New Approaches for Geographic Location Propagation in Digital

Photograph Collections

Davi Oliveira Serrano de Andrade

, Hugo Feitosa de Figueirêdo

1,2

, Cláudio de Souza Baptista

and Anselmo Cardoso de Paiva

Department of Systems and Computation, University of Campina Grande, Campina Grande, Paraíba, Brazil

Federal Institute of Education, Science and Technology of Paraíba, Campus Monteiro, Monteiro, Paraíba, Brazil

Department of Informatics, Federal University of Maranhão, São Luís, Maranhão, Brazil

Keywords:

Geotag, Photograph Metadata, Geotag Propagation, Digital Photograph Collection.

Abstract:

The integration of GPS in smartphones, tablets and digital cameras becomes more present, resulting in a large

amount of multimedia ﬁles. As GPS receivers may not work well indoors, this problem may generate incorrect

locations, very distant from the real location where the picture was taken, or even generate no location at all.

So, to deal with these inconsistencies, this work proposes two novel location propagation techniques. These

techniques were validated through a comparative analysis with two other techniques. Some metrics were used

to validate the techniques: precision, recall and accuracy in the photographs location propagation. The results

prove that the correct choice for location propagation technique depends on the importance of each metric and

on the system user proﬁle. Besides the choice of the correct technique, we also show that the order of the

photographs that will receive the location propagation must be random.

1 INTRODUCTION

The technological advances in the last years have al-

lowed a widespread of electronic devices, such as:

digital cameras, smartphones and tablets. As a result,

people produce a large amount of multimedia ﬁles,

hindering the task of annotation and cataloging such

ﬁles. For example, manually organizing a collection

with thousands of photographs taken during a vaca-

tion trip is a very costly task.

Many approaches (Figueirêdo et al., 2012a; Naa-

man et al., 2004; Cooper, 2011; Tsay et al., 2009)

have been suggested for automatic organization of

photographs in order to reduce users’ effort. Such

systems use information contained in the metadata of

the photographs to help their automatic organization.

The metadata information used include date, time, ge-

ographic location and tags.

Some studies point that the place where the pho-

tograph was taken is one of the ﬁrst things that peo-

ple remember when they want to retrieve that photo-

graph (Naaman et al., 2004). This implies that the

geographic location of the camera at the moment that

the picture was taken is very important for the pho-

tographs organization process.

As time passes, the integration of GPS to smart-

phones, tablets and digital cameras becomes more and

more present, allowing the automatic storage of the

geographic location of the photographs in their meta-

data. Nevertheless, GPS receivers do not work well

indoors, with the possibility of generating no informa-

tion at all. Moreover, it takes a few moments for the

receiver to obtain the signal indicating the geographic

location of the photograph. If the signal is not cap-

tured, the camera will not georeference the picture, or

will use the location of the last one, which will prob-

ably be incorrect. Some smartphones use the A-GPS

(Djuknic and Richton, 2001) technology in order to

minimize these problems, but the locations indicated

may be very distant from the actual position where the

photograph was taken.

In order to optimize the task of organizing user’s

photographs, geographic information is needed. The

use of GPS allows the easy obtainment of the loca-

tion of the photographs, but the failures of the receiver

give the user an extra work. Manually correcting fail-

ures caused by the GPS receiver is a hard task, and for

this reason some location predicting techniques and

geotag propagation techniques were developed (Hays

and Efros, 2008; Ivanov et al., 2012; Gong et al.,

Oliveira Serrano de Andrade D., Feitosa de Figueirêdo H., de Souza Baptista C. and Cardoso de Paiva A..

New Approaches for Geographic Location Propagation in Digital Photograph Collections.

DOI: 10.5220/0004895000920099

In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 92-99

ISBN: 978-989-758-029-1

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2011; Gao et al., 2012; Lacerda et al., 2013).

