In e na ional Jou nal o
Geo-In o ma ion
A icle
Using Objec De ec ion on Social Media Images o U ban
Bicycle In as uc u e Planning: A Case S udy o D esden
Ma in Knu a 1,*,† , Flo ian Kluge 2,† , Mo is Zah ila 3,† , Jochen Schiewe 1, Bodo Rosenhahn 2
and Di k Bu gha d 3


Ci a ion: Knu a, M.; Kluge , F.;
Zah ila, M.; Schiewe, J.; Rosenhahn,
B.; Bu gha d , D. Using Objec
De ec ion on Social Media Images o
U ban Bicycle In as uc u e
Planning: A Case S udy o D esden.
ISPRS In . J. Geo-In . 2021,10, 733.
h ps://doi.o g/10.3390/ijgi10110733
Academic Edi o : Jean-Claude Thill,
Ran Tao, Zhaoya Gong and Wol gang
Kainz
Recei ed: 17 Sep embe 2021
Accep ed: 23 Oc obe 2021
Published: 28 Oc obe 2021
Publishe ’s No e: MDPI s ays neu al
wi h ega d o ju isdic ional claims in
published maps and ins i u ional a il-
ia ions.
Copy igh : © 2021 by he au ho s.
Licensee MDPI, Basel, Swi ze land.
This a icle is an open access a icle
dis ibu ed unde he e ms and
condi ions o he C ea i e Commons
A ibu ion (CC BY) license (h ps://
c ea i ecommons.o g/licenses/by/
4.0/).
1Lab o Geoin o ma ics and Geo isualiza ion (g2lab), Ha enCi y Uni e si y Hambu g,
Henning-Vosche au-Pla z 1, 20457 Hambu g, Ge many; jochen.schiewe@hcu-hambu g.de
2Ins i u e o In o ma ion P ocessing, Leibniz Uni e si y Hanno e , Appels . 9a, 30167 Hanno e , Ge many;
[email p o ec ed].de (F.K.); [email p o ec ed].de (B.R.)
3Ins i u e o Ca og aphy, D esden Uni e si y o Technology, Helmhol zs . 10, 01062 D esden, Ge many;
[email p o ec ed] (M.Z.); di k.bu gha [email p o ec ed] (D.B.)
*Co espondence: ma in.knu a@hcu-hambu g.de
† These au ho s con ibu ed equally o his wo k.
Abs ac :
Wi h ci ies ein o cing g eene ways o u ban mobili y, encou aging u ban cycling helps
o educe he numbe o mo o ized ehicles on he s ee s. Howe e , ha also leads o a signi ican
inc ease in he numbe o bicycles in u ban a eas, making he ques ion o planning he cycling
in as uc u e an impo an opic. In his pape , we in oduce a new me hod o analyzing he
demand o bicycle pa king acili ies in u ban a eas based on objec de ec ion o social media images.
We use a subse o he YFCC100m da ase , a collec ion o pos s om he social media pla o m Flick ,
and u ilize a s a e-o - he-a objec de ec ion algo i hm o de ec and classi y mo ing and pa ked
bicycles in he ci y o D esden, Ge many. We we e able o e ie e he as majo i y o bicycles while
gene a ing ew alse posi i es and classi y hem as ei he mo ing o s a iona y. We hen conduc ed
a case s udy in which we compa e a eas wi h a high densi y o pa ked bicycles wi h he numbe
o cu en ly a ailable pa king spo s in he same a eas and iden i y po en ial loca ions whe e new
bicycle pa king acili ies can be in oduced. Wi h he esul s o he case s udy, we show ha ou
app oach is a use ul addi ional da a sou ce o u ban bicycle in as uc u e planning because i
p o ides in o ma ion ha is o he wise ha d o ob ain.
Keywo ds:
objec de ec ion; social media; u ban planning; bicycle in as uc u e; compu e ision;
olun ee ed geog aphical in o ma ion; isual analy ics
1. In oduc ion
Today, as ci ies g ow and de elop a high-speed a es, many o hem ac i ely ein o ce
g eene ways o u ban mobili y o igh agains pollu ion, a ic jams, noise, e c. [
1
]. One
o he encou aged ways o u ban commu ing is cycling o i s bene icial e ec on bo h he
en i onmen and pe sonal heal h [
2
]. While his s a egy helps o educe he numbe o
mo o ized ehicles on s ee s, i also leads o a signi ican inc ease in he numbe o bicycles
in u ban a eas, which aises he impo ance o bicycle in as uc u e planning as a opic.
Ci ies usually pay a en ion o p o iding a la ge numbe o bicycle pa king a popula
loca ions such as ain s a ions, shopping malls and highly equen ed squa es; howe e ,
nume ous bicycles a e o en andomly pa ked in he su ounding a eas whe e he e a e
ewe bicycle acks a ailable [
3
]. Acco dingly, planning he cycling in as uc u e is an
impo an opic o bo h u ban planne s and cyclis s.
T adi ional me hods o collec ing ield da a in u ban planning a e spo obse a ions
and su eys. Spo obse a ions a e usually conduc ed by coun ing objec s (e.g., passenge s,
bicycles, e c.) a u ban loca ions o in e es . This me hod is esou ce-consuming ega ding
ime and s a , so he collec ed da a usually do no co e longe ime in e als. In con as ,
ISPRS In . J. Geo-In . 2021,10, 733. h ps://doi.o g/10.3390/ijgi10110733 h ps://www.mdpi.com/jou nal/ijgi
ISPRS In . J. Geo-In . 2021,10, 733 2 o 25
conduc ing su eys usually implies ec ui ing passenge s and asking hem o answe a
ques ionnai e, which makes i di icul o collec a high numbe o answe s and in oduces
a bias owa ds ci izens who ha e a posi i e gene al a i ude owa ds pa icipa ion in a
su ey. In bo h cases, spo obse a ions and su eys, he e is a clea gap ha conce ns
he a ailabili y o he da a a ailable o u ban planne s, and hus also o planning he
bicycle in as uc u e.
To ackle he opic o cycling in ci ies and imp o e he quali y o in o ma ion ha u ban
planne s use o making decisions ela ed o spa ial in es men s, esea che s la gely u ned
o newly a ailable sou ces o da a. The majo i y o his esea ch ocuses on analyzing he
da a om bicycle-sha ing sys ems (BSS). The ela ed da a sou ces a e mainly he pooling
s a ions, which allow analyses based on he numbe s o a ailable bicycles and ee pa king
spo s a BSS s a ions [
4
] o check-ins and check-ou s on BSS s a ions [
5
–
7
]. While being
aluable o he logis ics o he BSSs, hese numbe s do no necessa ily gi e insigh s use ul
o planning he in as uc u e o ci izens who commu e by using p i a ely owned bicycles.
Planning o pa king acili ies o bicycles ha do no belong o he BSS is he e o e no
easible wi h hese da a alone. Ano he inexpensi e me hod o da a collec ion is GPS
acking o idden bicycles (e.g., ia sma phone), which is, howe e , mo e sui able o
analyzing ail pa e ns [
8
]. Some esea ch ela ed o u ban planning also analyzed he
mobile-phone da a gene a ed by mobile ne wo ks [
9
]. While collec ing elecommunica ions
ac i i y can p o ide an ex ensi e da ase , i does no di e en ia e bicycle use s om
o he passenge s.
In his pape , we add ess he gap in da a collec ion o u ban planning by ocusing on
social media da a. Because o he con inuous inc ease in he numbe s o sma phone owne s
and social media use s, we iden i y ano he oppo uni y o collec he needed bicycle-
ela ed da a and de elop a no el me hod o analyzing he demand o bicycle pa king
in as uc u e in u ban a eas. We p opose o use his me hod alongside o he s es ablished
in u ban planning in o de o en ich he da a co e age and p o ide mo e comp ehensi e
in o ma ion o making decisions ela ed o u ban in as uc u e in es men s, e.g., bicycle
pa king. We s a in Sec ion 2wi h he hypo hesis ha social media pos s can be use ul
o analyzing loca ions in ci ies in ela ion o bicycle usage. Sec ion 3in oduces ou
me hod and da a used o de ec bicycles on pho os om social media pos s, and
Sec ion 4
p esen s he p elimina y esul s o he bicycle de ec ion p ocess. Wi hin ou case s udy in
Sec ion 5
, we show ha ou da a p ocessing can p o ide subs an ial alue o planning
bicycle pa king acili ies in he ci y o D esden and discuss ad an ages and d awbacks o
ou app oach in Sec ion 6be o e concluding in Sec ion 7.
2. Rela ed Wo k
Fo he mos e ec i e p omo ion o cycling in a ci y, planning bicycle in as uc u e
should be demand-d i en, and so u ban planne s need o know he main cha ac e is ics
o he bicycle a ic lows in hei ci y [
10
]. Al hough he e is an inc easing in e es in
bicycles as pa o mul i-modal u ban in as uc u e, bicycle- ela ed esea ch in ecen
yea s ocused mainly on aspec s such as bicycle sa e y [
11
], posi i e impac s o cycling on
public heal h [
12
], a el mode choice [
13
] and ou e choice analysis [
10
,
14
]. By con as ,
less a en ion is paid o a ic enginee ing opics such as a ic coun s, a el imes, and
capaci ies [15].
In gene al, bicycle a ic olume da a a e ha d o ob ain. As opposed o mo o ized
a ic, bicycle a ic olume is s ongly a ec ed by he p esence and quali ies o bicycle
in as uc u e, ele a ion, mo o ized ehicles, wea he condi ions, e c. [
16
]. His o ically, e-
sea ch on cycling ac i i y elied on indi idual-le el su eys on household a el—me hods
ha a e esou ce in ensi e and can p oduce s a is ically un ep esen a i e samples dis-
o ing he indings o he quali a i e analysis [
17
]. Mode n me hods o bicycle a ic
es ima ion all in o wo ca ego ies: long- e m coun e s ha un con inuously, and sho -
e m measu emen s o ypically 1 o 28 days. To de i e obus demands based on sho e
obse a ion pe iods, he alues can be mul iplied by scaling ac o s and ac o g oups
ISPRS In . J. Geo-In . 2021,10, 733 3 o 25
accoun ing o daily, weekly and seasonal bicycle olume a iance gained h ough he
con inuous measu emen s [18].
