and the targeting of

neighborhoods is undeniable.

Through the perspective of the broken window

theory of Kelling and Wilson (Kelling et al., 1982) —

a professor at Harvard and a former policeman —

grafﬁti has been interpreted as an essential social dis-

ordering factor that can lead to inappropriate behav-

iors. For example, unauthorized grafﬁti can reject so-

cial bonds. This theory also plays a signiﬁcant role

in getting police attention to social factors and other

offenses; this idea has been broadly applied in the

United States criminal justice system (Jenkins, 2020).

A study in Groningen, Netherlands, showed that the

presence of tags-type grafﬁti more than double the

number of littering or stealing (Keizer et al., 2008).

Since late 1994, grafﬁti has been removed in Paris,

France, focusing mainly on recurrent visual signs of

disruption on the facades, reserving severe punish-

ment for any sort of vandalism in public places (De-

nis and Pontille, 2021). On the other hand, a recent

study showed that the authorized grafﬁti in Belo Hor-

izonte, Brazil, presented no relation to local crime oc-

currence (Diniz and Stafford, 2021).

It is well-known that there is a strong connection

between a city’s appearance and crime activity (Har-

court, 1998; Sampson and Raudenbush, 2001; Gomes

et al., 2021; Khorshidi et al., 2021). In fact, fac-

tors such as urban area conservation, health, educa-

tion, and mobility have high effects on the rates of

alcoholism, obesity, and the spread of STDs (Naik

et al., 2016). Therefore, grafﬁti is an essential spa-

tial phenomenon that must be carefully studied and

analyzed (Biljecki and Ito, 2021).

The contribution of this paper can be seen in three

fronts: (i) we introduce a collected, organized, and

968

Lavi, B., Tokuda, E., Moreno-Vera, F., Nonato, L., Silva, C. and Poco, J.

17K-Grafﬁti: Spatial and Crime Data Assessments in São Paulo City.

DOI: 10.5220/0010883300003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 4: VISAPP, pages

968-975

ISBN: 978-989-758-555-5; ISSN: 2184-4321

annotated grafﬁti dataset to train a robust model able

to identify grafﬁti from images; (ii) considering graf-

ﬁti as a spatial element, we run a case study to inves-

tigate the relation between crimes in the city of S

Paulo, Brazil (referred as S

ao Paulo from now on) and

grafﬁti; and ﬁnally (iii) we ﬁnd a substantial correla-

tion between crime events against pedestrians and the

presence of grafﬁti in S

ao Paulo.

2 RELATED WORK

We ﬁrst consider some works about urban percep-

tion and its relation with urban phenomena like crime.

Then we report the existing works and methods on

grafﬁti detection task.

2.1 Urban Perception Analysis

The research area on urban perception aims at under-

standing the city environment and the behavior of its

inhabitants (Yoshimura et al., 2020). Several studies

with different approaches (e.g., urban planning, ur-

ban designing, social development) have shown that

the city’s visual components inﬂuence human per-

ception and might impact the behavior of a popu-

lation. Lynch et al. (Lynch, 1984) report a simi-

lar metamorphosis in the main cities of the USA.

The work compares urban physical factors such as

shape, pattern, and texture of the buildings and urban

planning attributes (distribution of buildings, parks,

supermarkets, police stations, etc.), showing a rela-

tive relationship between the urban factors and the

behavior of inhabitants of speciﬁc neighborhoods.

Megler et al. (Megler et al., 2014) combine census

and city data in San Francisco, USA, to investigate

which urban factors signiﬁcantly correlate with graf-

ﬁti reports. Then, those factors are used to build a

regression model for predicting the grafﬁti incidence

in different city neighborhoods.

Schroeder and Anderson (Schroeder and Ander-

son, 1984) investigate the correlation of the city’s vi-

sual components (e.g., grafﬁti, garbage, trash) and

violations in Chicago, Georgia, and Michigan. The

work by (Arietta et al., 2014) studied the correla-

tion between visual and non-visual-attributes with the

statistic of crimes, house pricing, population den-

sity, grafﬁti presence, and a perception survey. Some

studies focus on extracting objects from street-view

images to understand the urban perception in some

populations (Ordonez and Berg, 2014; Naik et al.,

2014; Zhang et al., 2018; Moreno-Vera et al., 2021b;

Moreno-Vera et al., 2021a).