Aiming at the need for reducing the efforts in-

volved in the geographic annotation task, the main

contributions of this work are:

• The Trajectory technique: uses the photograph

owner’s trajectory;

• The Shared Events technique: uses shared events

captured by distinct cameras;

• A comparative analysis between the proposed

techniques and two existing propagation tech-

niques: Temporal Clustering and Social Correla-

tion, where the ﬁrst one is the reproduction of the

work by Lacerda et al. and the second one re-

produces the work by Gong et al. (Lacerda et al.,

2013; Gong et al., 2011).

The comparison is performed through an experiment,

concerning the following metrics: precision, recall

and accuracy. To the best of our knowledge, there has

not been done an analysis comparing the geotag prop-

agation techniques in the literature yet. The experi-

ments show that the four studied techniques present

good results for different scenarios. So, the most suit-

able propagation technique will depend on the nature

of data and on users’ proﬁle.

The rest of this paper is organized as follows. In

Section 2 we discuss some related works. In Section 3

we focus on two new technique for photograph loca-

tion propagation and present two existing techniques

used to validate the proposed ones. Section 4 de-

scribes the database, the metrics used, and the way

the techniques were compared. Section 5 presents

and discusses the results. Finally, Section 6 presents

the conclusions and discuss further work to be under-

taken.

2 RELATED WORK

In this section we present some works related to the

problem of annotating geographic location into pho-

tographs. The related works are organized as follows:

concerned with the prediction of the users’ location;

concerned with the use of textual tags; works that deal

with image processing; and ﬁnally works that deal

with the propagation of geotags.

Gong et al. work to predicting the location of peo-

ple in order to improve mobile phone services (Gong

et al., 2011). The proposed idea used a social corre-

lation model to predict the location of people based

on the location of their personal contacts. Due to the

simplicity of this model, it is possible to adapt it for

use in the context of digital photograph collections.

Gao et al. address the problem of predicting the

location as a whole in order to overcome problems

such as trafﬁc planning and directed advertising (Gao

et al., 2012). The authors proposed a model to predict

user location based on a visitation record taking the

spatiotemporal context into account. Despite the sim-

plicity of the model, the idea is not simple considering

geotags described as latitude and longitude.

Hays and Efros proposed an algorithm called

“im2gps”, that estimates the geographic location of a

photograph based on the geographic location of other

images with high visual similarity (Hays and Efros,

2008). Ivanov et al. proposed the propagation of

geotags based on the combination of detection of re-

peated objects and user’s trust modeling (Ivanov et al.,

2012). The idea is the propagation of geotags using

other photographs with geotags.

These approaches (Hays and Efros, 2008; Ivanov

et al., 2012) can become complicated when applied to

personal photograph collections, since many times the

people present in the image occupy the larger area of

the picture. Considering the people do not represent

a location, the ideas of visual similarity and detection

of repeated objects are threatened.

Hollenstein and Purves suggest that the geo-

graphic location must be deﬁned by the manner in

which the users describe the place instead of lati-

tude and longitude (Hollenstein and Purves, 2013).

Tags like “Cristo Redentor”(“Christ the Redeemer”)

should be added to the location. The use of tags re-

lated to the geographic location is unreliable when the

tags are deﬁned by the user, and are also hard to pro-

pose when not deﬁned by the user.

Lacerda et al. focus on inconsistencies in the loca-

tions of photographs as well as in the propagation of

locations. The proposed idea uses a temporal cluster-

ing to ﬁnd the location based on the temporally closer

photograph (Lacerda et al., 2013).

CrEve (Zigkolis et al., 2012) is a collaborative

event annotation framework that used the content of

photographs found in social media sites. One of

the themes approached is the inconsistency of pho-

tographs metadata, but the user must create an event

using social media, and this framework does not focus

on personal photographs.

The present work proposes two propagation tech-

niques and performs a comparison between the pro-

posed techniques and two existing ones.