In p ac ice, he ac ual coun ing and acing o bicycles can be execu ed using di e en
da a collec ion me hods, which should ollow quali y assu ance p ocedu es [
19
]. These
me hods include:
• Use o s a iona y senso s o coun passing bicycles,
• Analysis o public su eillance ideos h ough objec de ec ion,
• GPS- acking h ough de ices used by cyclis s,
• T acking o GPS de ices di ec ly moun ed on bicycles.
Adap ing adi ional me hods o mo o ehicle a ic moni o ing, nume ous echnical
solu ions and comme cial p oduc s o coun cyclis s wi h s a iona y senso s a e a ailable.
Fo example, e . [
20
] used da a om pneuma ic ubes on s ee s and adio beams on cycle
pa hs o hei s udy, while he da a o [
18
] we e ob ained om induc i e loop coun e s.
Simila o senso s, he isual de ec ion o bicycles using s a iona y came as is adap ing
well-es ablished echniques, in his case om compu e ision [
15
,
21
], and can p o ide
ehicle de ec ion, classi ica ion, coun ing and speed measu emen s in eal- ime [22].
In con as o s a iona y senso s o came as, GPS acking de ices a e capable o
collec ing da a om a comple e jou ney and can be di ided in o wo ca ego ies, depending
on he me hod o ob aining he da a. Fi s , he GPS da a can be ob ained h ough a de ice
used by he cyclis s, e.g., a sma phone applica ion. These da a a e no mally sha ed om
olun ee s ei he o scien i ic esea ch [
10
,
23
], o o comme cial use h ough mobile apps
such as S a a [
11
]. Second, he acking de ice can be di ec ly moun ed on o he bicycle.
Dockless BSS p o ides eal- ime GPS da a o e e y bicycle, which p o ides de ailed
insigh s in o bicycle-sha ing use s’ empo al and spa ial mobili y pa e ns [
24
,
25
]. Fo
example, e . [16] used GPS da a o BSS p o ided by he company Wa elo.
S a iona y senso s, a ic su eillance came as and GPS acking de ices di e no
only in he me hod o ob aining in o ma ion bu also in he cha ac e is ics o he da a
hey p o ide, namely he gained spa ial in o ma ion, he co e age o a ge g oups and
he su eyed bicycle s a e. In gene al, s a iona y senso s ha e he ad an age o co e ing
e e y single passing bicycle, while hey a e ixed o a de ined loca ion and he e o e only
p o ide poin - ela ed da a o mo ing bicycles. Visual analysis o a ic ideos can iden i y
mo ing and pa ked bicycles wi hin he ame and is able o gene a e ajec o y da a wi hin
he co e ed a ea when cyclis s a e acked o e consecu i e ames. While bo h ypes
o GPS acking de ices p o ide ajec o y da a o he whole ip, a majo disad an age
o his da a collec ion me hod is he c ea ion o biased “ olun a y esponse samples”,
because i only includes da a o people who ha e chosen o olun ee [
26
]. Fu he mo e,
as he acking de ice is no moun ed on o he bicycle, he s a us and posi ion o he
bicycle while he bicyclis is no using he bicycle is unknown. In con as , he s a us and
posi ion o in eg a ed GPS de ices can cons an ly be measu ed, which allows de ec ing he
loca ion whe e he bicycle is pa ked. Howe e , unlike in Asia, u ban mobili y planning
policies in Eu ope ocus on p i a e bicycle use [
27
], and public bicycles om BSS a e mo e
equen ly used o i s - and las -mile connec ion and leisu e ac i i ies and less equen ly
o commu ing [24,28].
As shown abo e, ob aining da a on pa ked bicycles is s ill challenging. In o ma ion
e ie al using social media da a can be a complemen a y way o da a collec ion. Social
media usage is widesp ead geog aphically as well as empo ally and has become a na -
u al pa o people’s daily li es. As a consequence, da a a e gene a ed implici ly by he
use s, p o iding an “in- he-wild sensing” o he ci y wi hou es ic ions o labo a o y
en i onmen s [29].
Social media pos s usually con ain ex and ime in o ma ion, wi h
po en ially mo e isual (images, ideos) and spa ial da a a ached, which allows loca ion
ex ac ion [30,31].
While he e a e se e al app oaches o ecognize low-le el (e.g., walking, si ing, e c.)
and high-le el (e.g., ea ing, shopping, e c.) ac i i ies mainly based on di e en sou ces
o social media da a [
32
,
33
], we wan o ex ac bicycle- ela ed in o ma ion solely using
ISPRS In . J. Geo-In . 2021,10, 733 4 o 25
images om social media pos s. Rega ding he iden i ied da a cha ac e is ics o bicycle-
ela ed measu es s a ed abo e, his has wo ad an ages. Fi s , using images om social
media po en ially allows us o co e he whole a ea o he ci y, depending on he equency
o pos s, and ob ain in o ma ion on a la ge a ie y o bicycle usage. Second, we can
dis inguish be ween mo ing and pa ked bicycles using simila objec de ec ion me hods as
implemen ed o s a iona y a ic su eillance ideos.
3. Me hod
Ou app oach o coun ing bicycles in images om social media pos s consis s o wo
s eps. Fi s , we applied a s a e-o - he-a objec de ec ion algo i hm (Sec ion 3.2) in o de o
de ec and localize bicycles and pe sons in each image. Using he de ec ed pe sons, we hen
classi ied each de ec ed bicycle as ei he mo ing o s a iona y (Sec ion 3.2.1). Fo e alua ion
(Sec ion 3.3) and pa ame e selec ion (Sec ion 3.4), we u he mo e labeled an app op ia e
da ase (Sec ion 3.1).
3.1. Da ase
In o de o quan i a i ely e alua e he easibili y o using social media da a o bicycle
a ic analysis, we used he YFCC100m [
34
] da ase because i is one o he la ges open-
sou ce da ase s o i s kind wi h a collec ion o 100 million pos s om he social media si e
Flick . Each pos con ains an image o ideo as well as addi ional in o ma ion, such as
loca ion, ime o cap u e and ags. All images we e aken in he yea s be ween 2004 and
2014 and a e sca e ed ac oss he whole wo ld. As we a e mainly in e es ed in da a om
u ban a eas, we selec ed a subse o images aken in a single ci y. This subse con ains
30,922 images
wi h loca ion me ada a indica ing ha hey we e eco ded in he ci y o
D esden, Ge many.
3.1.1. Bicycle Anno a ions
We manually anno a ed all bicycles in he subse o images. Each bicycle is labeled
wi h a bounding box and assigned one o wo ca ego ies: s a iona y i he bicycle is cu en ly
pa ked, o mo ing i i is being idden, wheeled o o he wise in use. O he 30,922 images,
2219
(
7.2%
)
con ain a leas one bicycle, wi h 1457
(
4.7%
)
images con aining s a iona y
bicycles and 976
(
3.2%
)
images con aining mo ing bicycles. As Figu e 1shows, mos
images (1204, 54.3%) con ain exac ly one bicycle. Howe e , images wi h signi ican ly la ge
numbe s o bicycles occu as well, e.g., 100 images (4.5%) depic mo e han six bicycles. In
o al, we labeled 4913 bicycles, o which 3038 (61.8%)a e s a iona y and 1875 (38.2%)a e
mo ing. Figu e 2shows a ew examples.
Figu e 1.
This his og am shows he numbe s o images in he D esden subse o he YFCC100m
da ase con aining be ween one and six, and mo e han six bicycles.
ISPRS In . J. Geo-In . 2021,10, 733 5 o 25
Figu e 2.
Examples om ou anno a ed D esden subse o he YFCC100m da ase . We manually
labeled bo h mo ing (yellow boxes) and s a iona y (cyan boxes) bicycles.
3.2. Objec De ec ion
In o de o au oma ically and eliably coun he numbe o mo ing and s a iona y
bicycles in an image, we u ilized a s a e-o - he-a objec de ec ion algo i hm. The ask
o objec de ec ion comp ises localiza ion o objec s in he image, usually by es ima ing
he coo dina es o bounding boxes aming he objec s, as well as classi ying each objec
using a se o p ede ined ca ego ies. Nume ous app oaches o objec de ec ion ha e been
p esen ed in ecen yea s [
35
–
39
]. They all use con olu ional neu al ne wo ks (CNNs) and
a e ained on he la ge-scale COCO (Common Objec s in Con ex ) [
40
] da ase . COCO
con ains mo e han 200,000 images labeled wi h objec bounding boxes o 80 di e en
ca ego ies such as ca , bicycle, pe son, couch, o ange, e c. Fo all expe imen s in his
wo k, we used he ecen ly p esen ed E icien De [
35
] objec de ec ion algo i hm, which
has been p e- ained on he COCO da ase , as i p o ides s a e-o - he-a pe o mance.
Compa ed o he p e ious bes me hod [
41
], E icien De achie es a signi ican ly highe
mean a e age p ecision (mAP) on he challenging COCO da ase (54.4% s. 50.7%) while
being compu a ionally mo e e icien . Compu ing objec de ec ions o one image on an
N idia Ti an V GPU akes 285 ms wi h E icien De , while [
41
] equi es 489 ms, i.e., almos
wice as long.
Gi en an image
I
, he objec de ec ion algo i hm compu es a se
P
o objec p oposals
Pi= (bi
,
ci
,
si)∈ P
. Each objec p oposal is de ined by a bounding box ( ec angle)
bi
wi h
image coo dina es
[xi,1
,
yi,1
,
xi,2
,
yi,2]
, an objec class
c
(e.g., bicycle), and a con idence sco e,
s
which can be loosely in e p e ed as an es ima e o he likelihood ha he objec p oposal
is co ec . In p ac ice, objec p oposals ha ha e a con idence sco e below a h eshold
θs
a e
disca ded. This h eshold mus be chosen app op ia ely in o de o minimize he numbe
o alse de ec ions while maximizing he numbe o co ec de ec ions.