Table 1: Number of images and boundary box instances for

Train-and-Test sets. The table also reports the total value

over each column and row.

Boundary box

Set Images Single-boundary Multi-boundary Total

Train 6,956 4,115 9,704 13,819

Test 1,737 1,004 2,008 3,012

Total 8,693 5,119 11,712 16,831

2.2 Existing Grafﬁti Detection Methods

To the best of our knowledge, there is only a sin-

gle publicly available dataset for grafﬁti detection.

STORM (Charalampos et al., 2019) specializes in de-

tecting grafﬁti and was mainly collected in Greece. It

contains only instances with tags-type grafﬁti. The

dataset is made up of about 1K image samples ac-

quired at street level. Alzate et al. (Alzate et al.,

2021) later extended the STORM dataset by append-

ing three hundred images from Google Street View

(GSV) images, which were annotated for the graf-

ﬁti detection task. We compare our grafﬁti detec-

tor with those works in the experimental section.

Tokuda et al. (Tokuda et al., 2019) also performed

grafﬁti prediction on ground-level images, but on a

larger scale; however, the performance of the predic-

tion was modest (∼ 57% of average precision), not

providing any association with city indicators such as

crime.

3 GRAFFITI DETECTION

In the following section, we ﬁrst introduce our 17K-

Grafﬁti dataset and explain its annotation procedure,

aiming its use to train models to identify grafﬁti from

images. We also discuss the grafﬁti detection model

considered in this work, assessing its performance.

3.1 Dataset Collection and Annotation

Undoubtedly, Flickr.com is well-known as a rich re-

source of photo-sharing website. It provides an Appli-

cation Programming Interface (API) that allows users

to retrieve and download vast amounts of images sup-

plied by photographs (alongside the meta-data) for

further analyses.

Flickr.com also contains various types and a large

number of grafﬁti images. We retrieved the grafﬁti

images through the keyword “grafﬁti”. In our initial

stage, we recovered 15K photos of grafﬁti. We ex-

amined the initial pool and removed the duplicates,

which resulted in a ﬁnal dataset of about 9K im-

ages. The collection is rich and covers different types

17K-Grafﬁti: Spatial and Crime Data Assessments in São Paulo City

969

Figure 1: Histogram of grafﬁti instance distribution on our

collection set.

of grafﬁti (e.g., tags, mural, throw-ups, blockbuster,

stencil, etc.). Then we take another step to manually

annotate the boundary box over each grafﬁti instance

observed on each image in the dataset. To train a de-

tection model is essential to provide the boundary box

of the region of interest (ROI) — grafﬁti in this task.

It can offer single- or multi-instance boundary boxes

per image. Each ground truth boundary box repre-

sents two points in a 2D space: top-left coordinate

∗

) and (w

∗

) width and height values, respec-

tively. After the annotation process, the dataset con-

tains nearly 17K grafﬁti instances (some images in-

clude more than one grafﬁti). We point out that, since

the pictures have different sizes, we annotated them

with their original dimensions.

We considered a fraction of the dataset for train-

ing (80%), remaining for testing. Table 1 provides

detailed information on the number of grafﬁti images

and the number of annotated boxes. Note that most of

the photos from the dataset presented multi-instance

grafﬁti; in which some images contains more than ten

instances. Fig. 1 shows the histogram distribution of

grafﬁti instances.

3.2 Faster R-CNN

Faster Region-based Convolutional Neural Network

(Faster R-CNN) (Ren et al., 2016) is a seminal ob-

ject detection framework, mainly popular due to its

simplicity and robustness. It is a multi-stage ob-

ject detection model (Zhao et al., 2019) trained with

multi-task loss. Unlike a single-stage object detection

model, like YOLO (Redmon et al., 2016)), in which

the model aims to predict the probability of an ob-

ject’s presence through the conditional class proba-

bility and boundary box regressions for the object that

its center points match within a grid cell. Therefore,

most predicted anchors rely upon the background, and

dramatically only a few ground-truth instances re-

ceive positive prediction probability from the target

ground-truth within the grid cell. Thus, the network

ultimately has to make a trade-off to determine the

most potential candidates to deliver them as the ob-

jects presence, which yields the network to weak per-

Figure 2: Architecture of the Faster R-CNN utilized for our

grafﬁti detection task.

formance on some queries. In contrast, the Faster R-

CNN overcomes this issue by considering the number

of anchors appearing in the grid cell. The single-stage

detector can also be a prefect match for real-time de-

tection task. However, in this work, we aim to use

a robust detector in order to localize the incidence of

grafﬁti observed on street level for which the Faster

R-CNN showed its promises on accurate detection.