3 PROPAGATION TECHNIQUES

For all the techniques that will be explained, consider:

• U: the set of users;

NewApproachesforGeographicLocationPropagationinDigitalPhotographCollections

• u: a person in U;

• F

: the set of all photographs of a person P;

• FC

: the set of all photographs that are not in F

and where the person P appears;

• F

: a photograph in F

represented by a tuple in

the form <geo, h, d, dh>, where "geo" is the longi-

tude and latitude of the photograph, "h" the hour

of the day, "d" is the day of the week in which the

photograph was taken and "dh" is the exact date

and time in which the photograph was taken;

• FC

: is a photograph in FC

with the same struc-

ture of F

;

• dif(d1,d2): the function that computes the differ-

ence, in hours, between two dates;

• LT: time threshold with value of 24 hours;

• LS: similarity threshold between events with

value 60%.

3.1 Existing Techniques Analysed

In this subsection we explain the two techniques used

to validate the two proposed ones: Temporal Cluster-

ing and Social Correlation. Temporal Clustering, as

the name suggests, creates temporal clusters of pho-

tographs to ﬁnd the location where the photograph

was taken. Social Correlation ﬁnds the location based

on the location of the most correlated contact of the

user. Next we present the formalization of both tech-

niques.

3.1.1 Temporal Clustering

The Temporal Clustering technique (Lacerda et al.,

2013) makes a temporal segmentation in F

using a

time t

max

, in minutes, passed as parameter. Consider-

ing t

max

, the technique will separate F

in k clusters

of photographs. So, a cluster g

is a single subset of

such that:

[

j=1

= F

(1)

The clusters have photographs with temporal dis-

tances smaller than or equal to t

max

minutes, consid-

ering consecutive photographs, and can be deﬁned in

the following way:

∈ g

; (2)

i+1

∈ g

i f (t

i+1

−t

) ≤ t

max

; (3)

i+1

∈ g

s+1

i f (t

i+1

−t

) > t

max

. (4)

For each cluster g

, the iteration will search for

photographs that have no location. For each photo-

graph with no location, it is sought a photograph F

with location that is temporally closer in the same

cluster of F

and then the location of F

is propagated

to F

3.1.2 Social Correlation

The Social Correlation technique (Gong et al., 2011)

uses the geographic locations of possible neighbors

to ﬁnd the location of the user in the exact moment

where the photograph was taken.

Before ﬁnding the location of a photograph, we

need to ﬁnd the person who has the highest social cor-

relation with the user who took the photograph F

Let Q

be the number of photographs in which the

user u

and the user u

appear. The correlation be-

tween the user i and the user j is R

, deﬁned as fol-

lows:

(5)

This way, R

, indicates the percentage of pho-

tographs in which the user u

participates with respect

to the photographs of the user u

. Let C

be the set of

neighbor contacts of a user u. Considering that we

will suggest the location of a photograph F

, we must

ﬁnd the contact u with the highest social correlation

and who has at least one photograph inside LT, con-

sidering F

. We have:

∀ j ∈ C

→ R

≤ R

; (6)

∃F

∈ F

| di f (dh(F

), dh(F

)) ≤ LT. (7)

Let before(v,F

) be the function that ﬁnds the last

photograph taken before F

in which the neighbor v

appears and let after(v,F

) be the function that ﬁnds

the ﬁrst photograph taken after F

where the neighbor

v appears. These two functions look for a photograph

in LT with respect to the photograph passed as param-

eter, and in the case the photograph does not exist,

returns empty. So, the social correlation is deﬁned as:

= Centroid(be f ore(u, F

), a f ter(u, F

be f ore(u, F

), a f ter(u, F

)); (8)

3.2 Proposed Techniques

In this subsection we explain the Trajectory and the

Shared Events techniques. The ﬁrst one follows the

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

idea that a user keeps a straight trajectory between a

previous photograph and the next one. The second

technique uses the idea that the photographs belong-

ing to a same event are in the same geographic area.

The photographs used in the experiment contain the

following information for using the techniques: date

and time, geographic location and annotated people.