In he ollowing, we a e only in e es ed in bicycle de ec ions
Pb∈ Pb⊆ P
and pe son
de ec ions Pp∈ Pp⊆ P, wi h Pb∩ Pp=∅:
∀Pi∈ P :(ci=bicycle ⇔Pi∈ Pb)
∧(ci=pe son ⇔Pi∈ Pp).(1)
3.2.1. Mo ing Bicycles
In o de o di e en ia e be ween mo ing and s a iona y bicycles, we le e age he
abili y o he objec de ec o o localize people in addi ion o bicycles. We assume ha

ISPRS In . J. Geo-In . 2021,10, 733 6 o 25
i a bicycle is loca ed igh below a pe son o igh nex o a pe son, his bicycle is being
handled by ha pe son and is hus non-s a iona y o mo ing. In ha case, he cen e o he
bounding box o a de ec ed pe son
Pp
mus be loca ed abo e he bounding box cen e o
bicycle Pb. We desc ibe his ela ion ia he ollowing indica o unc ion:
χ(Pb,Pp) = (1 i yp,1 +yp,2 >yb,1 +yb,2 ,
0 else. (2)
Since a pe son mus be loca ed in e y close p oximi y o a mo ing bicycle, we assume
a minimal o e lap o hei espec i e bounding boxes. We measu ed his o e lap using he
in e sec ion-o e -union (IoU) me ic, which compu es he a io o he o e lapping a ea o
he bounding boxes o hei uni ied a ea:
IoU(Pb,Pp) = A(bb∩bp)
A(bb∪bp)∈[0,1]. (3)
Fo e e y bicycle de ec ion
Pb∈ Pb
and e e y pe son de ec ion
Pp∈ Pp
, we de ine
an o e lap ma ix Cwi h:
Cbp =χ(Pb,Pp)·IoU(Pb,Pp). (4)
Using he Hunga ian me hod [
42
], we ind a maximum o e lap assignmen
H
based
on
C
. I a bicycle
Pb
is assigned o a pe son
Pp
wi h
Cbp >θp
, we de ine he bicycle as
mo ing and as s a iona y o he wise:
∀Pi∈ Pb:((∃Pj∈ Pp([i,j]∈ H ∧ Cij >θp)) ⇔Pi∈ Pbm)
∧(Pi/∈ Pbm ⇔Pi∈ Pbs),(5)
wi h
Pbm
and
Pbs
deno ing he se s o mo ing and s a iona y bicycle de ec ions, espec-
i ely. Figu e 3shows a ew examples o bicycles ha ha e been classi ied as s a iona y o
mo ing using his p ocedu e. We deno e he maximum assigned o e lap wi h a pe son o
a bicycle de ec ion Pbas Cb=maxpCbp.
Figu e 3.
Examples o co ec ly iden i ied s a iona y ( op, cyan boxes) and mo ing (bo om, yellow
boxes) bicycles, wi h all de ec ed pe sons ma ked in magen a boxes.
ISPRS In . J. Geo-In . 2021,10, 733 7 o 25
3.3. E alua ing De ec ions
In o de o e alua e he bicycle de ec ion me hod and o op imize i s pa ame e s, we
compa ed he p oposed bicycle de ec ions wi h he g ound u h anno a ions (c .
Sec ion 3.1.1
).
Gi en g ound u h anno a ions
Ti= ( ˆ
bi
,
ˆ
ci
,
ˆ
si)∈ T
and p oposed de ec ions
Pi∈ P
o
he same image, we de ine an o e lap ma ix Dwi h:
Dij =IoU(Pi,Tj)∀Pi∈ P,Tj∈ T . (6)
We ind a maximum o e lap assignmen based on
D
using he Hunga ian me hod [
42
].
I a p edic ion
Pi
is assigned o an anno a ion
Tj
wi h
Dij >θIoU
and same objec class
ci=ˆ
cj
, we ega d i as a ue posi i e
Pi∈ P p
. O he wise, i is a alse posi i e
Pi∈ P p
.
Likewise, i an anno a ion
Tj
is no assigned o a p edic ion, i coun s as a alse nega i e
Tj∈ T n
. As localiza ion accu acy is o li le ele ance o ou applica ion—we only
need o know he numbe o bicycles in an image—we se he IoU h eshold ela i ely
low, i.e.,
θIoU =
0.1. A e assigning p edic ions and anno a ions o each image, we can
compu e p ecision and ecall o e all images in o de o asses he quali y o he p edic ions.
P ecision is de ined as he a io o he numbe o co ec ly de ec ed objec s ( ue posi i es)
o he numbe o all de ec ions ( ue posi i es and alse posi i es):
p ecision =|P p|
|P| =|P p|
|P p|+|P p|. (7)
Recall is he a io o he numbe o co ec ly de ec ed objec s ( ue posi i es) o he
numbe o all p esen objec s ( ue posi i es and alse nega i es), i.e., all anno a ed objec s:
ecall =|P p|
|T | =|P p|
|P p|+|T n|. (8)
We pu posely do no use he mean a e age p ecision me ic (mAP, c . Sec ion 3.2) com-
monly u ilized in objec de ec ion li e a u e o e alua ion wi h he COCO
da ase [35,41].
The mAP me ic compu es he mean o he a ea unde he p ecision- ecall cu e o e a
ange o
θIoU ∈[
0.5,0.95
]
. While his me ic is well sui ed o compa ing he pe o mance
o objec de ec ion algo i hms independen o con idence h eshold
θs
and pa ially inde-
penden o IoU h eshold
θIoU
, i does no p o ide in o ma ion abou he accu acy o an
algo i hm in a p ac ical se ing, whe e hese h esholds mus be se o a speci ic alue.
3.4. De e mining Th esholds
We empi ically de e mine a con idence h eshold
θs
and a pe son assignmen h eshold
θpin o de o s ike an op imal balance be ween p ecision and ecall.
3.4.1. Con idence Th eshold
We adjus p ecision and ecall o all bicycle de ec ions—bo h mo ing and s a iona y—
by changing he con idence h eshold
θs
. We compu e ecall and p ecision o all alues o
θs∈[
0,1
]
and show he esul s in Figu e 4. The i s g aph in Figu e 4shows co esponding
ecall and p ecision alues, and he second and hi d g aphs show ecall and p ecision
alues co esponding o di e en h eshold alues. We iden i y a poin on he ecall-
p ecision cu e which is as close o he op- igh co ne as possible, i.e., maximizing bo h
p ecision and ecall. This poin co esponds o a h eshold o oughly
θs=
0.4, esul ing in
a p ecision o 0.96 and ecall o 0.81.
ISPRS In . J. Geo-In . 2021,10, 733 8 o 25
Figu e 4.
P ecision and ecall o bicycle de ec ions in ela ion o he con idence h eshold
θs
: he i s
g aph shows he ecall–p ecision cu e, while he second and hi d g aphs ep esen he ela ionships
o ecall and p ecision o he con idence h eshold sepa a ely.
3.4.2. Pe son Assignmen Th eshold
In o de o de e mine an op imal pe son assignmen h eshold
θp
, we conside ed all
ue posi i e bicycle de ec ions
Pb∈ P p
and hei assigned maximum pe son o e lap
Cb
.
Fo all h esholds
θp∈[
0,1
]
, we compu ed he ac ion o de ec ions ha a e co ec ly
classi ied as ei he mo ing o s a iona y. As Figu e 5shows, his classi ica ion accu acy
peaks a oughly 89.5%. We hus se he pe son assignmen h eshold o he co esponding
alue o θp=0.15.
Figu e 5.
Mo ing s. s a iona y: we plo he classi ica ion accu acy o a ange o pe son assignmen
h esholds θpin o de o iden i y an op imal alue.
4. Bicycle De ec ion Resul s
4.1. De ec ion Accu acy
In o de o assess he o e all accu acy o ou app oach, we compa e ou bicycle
de ec ions wi h he g ound u h bicycle anno a ions om ou da ase (c . Sec ion 3.1). As
he con usion ma ix in Table 1shows, we de ec ed a o al o 4157 bicycles in D esden, om
which 1589 we e classi ied as mo ing and 2568 as s a iona y. O hese 4157 de ec ions, only
160 we e inco ec , esul ing in a alse disco e y a e o 3.85% and equi alen ly a p ecision
o 96.1%. We co ec ly iden i ied 3997 o he 4913 bicycles in he da ase , hus achie ing a
ecall o 81.4%. This means ha we ha e adjus ed ou bicycle de ec ion me hod o ope a e
a he cau iously, i.e., he numbe o alse posi i es is signi ican ly lowe han he numbe
o alse nega i es. The as majo i y o alse nega i es can be di ided in o h ee ca ego ies:
small (i.e., low esolu ion) bicycles, pa ly occluded bicycles, and unusual pe spec i es.
Figu e 6shows one example image o each ca ego y. In such cases, he bicycles may be
di icul o ecognize e en o a human anno a o .
ISPRS In . J. Geo-In . 2021,10, 733 9 o 25
Table 1.
This con usion ma ix shows he numbe o bicycles o ce ain g ound u h classes ( ows)
being classi ied in o es ima ed classes (columns) by ou me hod. The
∑
-en ies indica e column- and
ow-wise sums. None indica es ei he no bicycle p esen o no co esponding bicycle de ec ed and
we e omi ed om he o e all sums.
T ue
Es ima ed Mo ing S a iona y None ∑
mo ing 1368 226 281 1875
s a iona y 194 2209 635 3038
none 27 133 - (160)
∑1589 2568 (916) 4157
4913
Figu e 6.
Mos common cases o alse nega i es, i.e., uniden i ied bicycles, om le o igh : small size o un a o able
ligh ing condi ions, pa ial occlusions, unusual pose o bicycle o came a.
The smalle numbe o alse posi i es all in o he ollowing ou ca ego ies: pa s o
comple e bicycles (i.e., possibly duplica es), o he wheeled objec s (such as mo o cycles,
wheelchai s o baby s olle s), a ic signs wi h bicycle pic og ams, and miscellaneous. We
p esen one example o each kind in Figu e 7.