Given an image I and its corresponding boundary

boxes (x

∗

) as the network’s input, it ulti-

mately returns spatial coordinates of boundary boxes

with its associated classes by performing a boundary

box regression. The network contains an additional

region proposal network (RPN) that seeks the region

of interest (ROI), enabling the network to detect ob-

jects in a wide range of scales and aspect ratios. The

RPN network is trained by minimizing the objective

function as deﬁned in (Girshick, 2015). However, the

boundary box regression followed by the four coordi-

nates is computed as

= (x − x

)/w

= (y − y

)/h

= log(w/w

),t

= log(h/h

∗

= (x

∗

− x

)/w

∗

= (y

∗

− y

)/h

∗

= log(w

∗

),t

∗

= log(h

∗

(1)

where (x, y,w, h) and (x

) are the coordi-

nates of the predicted boundary box and the anchor

box, respectively. Faster RCNN begins with one of

the standard CNNs (e.g., ResNet (He et al., 2016),

VGG (Simonyan and Zisserman, 2014)) as its back-

bone, which is followed by a Feature Pyramid Net-

work (FPN). The FPN plays as a bottleneck to obtain

multiple feature maps with different scales from the

features generated by the backbone CNN. Both FPN

and RPN then pass through the pooling ROI layer.

Fig. 2 depicts the whole pipeline of Faster R-CNN

considered in this work as our grafﬁti detector.

3.3 Detection Assessment

Average precision (AP) is a typical metric for mea-

suring the accuracy of an object detector and esti-

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

970

mates the precision for each value of recall at different

ranks. To estimate if the detected box is a true/false

positive, it is essential to calculate the area of overlap

between ground truth and detected boxes, known as

Intersection Over Union (IOU).

AP calculates the area under the precision-recall

curve for each given image, the ground-truth, and

the detected boxes, AP =

p(r)dr where p(r) is the

measured precision at recall rank r. Finally, a mean

average precision (mAP),

∑

i=1

, is reported as

the ﬁnal score of the detector. The mAP is com-

monly reported over different IOU criteria. This eval-

uation metric is the same as in-applied on MS-COCO

dataset (Lin et al., 2014).

4 EXPERIMENTS

This section describes the experimental setup, the

grafﬁti detection results, and the study focused on

investigating the association between grafﬁti and

crimes.

4.1 Experimental Setup

The grafﬁti detection model is built upon the Py-

torch framework (Paszke et al., 2019). We adopted

ResNet50 (He et al., 2016) as the backbone for the

Faster RCNN. Then, we re-scaled each input im-

age to 3 × 224 × 224, representing the image’s num-

ber of channels, width, and height, respectively. We

used a learning rate of 0.001 and ﬁne-tuned the net-

work using the pre-trained weights on the COCO

dataset (Lin et al., 2014). We then iterated the net-

work to learn 27K times. For each iteration, the net-

work was ﬁne-tuned with a randomly-sample batch

of 16 images. We use the momentum of 0.9 and

a weight decay of 0.0005 (Ren et al., 2016) while

training. We performed all the experiments in an

Nvidia Quadro RTX5000 GPU. The organized 17K-

Grafﬁti dataset, boundary box annotations, model im-

plementation, and pre-trained weights are available in

https://github.com/visual-ds/17K-Grafﬁti.