3.2.1 Trajectory

Lacerda et al. proposes the use of the speed of the

user to ﬁnd inconsistencies in the locations of the

photographs (Lacerda et al., 2013). Inspired in the

idea of the users’ speed, the Trajectory technique as-

sumes that the trajectory between known locations is

a straight line to propagate the location by means of

this trajectory. If the technique is trying to ﬁnd a lo-

cation for a photograph present in the collection of a

user u, named F

, it uses the trajectory of user u and

the trajectory of all contacts present in a photograph

. Let F be the set of all photographs in the image

database, and F

∈ F is a photograph that we want to

compute the location.

Consider for each person P present in F

= F

∪ FC

(9)

We may decompose FA

into two subsets:

• FA

geo

: the set of all photographs in FA

that have

a geographic location;

• FA

noGeo

: the set of all photographs in FA

that do

not have a geographic location;

To propagate a geographic location for the photo-

graph F

∈ FA

noGeo

we must ﬁnd the photographs f

∗

and f

∗

where:

, f

∈ FA

geo

; (10)

dh( f

) < dh(F

); (11)

dh( f

) > dh(F

); (12)

di f (dh( f

), dh(F

)) ≤ LT ; (13)

di f (dh( f

), dh(F

)) ≤ LT ; (14)

∗

= MIN(dh(F

) − dh( f

)); (15)

∗

= MIN(dh( f

) − dh(F

)); (16)

To compute the possible location geo

for F

con-

sidering the trajectory of P, we use a linear interpola-

tion based on:

z =

dh(F

) − dh( f

∗

)

dh( f

∗

) − dh( f

∗

)

; (17)

geo

= f

∗

.(1 − z) + f

∗

.z; (18)

The linear interpolation does not consider earth’s

curvature, but it is used for small distances, so the

error will be not relevant.

We repeat this process for each person P in F

and

compute the location of F

as the centroid of the set

composed by all geo

ﬁnded.

3.2.2 Shared Events

The work by Figueirêdo et al. proposes a technique

for detection of event in digital photographs captured

by distinct cameras (Figueirêdo et al., 2012b). The

detection of events separates all the possible events in

the collection of photographs and computes a similar-

ity index between these events. We deﬁne:

• E : the set of all events;

• E

∈ E: the event, owned by the user u, that con-

tains the photograph F

;

• SM

i,k

, which represents the similarity between the

events E

and E

;

• EC ⊂ E: the set of all events that have SM

i,k

≥ LS;

• FE: the set of all photographs with geographic

location in all events of EC;

It is proposed the use of these events found by

the work of Figueirêdo et al. to ﬁnd the location

of the photograph F

, assuming that the photographs

present in a same event are in a short distance. To

propagate a geographic location for the photograph

we need to ﬁnd the photograph p

∗

where:

p ∈ FE; (19)

∗

= MIN(di f (dh(p), dh(F

))); (20)

The geographic location of p

∗

is propagated to F

3.3 Examples

In this subsection we expose an example for each of

the proposed techniques. First we present an exam-

ple for the Trajectory technique and then we show an

example for the Shared Events technique.

3.3.1 Trajectory

Figure 1: Trajectory Example.

Considering that some friends took photographs in

Lisbon and that only one of the photographs is miss-

ing the location. People p1, p2, p3 and p4 appear in

NewApproachesforGeographicLocationPropagationinDigitalPhotographCollections

, the missing location photograph, but only p1, p2

and p3 have a related trajectory in T

. Figure 1 shows

the trajectories of p1, p2 and p3 as a straight black

line between the markers. The triangle’s vertices are

the possible locations related to the trajectories. The

point inside the triangle would be the centroid of all

the possible locations, and that point is the location

propagated.

3.3.2 Shared Events

Two friends, f1 and f2, went to a 5 hours party. The

ﬁrst likes to take several photographs and the sec-

ond just likes to picture the highlight moments of the

party. Supose the photographs taken by f2 have f1

and lack location. At the end of the party f1 took pho-

tographs every 20 minutes and f2 took every 1 hour.