Figu e 7.
Mos common cases o alse posi i es, i.e., w ongly de ec ed bicycles, om le o igh : smalle pa s o comple e
bicycles; o he wheeled objec s such as baby s olle s, wheelchai s and mo o cycles; a ic signs wi h bicycle pic og ams;
miscellaneous objec s such as musical ins umen s, chai s and came a ipods.
Wi hin he se o co ec ly iden i ied bicycles, we classi y mos o he mo ing bicycles
(1368 o 1594, 85.8%) and mos o he s a iona y bicycles (2209 o 2403, 91.9%) co ec ly.
False classi ica ions as mo ing mos commonly occu when a pe son is coinciden ally
loca ed in close p oximi y o a s a iona y bicycle, o when such a pe son is alsely de ec ed.
False classi ica ions as s a iona y occu when he o e lap be ween he bicycle and pe son
de ec ions is oo small, when he pe son was no de ec ed a all, o when he bicycle is
mo ing wi hou a pe son (e.g., moun ed on a ca ). We p o ide one example o each case
in Figu e 8.
ISPRS In . J. Geo-In . 2021,10, 733 16 o 25
5.2. Numbe o Pa ked Bicycles
NPB
s. Pe cen age o Pho os Con aining Pa ked Bicycles
PPB
s.
Numbe o A ailable Pa king Spo s NPS
A e iden i ying o which combina ions o
NPB
and
PPB
o e laps exis and whe e,
we u he analyze he o e laps in ela ion o he numbe o bicycle pa king spo s
NPS
(da a
downloaded om Open S ee Map ia he O e pass API. h ps://o e pass- u bo.eu/,
accessed on 5 June 2021) in o de o de ec loca ions o a po en ial pa king space de ici . Fo his
pu pose, we i s classi y g id cells in o h ee ca ego ies: (I) Cells con ain su icien pa king
capaci y; (II) Cells pa ially con ain pa king capaci y; (III) Cells con ain no pa king capaci y; and
one subca ego y: (a) Cells ha e close access o he neighbo ing cell’s pa king capaci y.
We classi y each cell in o one o he h ee ca ego ies o pa king capaci y based on he
ela ion o
NPB
and
NPS
o ha cell (Table 4). As
NPB
and
NPS
a e quan ized in o se s
o alue anges, we conside he uppe bounds o each pa king acili y. Fo example, i
NPS
is in he ange o 12 o 37, we assume
NPS =
37. I he uppe bound o
NPS
equals o
exceeds he uppe bound o
NPB
in he analyzed cell, hen we assume ha he cell con ains
su icien pa king capaci y (I). I he uppe bound o
NPS
is lowe han he uppe bound o
NPB
o a cell, we conside ha he cell pa ially includes pa king capaci y (II). O he wise,
we assume ha he cell con ains no pa king capaci y (III). Fo example, he e a e h ee
cells wi h
PPB
o 18.01 o 35.0% con aining 36 o 65 pa ked bicycles
NPB
. Two ou o h ee
cells do no con ain any pa king (
NPS =
0), and one con ains wo pa king acili ies: one o
capaci y up o 12 bicycles and he o he o capaci y 12 o 37 bicycles (Figu e 13). Since he
uppe bound o
NPS
is 49, i does no each he uppe bound o
NPB
o 65; in his manne ,
we conside he cell is pa ially p o ided wi h pa king capaci y.
Figu e 13.
Example o isual analysis o bicycle pa king capaci y in D esden by o e lapping 100
×
100 m g id cells o he NPB (g ay) and PPB (hashed) laye s in QGIS ( e sion 3.16).
Conside ing ha we implemen ed he abo e ca ego iza ion o iden i y how c i ical
he condi ion wi hin he cell ega ding he lack o pa king spo s gene ally is, we also
in oduced he subca ego y a o he cells ha ha e close access o he neighbo ing cell
pa king capaci y. The mo i a ion o his is he ac ha we we e able o iden i y cases
whe e he cell belonged o a ca ego y II o III bu pa king o he neighbo ing cell was
loca ed a he exac bo de be ween i s na i e cell and he analyzed cell. The e o e, he
subca ego y se es us o decide whe he he cell we analyze is less c i ical because he
pa king can be easily eached ou side he cell. Fo example, he e a e i e cells wi h a
PPB
o 8.1 o 18.0% con aining 36 o 65 de ec ions o pa ked bicycles
NPB
. Two ou o i e
cells con ain a su icien numbe o pa king spo s and h ee a e pa ially co e ed wi h

ISPRS In . J. Geo-In . 2021,10, 733 17 o 25
pa king capaci y. Howe e , wo ou o he h ee cells ha a e pa ially co e ed wi h pa king
capaci y allow easy access o he neighbo ing cell’s pa king. We conclude ha , e en hough
he whole ca ego y is o high ele ance, i is ai ly well co e ed wi h pa king acili ies
and, consequen ly, does no equi e immedia e a en ion. The esul s o he analysis a e
p esen ed in Tables 4and 5.
Table 4.
Numbe o 100
×
100 m g id cells ha : con ain su icien pa king capaci y (I); a e pa ially co e ed wi h pa king
capaci y (II); a e pa ially co e ed wi h pa king capaci y bu some cells ha e close access o he neighbo ing cell’s pa king
capaci y (IIa); con ain no pa king capaci y (III); o con ain no pa king capaci y bu some cells ha e close access o he
neighbo ing cell’s pa king capaci y (IIIa). Shades o g ay addi ionally show ele ances assigned in Table 2.
NPB
PPB [%] 0.1–3.0 3.1–8.0 8.1–18.0 18.1–35.0
1–3 / / / /
4–8
I: 30/61
II: 0/61
III: 31/61; IIIa: 6/31
I: 14/31
II: 0/31
III: 17/31; IIIa: 1/17
/ /
9–18
I: 3/8
II: 1/8
III: 4/8; IIIa: 2/4
I: 15/29
II: 2/29
III: 12/29; IIIa: 1/12
I: 1/7
II: 2/7
III: 4/7; IIIa: 2/4
/
19–35 / /
I: 6/17
II: 4/17; IIa: 1/4
III: 7/17; IIIa: 3/7
I: 0/1
II: 0/1
III: 1/1; IIIa: 1/1
36–65 /
I: 0/1
II: 0/1
III: 1/1; IIIa: 1/1
I: 2/5
II: 3/5; IIa: 2/3
III: 0/5
I: 0/3
II: 1/3
III: 2/3
Table 5.
Final designa ion o impo ance o 100
×
100 m g id cells in D esden (ma ked in shades o ed): low ( ed 10%),
mode a e ( ed 35%), and high impo ance ( ed 55%). The designa ion o impo ance is based on de ec ed su iciency o
pa king spo s ( his able) and assigned ele ance (Table 2).
NPB
PPB [%] 0.1–3.0 3.1–8.0 8.1–18.0 18.1–35.0
1–3 / / / /
4–8
Mode a ely insu icien
o insu icien numbe
o pa king spo s
Insu icien numbe o
pa king spo s / /
9–18
Mode a ely insu icien
numbe o pa king
spo s
Insu icien numbe o
pa king spo s
Mode a ely insu icien
o insu icien numbe
o pa king spo s
/
19–35 / /
Mode a ely insu icien
o insu icien numbe
o pa king spo s
Mode a ely insu icien o
insu icien numbe o
pa king spo s
36–65 /
Mode a ely insu icien
o insu icien numbe
o pa king spo s
Mode a ely insu icien
numbe o pa king
spo s
Insu icien numbe o
pa king spo s
5.3. Summa y o Resul s
Based on he in o ma ion gained om he p e ious s eps, we we e able o classi y
u ban a eas acco ding o he pa king su iciency in o a eas wi h mode a ely insu icien ,
mode a ely insu icien o insu icien o insu icien numbe o pa king spo s. Wi h his
app oach, we a e able o iden i y he mos c i ical a eas in D esden ela ing o bicycle
pa king. In Table 5we p esen he esul s o each combina ion o
NPB
and
PPB
: o each
cell, we epo he su iciency o pa king spo s and ma k in shades o ed whe he he a ea
inally has a low, mode a e, o high impo ance. By compa ing Tables 3and 4, we assess ha
he a eas a ound he cen al ain s a ion and he cen e o Al s ad a e he mos c i ical
in D esden. Fo emos , wi hin hese loca ions, he e is less han 50% o Ca ego y I cells
o he o e all cell numbe , while he numbe o Ca ego y III cells exceeds 50% o i e
ISPRS In . J. Geo-In . 2021,10, 733 18 o 25
ou o six ca ego ies. The e o e, we iden i y ha hese cells possess an insu icien o
mode a ely insu icien o insu icien numbe o pa king spo s. Second, we conside he
Flick da a ela ed o hese loca ions ele an because a signi ican pe cen age o pos ed
pho os con ains de ec ions o pa ked bicycles (up o 35%). Thi d, he e is a high numbe o
pa ked bicycle de ec ions in hese pho os (up o 65). Finally, we p e iously classi ied hose
cells as cells wi h high ele ance, which adds o he impo ance o esul s. We conclude ha
hese loca ions quali y as a p io i y o be u he inspec ed by u ban planne s in D esden
using o he a ailable da a sou ces. Following he same app oach, i is equally possible o
classi y each g id cell o a mo e de ailed o e iew, which we skip o his pape .
Alongside, we iden i y ha some a eas also appea c i ical acco ding o he numbe o
bicycle pa king spo s, bu due o he low pe cen age o pho os con aining pa ked bicycle
de ec ion, we classi y hem in o low ele ance cells. Subsequen ly, we conside hem much
less c i ical, and, inally, assign hem low impo ance. Needless o say, his does no mean
ha hose cells canno be inspec ed u he a e he mo e c i ical loca ions ha e been
esol ed.
Ou esul s also show ha some cells con ain signi ican ly mo e han enough pa king
capaci y ha i appea s o be in demand (Figu e 14). We de ec six cells o ha ype and
sugges hem as loca ions ha could be u he inspec ed o gain insigh s ha could p o e
use ul o imp o ing u u e decisions ela ed o he planning o he bicycle pa king.