4.2 Grafﬁti Detection Results

We report the performance of our grafﬁti detector us-

ing mAP (%). We evaluate the model also on the

STORM dataset (Charalampos et al., 2019), compar-

ing it with the work in (Alzate et al., 2021). Table 2

reports the performance of our grafﬁti detector as

mAP over different criteria on IOU. As expected, the

stricter the IOU criterion (higher value), the smaller

the mAP observed. The table also reports the com-

parison with a few works in the literature. The re-

sults from previous works have been reported just for

IOU of 0.5, in which case our proposed method pre-

sented the best performance, with 85% of mAP. Other

works (Tokuda et al., 2019) have also reported detec-

tion performance values, but they considered different

datasets and thus, their results have not been included

in this comparison.

4.3 Study on Spatial and Crime Data

The ultimate goal of the study presented in this sec-

tion is to determine whether Grafﬁti correlates with

any sort of crime. In other words, the goal is to an-

alyze if Grafﬁti impacts the occurrence of crimes in

speciﬁc neighborhoods. We evaluate this hypothesis

on a large number of images collected from Google

Street View (GSV), along with criminal records. We

consider S

ao Paulo as the case study of this experi-

ment.

Google Street View Images

Ground-level images provide a valuable resource for

exploring how features vary across regions, such as

the amount of green and buildings (Li et al., 2015;

Torii et al., 2009). In particular, GSV is a service

that provides ground-level images for public access

and with comprehensive spatial coverage. For these

reasons, we have considered their images in our case

study.

GSV maps cover most of the big cities in the

world. In particular, the city of S

ao Paulo is satisfac-

torily mapped by GSV images. This, coupled with the

notorious widespread of grafﬁti (Iddings et al., 2011)

in S

ao Paulo, makes it an appropriate choice for our

study.

We uniformly sampled over the entire city, in a

grid-like fashion spaced by 102 meters. For each

point, four complementary images have been consid-

ered to cover the full 360

◦

view. In the end, after re-

moving corrupt and third-party-provided images, we

ended up with 275,339 images from the regions of in-

terest.

We evaluated our grafﬁti detection model in each

image. To safely assert the incidence of grafﬁti, we

established a hard prediction threshold of 94%, i.e.,

predictions with conﬁdence below this value were not

considered. In the end, we obtained 4,475 individ-

ual instances of grafﬁti across 4,268 affected images

(1.6% of the total number of images). Although this

corresponds to a small percentage, the density of graf-

ﬁti is considerably heterogeneous across the city, and

17K-Grafﬁti: Spatial and Crime Data Assessments in São Paulo City

971

Table 2: Mean average precision on different IOU criteria (in-percentage) over the detected boundaries on the test set of the

STORM dataset.

mAP

Detector dataset @[IOU=0.25] @[IOU=0.50] @[IOU=0.75]

(Alzate et al., 2021)

STORM - 58.30 -

STORM-Extended - 69.14 -

Ours

STORM 83.05 71.60 51.53

17K-Grafﬁti 89.13 85.20 62.64

Figure 3: (Top): geographical distributions of downloaded

GSV images, and detected Grafﬁti; (bottom): crime against

vehicle, and crime against pedestrian over 96 districts of

ao Paulo. We report the absolute numbers only for the case

of GSV images, while the remaining show the distribution

over the normalized values as described in Sec. 4.3.

there are neighborhoods with a higher incidence of

grafﬁti (Fig. 3).

Besides the validation results described in the pre-

vious section, to assess the performance of the trained

model in our case-study images, we addition-ally an-

notated 853 images. We evaluated the same detection

model to predict grafﬁti incidences. A remarkably

high precision value of 94% was achieved, attesting

that most predicted occurrences are grafﬁti.

In Fig. 4, we show some of the grafﬁti detected by

the method. The two ﬁrst rows correspond to correct

predictions and the bottom row to erroneous ones. We

notice that our technique can cope with different graf-

ﬁti styles, view perspectives, and image illuminations

from the correct prediction.

Crime Records

The police department of S

ao Paulo provides crime

records regarding robbery, leaving out drug-related

felony and sexual assault. Each record contains the

date, time, geographic coordinates where the offense

took place, and the type of crime. The dataset com-

prises crime records from 2006 to 2017, and the rob-

bery records are split into three types: pedestrian,

commercial, and vehicle robberies. In this study,

we consider only two types of crime, vehicle (32K

records) and pedestrian (104K records), from the year

2017 — this is a fair choice since our GSV images

were collected in 2017 as well.