The detection of similar events technique (Figueirêdo

et al., 2012b) will ﬁnd two events (for each user f1

and f2) with high similarity. The propagation will

occur for each photograph p1 of f2 with the nearest

photograph p2 of f1 related to p1. In this case the

photographs set of f2 will have ﬁve photograph rep-

resenting each hour (p1,p2,p3,p4,p5). The location of

the photograph taken at the hour 1 by f1 will be prop-

agated to the photograph p1, the location of the pho-

tograph taken at the hour 2 by f1 will be propagated

to the photograph p2, and so on.

4 VALIDATION

In this section we expose the method used to collect

data, the features of the photograph database used

to test the techniques; then we explain the metrics

adopted to compare the observed techniques, and ﬁ-

nally we show how the metrics were validated to gen-

erate the results.

4.1 Database

We used a database containing approximately 7900

photographs of 41 users. Each photograph has the

following metadata: date and time of capture, geo-

graphic location and people present. Only 13% of the

pictures came from cameras with integrated GPS, so

87% of the locations were indicated by the owners of

the photographs.

4.2 Metrics

The metrics adopted were precision, recall and accu-

racy, and each of them is measured with basis on three

numbers:

• Sc = Number of correct propagations;

• Se = Number of erroneous propagations;

• Sn = Number of photographs that did not receive

propagations;

The metrics are computed in the following man-

ner:

Recall :

(Sc + Se + Sn)

; (21)

Precision :

(Sc + Se)

; (22)

Accuracy :

(Sc + Sn)

(Sc + Se + Sn)

; (23)

4.3 Training Set

The choice of the photographs that make part of the

training set occurred in three ways: by photograph,

by user and by event.

By Photograph:

Considering the set F containing all the stored

photographs, the training set TS is formed by ran-

domly chosen photographs in F.

By User:

Considering the set U containing all the users and

the set RU containing the randomly chosen users, the

training set TS is formed by the union of all pho-

tographs of the users in RU.

∀u ∈ RU → F

⊂ T S; (24)

By Event:

Considering the set E containing all the events

and the set RE containing all of the randomly cho-

sen events, the training set TS is formed by the union

the all photographs related to the events of RE. Let

be the set of photographs related to some event:

∀e ∈ RE → FE

⊂ T S; (25)

4.4 Experiment

A validation of the results is needed to guarantee a

good representativity and also a good statistical sig-

niﬁcance. The training set used was of 50% of the

photographs present in the database, so, the other half

was used to compare the propagated location with the

original one. Each training set scenario (by user, by

photo and by event) has 30 replications, as the training

set is built randomly, each replication has a different

training set.

A suggestion is considered correct when its dis-

tance to the original location is smaller than or equal

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

to the error threshold adopted, in the case of this work,

100 meters. Before each technique is tested, the loca-

tions of all the photographs (that are not in the train-

ing set) are temporarily removed and right after that a

propagation attempt is made. With all the suggestions

of all replications stored, we can analyze the proposed

metrics.

Next we present the pseudo code that represents

how the experiment occurred:

for (r = 0; r < 30; r++) {

toRemoveTemp = getPhotosToRemove();

for(Tech t in techniques){

removeTemporarily(toRemoveTemp);

for(p in toRemoveTemp){

suggestedLoc = tec.suggestLocation(p);

salveSuggestion(suggestedLoc);

}

restoreOriginalLocation(toRemoveTemp);

}

As the pseudo code shows, for each replication the

function getPhotosToRemove() will return the pho-

tographs not in the training set after building it. With

the training set built, the variable toRemoveTemp

stores the photographs that will have their location re-

moved temporarily. Three steps must be done to test

each technique considering the training set built for

the current replication:

1. Remove temporarily the geographic location of all

photographs in toRemoveTemp.

2. Compute and save the suggested location of each

photograph in toRemoveTemp.

3. Restore the geographic location of all pho-

tographs in toRemoveTemp.