Figu e 14.
Example o a isually iden i ied 100
×
100 m g id cell (le ) ha con ains signi ican ly mo e
han enough bicycle pa king capaci y ha i appea s o be in demand. The e a e se en pa king a eas
o capaci y up o 12 bicycles, while he e a e only 4 o 8 bicycles de ec ed inside he cell (NPB).
We also obse ed ha social media da a p o ide mo e da a o some a eas and ba ely
any o some o he s. Tha indica es ha mo e popula loca ions o e mo e ele an bicycle-
ela ed da a, while unpopula a eas emain poo ly co e ed. This popula i y ela es o he
popula i y wi hin he used Flick da ase . Fo example, he pa o he Regie ungs ie el
dis ic ha is su ounded by he s ee s Albe s aße, Wiga ds aße, and Glaciss aße
demons a es a signi ican numbe o a ailable bicycle pa king spo s; howe e , we did no
de ec any pa ked bicycles he e. Conside ing he numbe o pa king spo s, we conclude
ha he a ea is egula ly equen ed by cyclis s bu no in e es ing enough o pos ing i
on Flick . The bicycle pa king acks he e a e ins alled a ound a ca pa king a ea, se e al
esiden ial buildings, schools, and ins i u ions, such as he Saxon S a e Minis y o Jus ice.
The same e ec is also isible, e.g., along he Holbeins aße and Ta zbe g s ee s in he
Johanns ad dis ic whe e he acks a e ins alled in on o esiden ial buildings, an
ISPRS In . J. Geo-In . 2021,10, 733 19 o 25
a hle ic acili y, esea ch ins i u es, and a ci y’s communal se ice company. Acco ding
o a signi ican numbe o ins alled acks, i is appa en ha hese loca ions a e also well
equen ed by cyclis s bu a ely pos ed on social media, such as Flick .
In his chap e , we p esen ed an analysis o he u ban space o D esden. We demon-
s a ed he way social media da a can be used o gain in o ma ion ela ed o he bicycle
pa king si ua ion in an u ban a ea. Addi ionally, we de ec ed u ban a eas ha , in ou
opinion, need p io i ized a en ion om u ban planning expe s in he sense o u he
analysis ega ding po en ially missing bicycle pa king. In he ollowing chap e , we discuss
he esul s o ou esea ch and in oduce hem in he con ex o hei use ulness o he
desc ibed ask in u ban planning.
6. Discussion
6.1. Objec De ec ion on Social Media Da a as an Addi ional Da a Sou ce o Ob aining
Bicycle-Rela ed In o ma ion in U ban A eas
We in oduced a new me hod o ob aining bicycle- ela ed da a in u ban a eas and
demons a ed i in he ci y o D esden. In he case s udy, we showed ha ou me hod
p o ides aluable in o ma ion o planning u ban in as uc u e as we we e able o iden i y
a eas wi h a lack o pa king acili ies o bicycles. Since ou app oach is ocused on images
om social media ins ead o comme cial da a sou ces o su eys, we a e able o co e bo h
indi idually owned and publicly sha ed bicycles wi hin a geog aphical sp ead o e he ci y
ha is ypical o social media da a [
45
]. In Sec ion 4.4, we b ie ly compa e he dis ibu ion
o de ec ed bicycles o o he da a sou ces and desc ibe he dockless bicycle-sha ing sys em
(BSS) da a as a biased in- he-wild sensing o he same a ea as ou me hod. Rega ding
he loca ion, some o he a eas ha we iden i ied as ha ing a mode a ely insu icien o
insu icien numbe o pa king acili ies a e also isible as dense bicycle clus e s in he da a
o MOBI, such as he cen al ain s a ion. Up o his poin , ou me hod can be seen as a
subs i u e o collec ing da a om bicycle sha ing sys ems. This can al eady be in e es ing
o ci ies wi hou a BSS o companies planning o expand hei BSS o a new ci y as a sou ce
o modeling a bicycle ne wo k and s a ion loca ion inding [46].
Compa ed o a BSS, we can p o ide addi ional alue o u ban planne s o a eas ha
ha e a highe equency o bicycle de ec ions bu a e no su icien ly co e ed by he BSS
se ice. This o en ela es o he mos scenic a eas whe e sha ed bicycles a e no allowed o
be d opped o . In D esden, o example, we iden i ied cen al a eas in he dis ic Al s ad
o ha e a mode a ely insu icien o insu icien numbe o bicycle pa king acili ies, bu
e u ning BSS bicycles was no allowed he e. In some a eas in D esden ha a e also no
co e ed by he e u n zones o he local BSS company, we did no iden i y lacking pa king
acili ies: hese a e, e.g., he ci y pa k G oße Ga en and he cycling ou e along he Elbe
i e . The e we e a eas ha we iden i ied in ou case s udy as no in e es ing enough o
pos ing hem on Flick (i.e., social media). We conside his o be a a he mino d awback
compa ed o BSS da a because only some a eas we e equen ly isible in he MOBI da ase
(e.g., Johanns ad dis ic ), while he e we e no BSS bicycles p esen in o he a eas (e.g., in
he Regie ungs ie el dis ic ).
Addi ionally, we used a Flick da ase ha con ains da a om 2004 o 2014, bu he
popula i y o social media has kep g owing e e since, conside ing he ac ha he numbe
o mobile subsc ip ions inc eased om 2.33 billion in 2014 o 6.4 billion in 2021 [
47
]. We
hus a gue ha ou me hod can de ini ely p o ide addi ional alue o u ban planne s,
especially by using he mos ecen da a and om mo e popula social media pla o ms ha
also ely on pos ing images, such as
Ins ag am [48,49]
. In ha case, i would be necessa y
o addi ionally p ep ocess he da a o comply wi h he p i acy egula ions by anonymizing
hem [50], e.g., as ecen ly p oposed by [51,52].
Rega ding he e iciency o he bicycle de ec ion, ou app oach achie es a ecall o
81.4% (c . Sec ion 4.1) and hus does no ind e e y bicycle, bu i is signi ican ly mo e
e icien han manual iden i ica ion by humans. We ound ha , on a e age, ca e ul manual
anno a ion (c . Sec ion 3.1) akes oughly 30s pe image. While such manual de ec ions
may p o ide a signi ican ly highe ecall han ou ully au oma ic app oach, i would be
ISPRS In . J. Geo-In . 2021,10, 733 20 o 25
e y cos ly o apply a a la ge-scale (e.g., housands o images) se ing. Ou me hod, on
he o he hand, only equi es a mode a ely powe ul compu e wi h compa a i ely small
unning cos s.
In gene al, da a a ailabili y is he only limi ing ac o o ans e abili y in ou ap-
p oach, emphasizing he oppo uni ies social media da a can b ing o u ban analysis [
53
].
Requi emen s ega ding compu ing capaci y a e a he low o applying he me hod o
o he ci ies, as he p e- ained objec de ec ion algo i hm is u ilized o iden i ying and
classi ying bicycles, and con en ional GIS ope a ions enable spa ial explo a ion, analysis
and isualiza ion [
54
]. In compa ison o he o he me hods o ob aining bicycle a ic
da a men ioned in Sec ion 2, ou app oach is comple ely easible wi hou any s uc u al
ins alla ion, e.g., s a iona y coun ing s a ions o bicycle-moun ed GPS senso s.
6.2. Rele ance o Time in Ou App oach
As we used he YFCC100m da ase , which co e s a ime span o 10 yea s, we only
wo ked wi h empo ally comp essed loca ion da a. Compa ed o o he bicycle- ela ed
da a sou ces, his is a i s a majo d awback o ou app oach. Ne e heless, di e en
spa io- empo al analyses would s ill be easible: Fi s , i is possible o analyze cumula ed
equencies o bicycle de ec ions in he da ase wi h espec o empo al ca ego ies such as
weekdays, mon hs o yea s [
55
]. These spa ial- empo al pa e ns can p o ide insigh s in o
daily and seasonal ou ines o u ban cyclis s, and complemen and e i y he con inuous
da a s eams o s a iona y coun e s and BSS da a [
25
,
56
]. Second, ou me hod could be
used as an indica o o he success o exis ing in as uc u e, such as pa king acili ies o
bicycle highways. I would be possible o compa e he spa ial pa e n o de ec ed bicycles
be o e and a e he da e o cons uc ion and, he e o e, de ec changes in bicycle- ela ed
a ic lows o he densi y o de ec ed bicycles o each g id cell we used in ou case s udy.
The capabili y o social media o acqui e in o ma ion abou in e en ions in di e en e en s
has al eady been shown in nume ous s udies, e en hough mos o hem examined mo e
signi ican in e en ions such as in cases o a ic inciden s o na u al disas e s [57,58].
6.3. In luence o Bicycle De ec ion E o s and Po en ial Imp o emen s
In Sec ion 4.1, we p esen he esul s o ou objec de ec ion me hodology o iden-
i ying mo ing and s a iona y bicycles. In addi ion o a quan i a i e e alua ion, we also
p o ide examples o ailu e cases and iden i y he mos equen ca ego ies o alse neg-
a i es and alse posi i es. Al hough we conclude—based on ou case s udy— ha he
in o ma ion ex ac ed om social media da a is sui able o iden i ying a lack o pa king
acili ies o bicycles, we a e awa e o he laws o ou objec de ec ion pipeline. Fu he on,
we discuss hei impac on he indings o he case s udy as well as po en ial imp o emen s
wi hin ou app oach.
Wi h ega ds o con en , he de ec ion e o s can be subdi ided in o s ic ly bicycle-
ela ed e o s (i.e., missed bicycles o duplica e de ec ions) and hose o con usion wi h
o he objec s such as a ic signs o chai s. Fo he o me , we a gue ha he impac o
he alse de ec ions on he indings o he case s udy is ela i ely small, as we uned he
algo i hm o ope a e using mo e conse a i e h esholds and, he e o e, we a he end o
unde es ima e he numbe o bicycles. Fo images wi h mul iple bicycles de ec ed, his will
dec ease he o al numbe o bicycles, bu he numbe o images wi h bicycle de ec ions
will no be a ec ed. Wi h Tables 2and 3in mind, changing a class in he e ical di ec ion
o he able will mos likely no change he assigned ele ance o he g id cell. On he o he
hand, a alse nega i e (i.e., missed de ec ion) o a lone bicycle in an image can change he
ele ance acco ding o Table 2, bu acco ding o he dis ibu ion o he classes in Table 3, a
ele ance change will mos likely occu only be ween classes o low and medium ele ance.