Correlation of Grafﬁti with Crime

Since the urban factors are independent, it is neces-

sary to normalize data appropriately to catch a ratio-

nal measurement with the data correlation. We treat

the normalization task for each factor independently.

The grafﬁti data is normalized by dividing it over the

number of GSV images from the same district. On the

other hand, the population is a promising normaliza-

tion factor for crime occurrence. By taking that into

account, we normalized the crime data of each district

by dividing over the district’s population.

We consider the Pearson correlation to analyze the

relationship between Grafﬁti (X) and each crime type

(Y ). The correlation shows that the quantiﬁed param-

eter X

is correlated with parameter Y

when both in-

dependent parameters are potentially affected by each

other (in this study, for i = 1,...,96 representing each

district of S

ao Paulo). Therefore an increase in one

should be associated with a rise in the other.

Fig. 3 (top-left) shows the geographical distribu-

tion of the images considered in this work aggre-

gated by each of the 96 districts of the city. As one

can observe, the districts have been unevenly sam-

pled, partly due to the different sizes of the neigh-

borhoods and partly due to the rural characteriza-

tion and consequently less coverage by the image

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

972

Figure 4: Two ﬁrst rows: examples of true positive samples, last row: examples of false positive samples. Original images

from Google Street View. For a privacy concern, we discard to report the geo location of the images.

Figure 5: Scatter plot and its ﬁtted regression over the data

points of Grafﬁti and two-type of crime records. Each point

in the plot represents the data record per district in S

Paulo.

provider in certain regions. Correspondingly, the top-

right map shows the geographical distribution of the

incidence of grafﬁti. A one-to-one correspondence

with the top-left map is not observed, which shows

the heterogeneity of the incidence of grafﬁti. Re-

gions with just a few images, such as in the cen-

ter of the top-left map, present a considerable inci-

dence of grafﬁti (top-right map). Conversely, regions

with a large number of images do not present much

grafﬁti. The criminality indices are shown in the

bottom row in Fig. 3, namely vehicle (bottom-right)

and pedestrian(bottom-left) robbery. Comparing both

crime maps, we see a clear difference in the distribu-

tion of the two types of crimes.

Fig. 5 shows scatter plots of the normalized fre-

quency of grafﬁti versus the normalized frequency of

crimes for every 96 districts. We can see the number

of crimes against vehicles on the left, and on the right,

the number of crimes against pedestrians. Whereas

we can not observe a correlation between grafﬁti and

vehicles’ robbery (6% in Table 3), we can observe

Table 3: This table reports the r value of Pearson correlation

on Grafﬁti against two-type of crime in S

ao Paulo.

Spatial infrastructure Crime r value

Grafﬁti vs.

Vehicle 0.06

Pedestrian 0.44

a 44% (Table 3) between grafﬁti and pedestrian rob-

bery. It is an exciting result because one could expect

that the eventual relationship between crime and graf-

ﬁti would be agnostic to the crime type. However, this

result corroborates with Fig. 3, which visually shows

an agreement between pedestrian robbery and graf-

ﬁti. We hypothesize that such a weak relationship be-

tween the distribution of crimes against vehicles and

the presence of grafﬁti is related to the fact that cars

are less vulnerable to the prevalent crimes in regions

with an abundant presence of grafﬁti. The same fac-

tors that favor grafﬁti, for example, the lack of illumi-

nation and low rates of police patrolling, might prefer

the occurrences of crimes against pedestrians.

5 LIMITATION AND FUTURE

WORK

A limitation arose that the developed detector was

treated as a binary classiﬁcation problem – aiming to

detect any grafﬁti incidence for a given image. It thus

limited us to study and analyze the effect of differ-

ent grafﬁti types and their relationship with crimes.

On the other hand, the inﬂuence of other urban indi-

cators (such as incidence of street light) along with

crime data records concerning over different time-

17K-Grafﬁti: Spatial and Crime Data Assessments in São Paulo City

973

period could assist us in enriching a potential discus-

sion overall in this study.

As future work, we recognize two discrete direc-

tion to extend the work on the hand. First, since our

grafﬁti data collection contains variant grafﬁti types,

one can strengthen the current detector into a multi-

class grafﬁti detection, enabling it to detect different

kinds of grafﬁti instances. It will allow us to exam-

ine more in-depth the presence of variant grafﬁti types

and crimes. Secondly, we investigate other spatial in-

frastructure data (e.g., population, health rate, educa-

tion) to broadly analyze each factor and its possible

relation with crimes.