At the end of the collecting we observed the

trust intervals of each technique with respect to the

four analyzed techniques. Considering a signiﬁcance

level α = 6%, we performed the statistical tests of

Wilcoxon and also the T-Test (Boslaugh, 2012), ac-

cording to the applicability of each one. The tests

were used to validate the results found and presented

in Section 5.

5 DISCUSSION AND RESULTS

In this section, we present the results achieved with

each technique with respect to all the observed met-

rics, and then we discuss these results. The techniques

are represented by acronyms, namely:

• SC: Social Correlation;

• SE: Shared Events;

• TT: Trajectory;

• TC: Temporal Clustering.

There are two moments at with the photographs

are randomly chosen:

1. Which photographs will be in the training set?

2. What is the next photograph without location that

will be analyzed by the technique?

The ﬁrst moment is intended to guarantee that the

training set has not a bias. The second one is intended

to improve the performance of the techniques. When

the photographs without location were ordered ac-

cording to the probability of receiving a correct prop-

agation (that has more photographs near in time with

a location), the performance of the techniques got

worse. For all the observed metrics, there were differ-

ences between 10% and 20% in favor of the random

propagation.

Figure 2: Results Recall.

Several systems adopt different error thresholds.

We know that the lower the error threshold with good

results, the higher the number of systems that would

adopt a certain propagation technique. The use of the

error threshold of 100 meters is favorable to the re-

sults achieved, since it considers the reality of several

systems that need geotag propagation.

This way, we present the result of each metric con-

sidering a random propagation sequence. Figures 2,

NewApproachesforGeographicLocationPropagationinDigitalPhotographCollections

3, 4 and 5 show the results, summarized by the me-

dian, of each observed metric.

Figure 3: Results Accuracy.

Figure 4: Results Precision.

Figure 5: Behavior by user.

Approach by Photograph:

The approach removing the location of the pho-

tographs randomly, without considering the users or

events, shows the same ranking of techniques for all

metrics. The ranking of the techniques is the follow-

ing: SC < TT < SE < TC.

Approach by User:

The removal of location considering the user

shows different rankings for the metrics. We have the

following rankings:

• Precision: TC < TT = SC < SE;

• Recall: TC < TT < SC = SE;

• Accuracy: SC < SE < TT < TC;

The Accuracy metric presented a reverse order,

because the techniques with the worst recall cause

the number of photographs without propagation to be

high, making the accuracy high.

Approach by Event:

The removal of the location considering the events

led to similar results to the approach by users. So, the

rankings are the same for all metrics.

Number of Photographs with Geotag:

At the end of the experiment, we analyzed the

percentage of photographs that had correct location,

comparing with the initial percentage (before the ap-

plication of the techniques). This analysis showed

that the removal approaches considering the users or

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

the events had the same rakings: TC = TT < SE = SC.

Differently from the other approaches, the one using

removal without considering the users or the events

present the same ranking found in the analysis of the

metrics: SC < TT < SE < TC.

Analysis of the Techniques per User:

Finally, we made an analysis on the behavior of

each technique for each user separately. The results

showed that the behavior of each technique varies for

each user. Figure 4 shows eight users for whom the

techniques presented differences among each other

for the Precision metric, considering the random re-

moval that does not use the users or the events. Each

group of four columns represents a user and each col-

umn represents a technique.

6 CONCLUSIONS

In this work, two geotag propagation techniques were

proposed and we also made a comparative analysis

between the two proposed techniques and two exist-

ing ones based on the metrics: precision, recall and

accuracy. With the initial tests, it became evident that

the order of the photographs that must receive prop-

agation must be kept random in order to achieve the

best results. Through the tests carried out throughout

the analysis, we can state that the choice of the cor-

rect propagation technique will depend on the system

type and on the importance of each metric. Consid-

ering a system in which the geographic annotation is

made randomly, making all the users to have at least

50% of the photographs, the technique of choice is

the Temporal Clustering. On the other hand, in a sys-

tem in which the geographic annotation depends on

the user’s proﬁle or on the events that take place, the

Shared Events and the Social Correlation techniques

are the most promising.