I we o e es ima e he numbe o bicycles in ce ain images, o occasionally de ec alse
posi i es o o he wheeled objec s such as mo o cycles, we a gue ha assigning a highe
ele ance o hese cells is an accep able d awback o he ocus o ou case s udy, whe e we
quali y loca ions as a p io i y o be u he inspec ed by u ban planne s.
ISPRS In . J. Geo-In . 2021,10, 733 21 o 25
In con as o bicycle- ela ed de ec ions, o imp o e ou app oach, we conside mini-
mizing he alse posi i e de ec ions o miscellaneous objec s such as a ic signs. As we
men ion in Sec ion 3.2, we use an objec de ec o ha has been ained on he e y di e se
COCO [
40
] da ase , which con ains a la ge numbe o objec classes in images aken ac oss
all con inen s. Ou a ge applica ion, howe e , is na owe : we a e only in e es ed in he
de ec ion o bicycles and pe sons, and in ou case s udy, he en i onmen is limi ed o
he ci y o D esden. In such a case, ine- uning he objec de ec ion neu al ne wo k on he
na owe a ge domain has been shown o imp o e pe o mance [
59
]. In addi ion, i can
be easible o combine ou me hod wi h o he de ec ion algo i hms om compu e ision.
Fo example, one can employ an algo i hm ha de ec s bicycle a ic signs [
60
,
61
] in o de
o il e ou his ca ego y o alse posi i e de ec ions. Simila ly, one could employ objec
de ec ion o image classi ica ion [
62
] algo i hms ained on o he classes han bicycle and
pe son in o de o es ima e he likelihood o a bicycle de ec ion being a alse posi i e due
o con usion. Bicycle de ec ions, bo h co ec and alse, may also occu in images aken
indoo s. As hese a e o no in e es o ou applica ion, i would be easonable o disca d
such images au oma ically using an algo i hm ha can dis inguish indoo om ou doo
scenes [
63
]. Fo imp o ing he mo ing s. s a iona y classi ica ion accu acy, i would be
easible o ain an objec de ec ion algo i hm o di ec ly p o ide his classi ica ion ins ead
o elying on pe son de ec ions as we desc ibe in Sec ion 3.2.1. Whe he his would ac ually
wo k be e , howe e , depends on a a ie y o ac o s, such as he size and di e si y o
he aining da ase , he a chi ec u e o he con olu ional neu al ne wo k, and app op ia e
da a augmen a ion du ing aining [64].
Imp o ing bicycle de ec ion om social media pos s migh also include p ocessing
ags, ex ual desc ip ions, and emo icons used in a pos o ex ac use eac ions [
65
].
Sen imen analysis and ca ego iza ion o emo ions associa ed wi h a pos can be applied
o iden i y i pos ings ela ed o bicycles a e mo e posi i e o nega i e conno ed and in
ha way ob ain de ailed con ex ual in o ma ion ela ed o a pos and speci ic a eas [
66
,
67
].
Lea ning abou he con ex o an image could allow o, e.g., elimina e alse de ec ions, such
as musical ins umen s and ipods on conce s ages.
I an image con ains bicycles ha a e loca ed a a a dis ance, using only he GPS
loca ion associa ed wi h he image may esul in poo localiza ion o he bicycles. I would
be easible, howe e , o use isual localiza ion [
68
] and geome ic cues [
69
] in o de o
es ima e he pose o he came a and, subsequen ly, ex ac 3D in o ma ion abou he scene
om he image [
70
–
72
]. This would p o ide a mo e p ecise localiza ion o he de ec ed
bicycles.
7. Conclusions
In his pape , we in oduced a new me hod o ob aining bicycle- ela ed da a om
social media pos s. In he i s s ep, we used a p e- ained s a e-o - he-a objec de ec ion
algo i hm o de ec bicycles on a egional subse o he YFCC100m da ase . In he second
s ep, we di e en ia ed be ween mo ing and s a iona y bicycles as we le e aged he abili y
o he de ec ion algo i hm o de ec people and assume ha i a bicycle is loca ed igh
nex o o below a pe son, i is non-s a iona y o mo ing. Wi h ou me hod, we de ec ed
4157 bicycles in he ci y o D esden wi h an o e all p ecision o 96.1% and a ecall o 81.4%,
and classi ied 85.8% o he mo ing bicycles and 91.9% o he s a iona y bicycles co ec ly.
We hen conduc ed a case s udy, whe e we analyzed he gene al si ua ion in D esden
ela ed o pa king acili ies o bicycles using he esul s o he objec de ec ion. As a esul ,
we we e able o classi y u ban a eas acco ding o he su iciency o pa king acili ies and
he eby iden i y a eas ha need p io i ized a en ion om u ban planne s. Using he same
app oach, i would be possible o de ec u he mic o-loca ions wi hin hese u ban a eas.
Ou me hod p o ed as ele an because we we e able o gain meaning ul insigh s in o
he u ban a ea o D esden using social media da a. We conclude ha i p o ides signi ican
alue in planning he bicycle in as uc u e in a ci y, pa icula ly conside ing ha da a o
ha pu pose is o he wise di icul and expensi e o collec . We we e able o sho en he

ISPRS In . J. Geo-In . 2021,10, 733 22 o 25
da a acquisi ion mul iple imes in compa ison o adi ional me hods o da a collec ion in
u ban planning. Addi ionally, we can also p o ide a much la ge empo al co e age.
Clea limi a ions o using social media da a lie in he ac ha he da a a ailabili y
and co e age la gely depend on he usage o social media. The spa ial co e age is wo se
o unpopula han o popula u ban a eas, and empo al co e age (i.e., days, mon hs,
yea s, e c.) is wo se o yea s when social media was less popula . Howe e , bo h da a
a ailabili y and co e age depend on usage ends o he social media pla o m whose da a
we e used. Ha ing ha in mind, usage o he newes social media da a should enable e en
mo e p ecise spa io- empo al analyses. This p ecision may addi ionally be inc eased by
implemen ing u he objec de ec ion algo i hms in da a p ocessing.
By choosing D esden as ou case s udy, we bene i ed om a manageable amoun o
da a, as well as om ou own local expe ise. In he u u e, we in end o po en ially enla ge
he amoun o bicycle de ec ions o deal wi h by ocusing on a la ge ci y. Ano he way o
enla ge ou da ase would be o in eg a e da a om mo e social media pla o ms such as
Ins ag am o Twi e , depending on he a ailabili y o he da a. We also conside imple-
men ing o he da abases o in e es , especially hose p o iding s ee -le el images such
as Mapilla y. Fu he mo e, dealing wi h a highe numbe o de ec ions and la ge u ban
a eas o in e es equi es mo e sophis ica ed app oaches o localiza ion and isualiza ion,
so we in end o wo k on me hods o o ien a e he images by ma ching de ec ed objec s and
ci y u ni u e (e.g., benches, ligh ning objec s, e c.) and imp o e he echniques o isual
analysis and explo a ion.
Au ho Con ibu ions:
Concep ualiza ion, Mo is Zah ila, Flo ian Kluge and Ma in Knu a; me hod-
ology, Mo is Zah ila, Flo ian Kluge and Ma in Knu a; so wa e, Flo ian Kluge ; alida ion, Mo is
Zah ila, Flo ian Kluge and Ma in Knu a; o mal analysis, Mo is Zah ila, Flo ian Kluge and
Ma in Knu a; in es iga ion, Mo is Zah ila, Flo ian Kluge and Ma in Knu a; esou ces, Bodo
Rosenhahn, Jochen Schiewe and Di k Bu gha d ; da a cu a ion, Flo ian Kluge and Ma in Knu a;
w i ing—o iginal d a p epa a ion, Mo is Zah ila, Flo ian Kluge and Ma in Knu a; w i ing—
e iew and edi ing, Mo is Zah ila, Flo ian Kluge , Ma in Knu a, Bodo Rosenhahn, Jochen Schiewe
and Di k Bu gha d ; isualiza ion, Mo is Zah ila and Ma in Knu a; supe ision, Bodo Rosenhahn,
Jochen Schiewe and Di k Bu gha d ; p ojec adminis a ion, Mo is Zah ila, Flo ian Kluge , Ma in
Knu a, Bodo Rosenhahn, Jochen Schiewe and Di k Bu gha d ; unding acquisi ion, Bodo Rosenhahn,
Jochen Schiewe and Di k Bu gha d . All au ho s ha e ead and ag eed o he published e sion o
he manusc ip .
Funding:
This collabo a ion was ealized wi hin he DFG P io i y P og amme (SPP 1894/2) and
suppo ed by g an s COVMAP (RO 2497/12-2), TOVIP (SCHI 1008/11-1) and EVA-VGI 2 (BU
2605/8-2).
Da a A ailabili y S a emen :
Da a and sou ce code is a ailable a h ps://gi hub.com/ kluge /
bicycle_de ec ion.
Con lic s o In e es : The au ho s decla e no con lic o in e es .
Re e ences
1.
Suman an, V.; Fine, C.; Gonsal ez, D. Fas e , Sma e , G eene : The Fu u e o he Ca and U ban Mobili y; The MIT P ess: Camb idge,
MA, USA, 2017. [C ossRe ]
2.
Qiu, L.Y.; He, L.Y. Bike Sha ing and he Economy, he En i onmen , and Heal h-Rela ed Ex e nali ies. Sus ainabili y
2018
,10,
1145. [C ossRe ]
3.
Puche , J.; Buehle , R. Making Cycling I esis ible: Lessons om The Ne he lands, Denma k and Ge many. T ansp. Re .
2008
,
28, 495–528. [C ossRe ]
4.