6 CONCLUSIONS

This paper presents a 17K-Grafﬁti dataset specialized

for the task of grafﬁti detection. The dataset com-

prises a rich pool of grafﬁti instances that were ade-

quate to train a robust object detection model, namely

Faster R-CNN. We compared the obtained detection

model with existing grafﬁti detection approaches and

obtained a signiﬁcant gain in performance. The model

was also evaluated on many GSV images of S

ao Paulo

and an exceptional performance was observed. We

manually assessed a sample of the predictions and

identiﬁed possible causes for the false predictions.

Finally, we considered grafﬁti as spatial infras-

tructure data and analyzed its effect with the criminal

records data provided by the Police department of S

Paulo. In particular, two types of offenses have been

considered: against vehicles and pedestrians. While

the results revealed no apparent association with the

former, a relatively high correlation across neighbor-

hoods was observed for the latter. We hypothesized

the causes of such effects, mainly related to the fac-

tors that favor grafﬁti production.

ACKNOWLEDGEMENTS

This work was supported by Getulio Vargas Foun-

dation; NSF awards CNS-1229185, CCF-1533564,

CNS-1544753, CNS-1730396, CNS-1828576,

CNS-1626098; FAPESP grants 2019/01077-3,

2015/22308-2, and 2013/07375-0; CNPq-Brazil

grants #303552/2017-4 and #312483/2018-0; Cien-

ciActiva of the National Council for Science, Tech-

nology and Technological Innovation (CONCYTEC-

PERU) with grant 234-2015-FONDECYT; and

CAPES. Any opinions, ﬁndings, conclusions, or

recommendations expressed in this material are those

of the authors and do not necessarily reﬂect the views

of the CAPES, CNPq, FAPESP, or FGV.

REFERENCES

Alzate, J. R., Tabares, M. S., and Vallejo, P. (2021). Graf-

ﬁti and government in smart cities: a deep learning

approach applied to medell

ın city, colombia. In Inter-

national Conference on Data Science, E-learning and

Information Systems 2021, pages 160–165.

Arietta, S. M., Efros, A. A., Ramamoorthi, R., and

Agrawala, M. (2014). City forensics: Using visual

elements to predict non-visual city attributes. IEEE

transactions on visualization and computer graphics,

20(12):2624–2633.

Biljecki, F. and Ito, K. (2021). Street view imagery in urban

analytics and gis: A review. Landscape and Urban

Planning, 215:104217.

Charalampos, P., Panagiotis, K., Lazaros, T., and Alexan-

dros, T. (2019). Storm grafﬁti/tagging detection

dataset.

Denis, J. and Pontille, D. (2021). Maintenance epistemol-

ogy and public order: Removing grafﬁti in paris. So-

cial Studies of Science, 51(2):233–258.

Diniz, A. M. A. and Stafford, M. C. (2021). Graf-

ﬁti and crime in belo horizonte, brazil: The broken

promises of broken windows theory. Applied Geogra-

phy, 131:102459.

Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE

international conference on computer vision, pages

1440–1448.

Gomes, H. S., Farrington, D. P., Defoe, I. N., and Maia,

(2021). Field experiments on dishonesty and stealing:

what have we learned in the last 40 years? Journal of

Experimental Criminology, pages 1–31.

Harcourt, B. E. (1998). Reﬂecting on the subject: A critique

of the social inﬂuence conception of deterrence, the

broken windows theory, and order-maintenance polic-

ing new york style. Mich. L. Rev., 97:291.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Iddings, A. C. D., McCafferty, S. G., and Da Silva, M. L. T.

(2011). Conscientizac¸

ao through grafﬁti literacies in

the streets of a s

ao paulo neighborhood: An ecosocial

semiotic perspective. Reading Research Quarterly,

46(1):5–21.

Jenkins, M. J. (2020). Fixing broken neighborhoods: How

police can ensure neighborhood safety and commu-

nity well-being. Advances in Social Science, Educa-

tion and Humanities Research, 420:7–13.

