=Paper=
{{Paper
|id=None
|storemode=property
|title=EURECOM @ MediaEval 2011 Social Event Detection Task
|pdfUrl=https://ceur-ws.org/Vol-807/Liu_SED_EURECOM_me11wn.pdf
|volume=Vol-807
|dblpUrl=https://dblp.org/rec/conf/mediaeval/LiuHT11
}}
==EURECOM @ MediaEval 2011 Social Event Detection Task==
https://ceur-ws.org/Vol-807/Liu_SED_EURECOM_me11wn.pdf
EURECOM @ MediaEval 2011 Social Event Detection Task
Xueliang Liu Benoit Huet Raphaël Troncy
EURECOM,France EURECOM,France EURECOM,France
xueliang.liu@eurecom.fr benoit.huet@eurecom.fr raphael.troncy@eurecom.fr
ABSTRACT mation online about such scheduled events. For example,
In this paper, we present our approach and results of the Me- FBLeague1 provides the official football games that regis-
diaEval 2011 social event detection (SED) task. We solve the tered in FIFA2 and UEFA3 . From this web site, we obtained
event detection problem in three steps. First, we query all 461 football games that occurred in May 2009, among which
event instances that happened given some condition. Then, 6 took place in Roma and Barcelona. These 6 soccer events
an event identification model is proposed to measure the re- are our prior knowledge for the challenge 1.
lationship between events and photos. Finally, visual prun- For challenge 2, we extract concerts information from event
ing and owner refining heuristics are employed to improve directories such as Last.fm4 , Eventful5 , and Upcoming 6 .
the results. After manual check, only Last.fm contains descriptions of
events held on the given conditions. Last.fm is a popular
music web site that records concert events held in more than
1. INTRODUCTION 190 countries. In addition, Last.fm provides an API for the
The Social Event Detection task at MediaEval 2011 aims developer to build their algorithm based on its data. Using
at detecting social events that occurred during May 2009 its public API, we found 68 events that took place in the
from a dataset composed of images shared on Flickr [2]. The Paradiso and 3 events in Parc del Forum in May 2009.
strategy we investigate is to find the event instances that
occurred during this period of time and then try to match 2.2 Event Identification Model
these event instances with photos from the Flickr dataset. With the prior knowledge of scheduled events description,
We also study how to employ the visual features and “owner” the event detection task changes to a matching problem
metadata from the photos to improve the performance. We where a model can be used to measure the relationship be-
first detail our approach (Section 2) before presenting and tween events and photos. Here, we consider events as some-
discussing our results (Section 3). Finally, we conclude the thing happening in some place during sometime. Therefore,
paper in Section 4. the title, time and location are three key factors that iden-
tify an event. The corresponding photo metadata are text
2. APPROACH DESCRIPTION description, taken time and place. Since these three factors
The challenge of the social event detection task is to find are independent, we can measure the probability of a given
the photo clusters that are relevant to events held on a given photo P to be relevant to an event E by
location during a particular period of time. We tackle this p(P |E) = p(P.text|E.title)p(P.time|E.time)p(P.geo|E.geo) (1)
problem in two steps: first, we attempt to retrieve all of
where: The first item measures the similarity of a photo text
the events that occurred at a given place and time; second,
description with an event title. Since both of them are short
we use the extracted information about these events and at-
and sparse, the most straightforward way to measure them
tempt to match them to the photos metadata in the dataset.
is:
All of the photos that are matched to the same event can
be grouped in one cluster. Besides these two main steps, we |T ext1 ∩ T ext2|
p(T ext1|T ext2) = (2)
also improve the detection results with visual feature and |T ext2|
“owner” metadata. Where the function | · | is the total number of words in a
text vector.
2.1 Prior knowledge acquisition The second item in Equation 1 measures the difference
We known that it is easier and more accurate for the com- between photo taken time and event held time. Here, we
puter to identify specific pattern compared with abstract measure the difference using the Dirac function.
concept. To find concert or soccer events that may be hid-
den in the dataset, we first look for all instances of these two date(T ime2 − T ime1)
p(T ime1|T ime2) = δ( ) (3)
kinds of events held in a given place and time. N
1
Soccer games and concerts are types of favorite activi- http://www.fbleague.com
2
ties in people’s daily life and one can find substantial infor- http://www.www.fifa.com
3
http://www.www.uefa.com
4
http://www.last.fm
5
Copyright is held by the author/owner(s). http://www.eventful.com
6
MediaEval 2010 Workshop, September 1-2, 2011, Pisa, Italy http://upcoming.yahoo.com
Where the function date(·) calculates the number of days • run3 the parameter N in Equation 3 is set to 3 to
for a time span, δ is the Dirac delta function that takes the reduce the impact from erroneous taken time, and the
value 1 when and only when the input parameter is zero, basic Event Identification Model is run.