F oehlich, J.; Neumann, J.; Oli e , N. Sensing and P edic ing he Pulse o he Ci y h ough Sha ed Bicycling. In P oceedings o
he 21s In e na ional Jon Con e ence on A i ical In elligence, Pasadena, CA, USA, 14–17 July 2009; pp. 1420–1426.
5.
Zhou, X. Unde s anding Spa io empo al Pa e ns o Biking Beha io by Analyzing Massi e Bike Sha ing Da a in Chicago. PLoS
ONE 2015,10, e0137922. [C ossRe ]
6.
Co co an, J.; Li, T.; Rohde, D.; Cha les-Edwa ds, E.; Ma eo-Babiano, D. Spa io- empo al pa e ns o a Public Bicycle Sha ing
P og am: The e ec o wea he and calenda e en s. J. T ansp. Geog . 2014,41, 292–305. [C ossRe ]
ISPRS In . J. Geo-In . 2021,10, 733 23 o 25
7.
E ienne, C.; La i a, O. Model-Based Coun Se ies Clus e ing o Bike Sha ing Sys em Usage Mining: A Case S udy wi h he VéLib’
Sys em o Pa is. ACM T ans. In ell. Sys . Technol. 2014,5, 1–21. [C ossRe ]
8.
Ko pilo, S.; Vi anen, T.; Leh ä i a, S. Sma phone GPS acking—Inexpensi e and e icien da a collec ion on ec ea ional
mo emen . Landsc. U ban Plan. 2017,157, 608–617.[C ossRe ]
9.
Reades, J.; Calab ese, F.; Se suk, A.; Ra i, C. Cellula Census: Explo a ions in U ban Da a Collec ion. IEEE Pe asi e Compu .
2007,6, 30–38. [C ossRe ]
10.
Hube , S.; Lissne , S.; Schnabel, A.; Lindemann, P.; F iedl, J. Modelling bicycle ou e choice in Ge man ci ies using open da a, MNL
and he bikeSim web-app. In P oceedings o he 2021 7 h In e na ional Con e ence on Models and Technologies o In elligen
T anspo a ion Sys ems (MT-ITS), He aklion, G eece, 16–17 June 2021. [C ossRe ]
11.
Chen, C.; Ande son, J.C.; Wang, H.; Wang, Y.; Vog , R.; He nandez, S. How bicycle le el o a ic s ess co ela e wi h epo ed
cyclis acciden s inju y se e i ies: A geospa ial and mixed logi analysis. Accid. Anal. P e .
2017
,108, 234–244. [C ossRe ]
[PubMed]
12.
Gö schi, T.; Ga a d, J.; Giles-Co i, B. Cycling as a Pa o Daily Li e: A Re iew o Heal h Pe spec i es. T ansp. Re .
2016
,
36, 45–71. [C ossRe ]
13.
Buehle , R. De e minan s o anspo mode choice: A compa ison o Ge many and he USA. J. T ansp. Geog .
2011
,19, 644–657.
[C ossRe ]
14.
Sco , D.M.; Lu, W.; B own, M.J. Rou e choice o bike sha e use s: Le e aging GPS da a o de i e choice se s. J. T ansp. Geog .
2021,90, 102903. [C ossRe ]
15.
Hoogendoo n, S.; Daamen, W. Bicycle Headway Modeling and I s Applica ions. T ansp. Res. Rec.
2016
,2587, 34–40. [C ossRe ]
16.
Pogodzinska, S.; Kiec, M.; D’Agos ino, C. Bicycle T a ic Volume Es ima ion Based on GPS Da a. T ansp. Res. P ocedia
2020
,
45, 874–881.
17.
Geh ke, S.R.; Rea don, T.G. Di ec demand modelling app oach o o ecas cycling ac i i y o a p oposed bike acili y. T ansp.
Plan. Technol. 2021,44, 1–15. [C ossRe ]
18.
Bei el, D.; McNee, S.; Mi anda-Mo eno, L.F. Quali y Measu e o Sho -Du a ion Bicycle Coun s. T ansp. Res. Rec.
2017
,2644, 64–71.
[C ossRe ]
19.
Tu ne , S.; Lasley, P. Quali y Coun s o Pedes ians and Bicyclis s: Quali y Assu ance P ocedu es o Nonmo o ized T a ic
Coun Da a. T ansp. Res. Rec. 2013,2339, 57–67. [C ossRe ]
20.
Hankey, S.; Lu, T.; Mondschein, A.; Buehle , R. Spa ial models o ac i e a el in small communi ies: Me ging he goals o a ic
moni o ing and di ec -demand modeling. J. T ansp. Heal h 2017,7, 149–159. [C ossRe ]
21.
Gup e, S.; Masoud, O.; Ma in, R.; Papanikolopoulos, N. De ec ion and classi ica ion o ehicles. IEEE T ans. In ell. T ansp. Sys .
2002,3, 37–47. [C ossRe ]
22.
Ghosh, A.; Sabuj, M.S.; Sone , H.H.; Sha abda, S.; Fa id, D.M. An Adap i e Video-based Vehicle De ec ion, Classi ica ion,
Coun ing, and Speed-measu emen Sys em o Real- ime T a ic Da a Collec ion. In P oceedings o he 2019 IEEE Region 10
Symposium (TENSYMP), Kolka a, India, 7–9 June 2019; pp. 541–546. [C ossRe ]
23.
Gillis, D.; Gau ama, S.; Van Gheluwe, C.; Semanjski, I.; Lopez, A.J.; Lauwe s, D. Measu ing Delays o Bicycles a Signalized
In e sec ions Using Sma phone GPS T acking Da a. ISPRS In . J. Geo-In . 2020,9, 174. [C ossRe ]
24.
Chen, Z.; an Lie op, D.; E ema, D. Dockless bike-sha ing sys ems: Wha a e he implica ions? T ansp. Re .
2020
,40, 333–353.
[C ossRe ]
25.
Ma, X.; Ji, Y.; Yuan, Y.; Van Oo , N.; Jin, Y.; Hoogendoo n, S. A compa ison in a el pa e ns and de e minan s o use demand
be ween docked and dockless bike-sha ing sys ems using mul i-sou ced da a. T ansp. Res. Pa A Policy P ac .
2020
,139, 148–173.
[C ossRe ]
26.
Basi i, A.; Haklay, M.; Foody, G.; Mooney, P. C owdsou ced geospa ial da a quali y: Challenges and u u e di ec ions. In . J.
Geog . In . Sci. 2019,33, 1588–1593. [C ossRe ]
27.
Wang, L. Planning o cycling in a g owing megaci y: Explo ing planne s’ pe cep ions and sha ed alues. Ci ies
2020
,106, 102857.
[C ossRe ]
28.
Weng, J.; Bäume , T.; Mülle , P. Bike-Sha ing Sys ems as In eg al Componen s o Inne -Ci y Mobili y Concep s: An Analysis
o he In ended Use Beha iou o Po en ial and Ac ual Bike-Sha ing Use s. In Inno a ions o Me opoli an A eas: In elligen
Solu ions o Mobili y, Logis ics and In as uc u e Designed o Ci izens; Planing, P., Mülle , P., Dehda i, P., Bäume , T., Eds.; Sp inge :
Be lin/Heidelbe g, Ge many, 2020; pp. 121–132. [C ossRe ]
29. Zhu, Z.; Blanke, U.; Cala oni, A.; T ös e , G. Human Ac i i y Recogni ion Using Social Media Da a. In P oceedings o he 12 h
In e na ional Con e ence on Mobile and Ubiqui ous Mul imedia, Luleå, Sweden, 2–5 Decembe 2013. [C ossRe ]
30. S ock, K. Mining loca ion om social media: A sys ema ic e iew. Compu . En i on. U ban Sys . 2018,71, 209–240. [C ossRe ]
31.
Middle on, S.E.; Ko dopa is-Zilos, G.; Papadopoulos, S.; Kompa sia is, Y. Loca ion Ex ac ion om Social Media: Geopa sing,
Loca ion Disambigua ion, and Geo agging. ACM T ans. In . Sys . 2018,36, 1–27. [C ossRe ]
32.
Gong, J.; Li, R.; Yao, H.; Kang, X.; Li, S. Recognizing Human Daily Ac i i y Using Social Media Senso s and Deep Lea ning. In . J.
En i on. Res. Public Heal h 2019,16, 3955. [C ossRe ]
33.
Cao, R.; Tu, W.; Yang, C.; Li, Q.; Liu, J.; Zhu, J.; Zhang, Q.; Li, Q.; Qiu, G. Deep lea ning-based emo e and social sensing da a
usion o u ban egion unc ion ecogni ion. ISPRS J. Pho og amm. Remo e Sens. 2020,163, 82–97. [C ossRe ]
ISPRS In . J. Geo-In . 2021,10, 733 24 o 25
34.
Thomee, B.; Shamma, D.A.; F iedland, G.; Elizalde, B.; Ni, K.; Poland, D.; Bo h, D.; Li, L.J. YFCC100M: The new da a in
mul imedia esea ch. Commun. ACM 2016,59, 64–73. [C ossRe ]
35.
Tan, M.; Pang, R.; Le, Q.V. E icien de : Scalable and e icien objec de ec ion. In P oceedings o he IEEE/CVF Con e ence on
Compu e Vision and Pa e n Recogni ion, Sea le, WA, USA, 16–18 June 2020; pp. 10781–10790.
36. Redmon, J.; Fa hadi, A. Yolo 3: An inc emen al imp o emen . a Xi 2018, a Xi :1804.02767.
37.
Du, X.; Lin, T.Y.; Jin, P.; Ghiasi, G.; Tan, M.; Cui, Y.; Le, Q.V.; Song, X. SpineNe : Lea ning scale-pe mu ed backbone o ecogni ion
and localiza ion. In P oceedings o he IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion, Sea le, WA, USA,
16–18 June 2020; pp. 11592–11601.
38.
He, K.; Gkioxa i, G.; Dollá , P.; Gi shick, R. Mask -cnn. In P oceedings o he IEEE In e na ional Con e ence on Compu e Vision,
Venice, I aly, 22–29 Oc obe 2017; pp. 2961–2969.