Keizer, K., Lindenberg, S., and Steg, L. (2008). The spread-

ing of disorder. Science, 322(5908):1681–1685.

Kelling, G. L., Wilson, J. Q., et al. (1982). Broken windows.

Atlantic monthly, 249(3):29–38.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

974

Khorshidi, S., Carter, J., Mohler, G., and Tita, G. (2021).

Explaining crime diversity with google street view.

Journal of Quantitative Criminology, 37(2):361–391.

Li, X., Zhang, C., Li, W., Ricard, R., Meng, Q., and Zhang,

W. (2015). Assessing street-level urban greenery us-

ing google street view and a modiﬁed green view in-

dex. Urban Forestry & Urban Greening, 14(3):675–

685.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Euro-

pean conference on computer vision, pages 740–755.

Springer.

Lynch, K. (1984). Reconsidering the image of the city. In

Cities of the Mind, pages 151–161. Springer.

Megler, V., Banis, D., and Chang, H. (2014). Spatial anal-

ysis of grafﬁti in san francisco. Applied Geography,

54:63–73.

Moreno-Vera, F., Lavi, B., and Poco, J. (2021a). Quantify-

ing urban safety perception on street view images. In

2021 IEEE/WIC/ACM International Joint Conference

on Web Intelligence and Intelligent Agent Technology

(WI-IAT). IEEE.

Moreno-Vera, F., Lavi, B., and Poco, J. (2021b). Urban per-

ception: Can we understand why a street is safe? In

Mexican International Conference on Artiﬁcial Intel-

ligence, pages 277–288. Springer.

Naik, N., Philipoom, J., Raskar, R., and Hidalgo, C. (2014).

StreetScore: predicting the perceived safety of one

million streetscapes. 2014 IEEE Conference on Com-

puter Vision and Pattern Recognition Workshops.

Naik, N., Raskar, R., and Hidalgo, C. A. (2016). Cities are

physical too: Using computer vision to measure the

quality and impact of urban appearance. American

Economic Review, 106(5):128–32.

Ordonez, V. and Berg, T. L. (2014). Learning high-level

judgments of urban perception. European Conference

on Computer Vision (ECCV).

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., et al. (2019). Pytorch: An imperative style,

high-performance deep learning library. Advances

in neural information processing systems, 32:8026–

8037.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time object

detection. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 779–

788.

Ren, S., He, K., Girshick, R., and Sun, J. (2016). Faster

r-cnn: towards real-time object detection with region

proposal networks. IEEE transactions on pattern

analysis and machine intelligence, 39(6):1137–1149.

Sampson, R. J. and Raudenbush, S. W. (2001). Disorder in

urban neighborhoods: Does it lead to crime. US De-

partment of Justice, Ofﬁce of Justice Programs, Na-

tional Institute of . . . .

Schroeder, H. W. and Anderson, L. M. (1984). Perception

of personal safety in urban recreation sites. Journal of

leisure research, 16(2):178–194.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Tokuda, E. K., Cesar, R. M., and Silva, C. T. (2019). Quan-

tifying the presence of grafﬁti in urban environments.

In 2019 IEEE International Conference on Big Data

and Smart Computing (BigComp), pages 1–4. IEEE.

Torii, A., Havlena, M., and Pajdla, T. (2009). From google

street view to 3d city models. In IEEE 12th Inter-

national Conference on Computer Vision workshops,

pages 2188–2195. IEEE.

Yoshimura, Y., He, S., Hack, G., Nagakura, T., and Ratti, C.

(2020). Quantifying memories: Mapping urban per-

ception. Mobile Networks and Applications, 25:1275–

1286.

Zhang, F., Zhou, B., Liu, L., Liu, Y., Fung, H. H., Lin, H.,

and Ratti, C. (2018). Measuring human perceptions

of a large-scale urban region using machine learning.

Landscape and Urban Planning, 180:148–160.

Zhao, Z.-Q., Zheng, P., Xu, S.-t., and Wu, X. (2019). Ob-

ject detection with deep learning: A review. IEEE

transactions on neural networks and learning systems,

30(11):3212–3232.

17K-Grafﬁti: Spatial and Crime Data Assessments in São Paulo City

975