and N is used for scaling (its value will be discussed in the
Section 3). • run4 Owner Refinement is performed on the results
The third item in Equation 1 measures the distance be- of run3.
tween photo geo tags and event location. The best distance • run5 Visual Pruning and Owner Refinement are per-
measure to use seems the L2 distance between the two loca- formed on the results of run3.
tions. However, an important amount of photos do not have
geo tags and when provided, GPS data in the Flickr dataset A summary of the results is detailed in the Table 1. As
can be inaccurate. Consequently, we just use the city/venue
name to measure the location feature and we use the textual
metric formalized in the Equation 2. Table 1: Event Detection Results
This method finds many photos with a clear description Results Evaluation
and association to events. However, text-based matching Run Events Photos P(%) R(%) F(%) NMI
brings also noise and it can not deal with photos without run 1.1 2 216 97,69 41,21 57,97 0,2420
any text description. We employ visual features to remove run 1.2 2 222 97,75 42,38 59,13 0,2472
the noisy photos and “owner” metadata to find out relevant run 2.1 18 1133 70,79 48,90 57,84 0,4516
photos without text description. run 2.2 18 1172 71,13 50,49 59,06 0,4697
run 2.3 24 1502 70,51 64,57 67,41 0,5987
2.3 Visual Pruning run 2.4 24 1556 70,99 67,01 68,95 0,6171
Visual pruning is employed to remove the noisy photos run 2.5 24 1546 71,00 66,59 68,72 0,6139
from the results of the Event Identification Model [1]. We
assume that the photos that are corresponding to the same
shown in the Table 1, 2 events are found for challenge 1 with
event should be similar visually. The method used here
216 photos identified by the Event Identification Model. 6
is quite straightforward. Given a set of the photo feature
additional photos are found by the “Owner Refinement” ap-
{fi , i ∈ [1, N ]}, the distance between each feature fi and its
proach. For the challenge 2, there are mainly two groups
mean vector m is measure by the L1 distance.
of runs. The first group (run1,run2) used the parameter
di = sum(|fi − m|) (4) N=1, and 18 events are found from the 69 events set previ-
ous obtained. In the second group (run3, run4, run5), 24
Photos are then sorted according to the distance di . The events are found with the parameter N=3. In general, the
bigger the distance and the less similar the photo is with results for the challenge 1 are just average since only 6 foot-
the photo cluster, so we prune the photos with such a large ball games were found as prior knowledge and we suppose
distance. Experimentally, we remove the 5% photos that are that several other games have been missed. For the chal-
far from the center in the visual feature space. lenge 2, the results are more promising and competitive.
2.4 Owner Refinement
4. CONCLUSION
Owner refinement is another way to improve the detection
results [1]. We assume that a person can not attend more In this paper, we propose a framework to detect social
than one event simultaneously. Therefore, all the photos events within a media dataset. In our approach, the events
that have been taken by the same owner during the event instances are retrieved first as prior knowledge, and then, an
duration should be assigned to the same cluster. Using this Event Identification Model is used to measure the similarity
heuristic, it is possible to retrieve photos which do not have of event and photos. In the solution, multi-modality feature
any textual description. such as text, time, visual feature and “owner” metadata are
used.
3. EXPERIMENTS AND RESULTS Acknowledgments
Based on the proposed approach and the events instances
This work is supported by the project AAL-2009-2-049 “Adapt-
obtained previously, we design our runs as follows:
able Ambient Living Assistant” (ALIAS) co-funded by the
Challenge 1 :
European Commission and the French Research Agency (ANR)
• run1 The parameter N in Equation 3 is set to 3, and in the Ambient Assisted Living (AAL) programme.
the basic Event Identification Model is run.
5. REFERENCES
• run2 Owner Refinement is performed on the results [1] X. Liu, R. Troncy, and B. Huet. Finding Media
of run1. Illustrating Events. In 1st ACM International
Conference on Multimedia Retrieval (ICMR’11),
Challenge 2 :
Trento, Italy, 2011.
• run1 the parameter N in Equation 3 is set to 1, and [2] S. Papadopoulos, R. Troncy, V. Mezaris, B. Huet, and
the basic Event Identification Model is run. I. Kompatsiaris. Social Event Detection at MediaEval
2011: Challenges, Dataset and Evaluation. In
• run2 Owner Refinement is performed on the results MediaEval 2011 Workshop, Pisa, Italy, September 1-2
of run1. 2011.