39.
Lin, T.Y.; Goyal, P.; Gi shick, R.; He, K.; Dollá , P. Focal loss o dense objec de ec ion. In P oceedings o he IEEE In e na ional
Con e ence on Compu e Vision, Venice, I aly, 22–29 Oc obe 2017; pp. 2980–2988.
40.
Lin, T.Y.; Mai e, M.; Belongie, S.; Hays, J.; Pe ona, P.; Ramanan, D.; Dollá , P.; Zi nick, C.L. Mic oso coco: Common objec s in
con ex . In Eu opean Con e ence on Compu e Vision; Sp inge : Be lin/Heidelbe g, Ge many, 2014; pp. 740–755.
41.
Zoph, B.; Cubuk, E.D.; Ghiasi, G.; Lin, T.Y.; Shlens, J.; Le, Q.V. Lea ning da a augmen a ion s a egies o objec de ec ion. In
Eu opean Con e ence on Compu e Vision; Sp inge : Be lin/Heidelbe g, Ge many, 2020; pp. 566–583.
42. Kuhn, H.W. The Hunga ian me hod o he assignmen p oblem. Na . Res. Logis . Q. 1955,2, 83–97. [C ossRe ]
43.
Landeshaup s ad D esden, S.u.T. Daue zähls ellen ü den Rad e keh . A ailable online: h p://www.d esden.de/media/pd /
S assenbau/Daue zaehls ellen_S ad plan.pd (accessed on 5 Augus 2021).
44. Jenks, G.F. The da a model concep in s a is ical mapping. In . Yea b. Ca og . 1967,7, 186–190.
45.
Jiang, B.; Ma, D.; Yin, J.; Sandbe g, M. Spa ial Dis ibu ion o Ci y Twee s and Thei Densi ies. Geog . Anal.
2016
,48, 337–351.
[C ossRe ]
46.
Bahado i, M.S.; Gonçal es, A.B.; Mou a, F. A Sys ema ic Re iew o S a ion Loca ion Techniques o Bicycle-Sha ing Sys ems
Planning and Ope a ion. ISPRS In . J. -Geo-In . 2021,10, 554. [C ossRe ]
47.
Tele onak iebolage LM E icsson E icsson Mobili y Visualize . A ailable online: h ps://www.e icsson.com/en/mobili y- epo /
mobili y- isualize ? =1& =1& =1& =8&s=1&u=1&y=2014,2021&c=1 (accessed on 14 Sep embe 2021).
48.
Domínguez, D.R.; Díaz Redondo, R.P.; Vilas, A.F.; Khali a, M.B. Sensing he ci y wi h Ins ag am: Clus e ing geoloca ed da a o
ou lie de ec ion. Expe Sys . Appl. 2017,78, 319–333. [C ossRe ]
49.
Gun e , U.; Önde , I. An Explo a o y Analysis o Geo agged Pho os F om Ins ag am o Residen s o and Visi o s o Vienna. J.
Hosp. Tou . Res. 2021,45, 373–398. [C ossRe ]
50.
Beigi, G.; Shu, K.; Zhang, Y.; Liu, H. Secu ing Social Media Use Da a: An Ad e sa ial App oach. In P oceedings o he 29 h on
Hype ex and Social Media, Bal imo e, MD, USA, 9–12 July 2018; pp. 165–173. [C ossRe ]
51.
Dunkel, A.; Löchne , M.; Bu gha d , D. P i acy-Awa e Visualiza ion o Volun ee ed Geog aphic In o ma ion (VGI) o Analyze
Spa ial Ac i i y: A Benchma k Implemen a ion. ISPRS In . J. Geo-In . 2020,9, 607. [C ossRe ]
52.
Löchne , M.; Fa hi, R.; Schmid, D.; Dunkel, A.; Bu gha d , D.; Fied ich, F.; Koch, S. Case S udy on P i acy-Awa e Social Media
Da a P ocessing in Disas e Managemen . ISPRS In . J. Geo-In . 2020,9, 709. [C ossRe ]
53.
Ma í, P.; Se ano-Es ada, L.; Nolasco-Ci ugeda, A. Social Media da a: Challenges, oppo uni ies and limi a ions in u ban
s udies. Compu . En i on. U ban Sys . 2019,74, 161–174. [C ossRe ]
54.
Ali and, M.; Hochmai , H.H. Spa io empo al analysis o pho o con ibu ion pa e ns o Pano amio and Flick . Ca og . Geog . In .
Sci. 2017,44, 170–184. [C ossRe ]
55.
McKenzie, G.; Janowicz, K.; Gao, S.; Yang, J.A.; Hu, Y. POI Pulse: A Mul i-g anula , Seman ic Signa u e–Based In o ma ion
Obse a o y o he In e ac i e Visualiza ion o Big Geosocial Da a. Ca og aphica 2015,50, 71–85. [C ossRe ]
56.
No dback, K.; Ma shall, W.E.; Janson, B.N.; S olz, E. Es ima ing Annual A e age Daily Bicyclis s: E o and Accu acy. T ansp. Res.
Rec. 2013,2339, 90–97. [C ossRe ]
57.
Rashidi, T.H.; Abbasi, A.; Magh ebi, M.; Hasan, S.; Walle , T.S. Explo ing he capaci y o social media da a o modelling a el
beha iou : Oppo uni ies and challenges. T ansp. Res. Pa C Eme g. Technol. 2017,75, 197–211. [C ossRe ]
58.
Resch, B.; Uslände , F.; Ha as, C. Combining machine-lea ning opic models and spa io empo al analysis o social media da a o
disas e oo p in and damage assessmen . Ca og . Geog . In . Sci. 2018,45, 362–376. [C ossRe ]
59.
Gi shick, R.; Donahue, J.; Da ell, T.; Malik, J. Rich ea u e hie a chies o accu a e objec de ec ion and seman ic segmen a ion.
In P oceedings o he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion, Columbus, OH, USA, 23–28 June 2014;
pp. 580–587.
60.
Reinde s, C.; Acke mann, H.; Yang, M.Y.; Rosenhahn, B. Objec ecogni ion om e y ew aining examples o enhancing
bicycle maps. In P oceedings o he 2018 IEEE In elligen Vehicles Symposium (IV), Suzhou, China, 26–30 June 2018; pp. 1–8.
61.
Reinde s, C.; Acke mann, H.; Yang, M.Y.; Rosenhahn, B. Lea ning con olu ional neu al ne wo ks o objec de ec ion wi h e y
li le aining da a. In Mul imodal Scene Unde s anding; Else ie : Ams e dam, The Ne he lands, 2019; pp. 65–100.
62.
K izhe sky, A.; Su ske e , I.; Hin on, G.E. Imagene classi ica ion wi h deep con olu ional neu al ne wo ks. Ad . Neu al In .
P ocess. Sys . 2012,25, 1097–1105. [C ossRe ]
63.
Zhou, B.; Laped iza, A.; Khosla, A.; Oli a, A.; To alba, A. Places: A 10 million Image Da abase o Scene Recogni ion. IEEE
TRansac ions Pa e n Anal. Mach. In ell. 2017,40, 1452–1464. [C ossRe ]
ISPRS In . J. Geo-In . 2021,10, 733 25 o 25
64.
Kluge , F.; Reinde s, C.; Rae z, K.; Schelske, P.; Wand , B.; Acke mann, H.; Rosenhahn, B. Region-based cycle-consis en da a
augmen a ion o objec de ec ion. In P oceedings o he 2018 IEEE In e na ional Con e ence on Big Da a (Big Da a), Sea le, WA,
USA, 10–13 Decembe 2018; pp. 5205–5211.
65.
Dunkel, A.; And ienko, G.; And ienko, N.; Bu gha d , D.; Hau hal, E.; Pu es, R. A concep ual amewo k o s udying collec i e
eac ions o e en s in loca ion-based social media. In . J. Geog . In . Sci. 2019,33, 780–804. [C ossRe ]
66.
Hau hal, E.; Bu gha d , D. Mapping Space-Rela ed Emo ions ou o Use -Gene a ed Pho o Me ada a Conside ing G amma ical
Issues. Ca og . J. 2016,53, 78–90. [C ossRe ]
67. Hau hal, E.; Bu gha d , D.; Dunkel, A. Analyzing and Visualizing Emo ional Reac ions Exp essed by Emojis in Loca ion-Based
Social Media. ISPRS In . J. Geo-In . 2019,8, 113. [C ossRe ]
68.
Sa lin, P.E.; Cadena, C.; Siegwa , R.; Dymczyk, M. F om coa se o ine: Robus hie a chical localiza ion a la ge scale. In
P oceedings o he IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion, Long Beach, CA, USA, 16–20 June 2019;
pp. 12716–12725.
69.
Kluge , F.; Acke mann, H.; Yang, M.Y.; Rosenhahn, B. Tempo ally consis en ho izon lines. In P oceedings o he 2020 IEEE
In e na ional Con e ence on Robo ics and Au oma ion (ICRA), Pa is, F ance, 31 May–31 Augus 2020; pp. 3161–3167.
70.
Kluge , F.; Acke mann, H.; Yang, M.Y.; Rosenhahn, B. Deep lea ning o anishing poin de ec ion using an in e se gnomonic
p ojec ion. In Ge man Con e ence on Pa e n Recogni ion; Sp inge : Be lin/Heidelbe g, Ge many, 2017; pp. 17–28.
71.
Kluge , F.; B achmann, E.; Acke mann, H.; Ro he , C.; Yang, M.Y.; Rosenhahn, B. Consac: Robus mul i-model i ing by
condi ional sample consensus. In P oceedings o he IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion, Sea le,
WA, USA, 14–19 June 2020; pp. 4634–4643.
72.
Kluge , F.; Acke mann, H.; B achmann, E.; Yang, M.Y.; Rosenhahn, B. Cuboids Re isi ed: Lea ning Robus 3D Shape Fi ing o
Single RGB Images. In P oceedings o he IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion, Nash ille, TN,
USA, 19–25 June 2021; pp. 13070–13079.