We may highlight as future work the proposal of

some way of combining the techniques (with an in-

dividual weighting for each user) in order to improve

the results per user. Besides the possibility of com-

bining the techniques, it can be also considered a lo-

cation made with basis on the way the users describe

the place, instead of latitude and longitude.

REFERENCES

Boslaugh, S. (2012). Statistics in a nutshell. O’Reilly Me-

dia, Inc.

Cooper, M. L. (2011). Clustering geo-tagged photo collec-

tions using dynamic programming. In Proceedings of

the 19th ACM international conference on Multime-

dia, pages 1025–1028. ACM.

Djuknic, G. M. and Richton, R. E. (2001). Geolocation and

assisted gps. Computer, 34(2):123–125.

Figueirêdo, H. F. d., Lacerda, Y. A., Paiva, A. C. d.,

Casanova, M. A., and Baptista, C. d. S. (2012a). Pho-

togeo: a photo digital library with spatial-temporal

support and self-annotation. Multimedia Tools and

Applications, 59(1):279–305.

Figueirêdo, H. F. d., Silva, J. P. R. d., Leite, D. F. B., and

Baptista, C. d. S. (2012b). Detection of photos from

the same event captured by distinct cameras. In Pro-

ceedings of the 18th Brazilian symposium on Multi-

media and the web, pages 51–58. ACM.

Gao, H., Tang, J., and Liu, H. (2012). Mobile location pre-

diction in spatio-temporal context. In Nokia Mobile

Data Challenge 2012 Workshop. p. Dedicated task.

Gong, Y., Li, Y., Jin, D., Su, L., and Zeng, L. (2011). A lo-

cation prediction scheme based on social correlation.

In Vehicular Technology Conference (VTC Spring),

2011 IEEE 73rd, pages 1–5. IEEE.

Hays, J. and Efros, A. A. (2008). Im2gps: estimating geo-

graphic information from a single image. In Computer

Vision and Pattern Recognition, 2008. CVPR 2008.

IEEE Conference on, pages 1–8. IEEE.

Hollenstein, L. and Purves, R. (2013). Exploring place

through user-generated content: Using ﬂickr tags to

describe city cores. Journal of Spatial Information

Science, pages 21–48.

Ivanov, I., Vajda, P., Lee, J.-S., Goldmann, L., and

Ebrahimi, T. (2012). Geotag propagation in social net-

works based on user trust model. Multimedia Tools

and Applications, 56(1):155–177.

Lacerda, Y. A., Figueirêdo, H. F. d., Silva, J. P. R. d., Leite,

D. F. B., Paiva, A. C. d., and Baptista, C. d. S. (2013).

On improving geotag quality in photo collections. In

GEOProcessing 2013, The Fifth International Confer-

ence on Advanced Geographic Information Systems,

Applications, and Services, pages 139–144.

Naaman, M., Song, Y. J., Paepcke, A., and Garcia-

Molina, H. (2004). Automatic organization for dig-

ital photographs with geographic coordinates. In Dig-

ital Libraries, 2004. Proceedings of the 2004 Joint

ACM/IEEE Conference on, pages 53–62. IEEE.

Tsay, K.-E., Wu, Y.-L., Hor, M.-K., and Tang, C.-Y. (2009).

Personal photo organizer based on automated annota-

tion framework. In Intelligent Information Hiding and

Multimedia Signal Processing, 2009. IIH-MSP’09.

Fifth International Conference on, pages 507–510.

IEEE.

Zigkolis, C., Papadopoulos, S., Filippou, G., Kompatsiaris,

Y., and Vakali, A. (2012). Collaborative event anno-

tation in tagged photo collections. Multimedia Tools

and Applications, pages 1–30.

NewApproachesforGeographicLocationPropagationinDigitalPhotographCollections