There is currently a trend in media management and the semantic web to develop new media processing methods and knowledge representation techniques to organise and structure media around events. While this increased interest for events as the central aggregator when organising media is supported by strong research in the elds of knowledge representation and computer vision; it is not yet clear how the digital era users use events when sharing their personal media collection. In this paper, we explore how users share photos online and discuss the results of a preliminary automatic processing of the data collected. We show that while media sharing services do not support events as yet, users still share their media around personal events, either by providing explicit spatio-temporal metadata, or by using an event-centric vocabulary.
With the increased availability of digital capturing devices, people now build large personal media collections; what is then done with these media has drastically changed in the last years with the emergence of popular photo sharing services like Flickr1 or Picasa2 and social networks sites such as Facebook3. Understanding how people organise and share their digital collections is key to building better tools that accommodate for the users' needs instead of forcing them to change their mental model to t a xed software work ow. This workow has changed with the introduction of new technologies and ways to interact and share media online; the user's goal is now, not only to archive media for a personal use but also to share them with relevant contacts online. Therefore the issue is not only of organising the user personal media collection for better future search and retrieval, but also the one of organising shared media for visibility to the relevant people and future search and retrieval of these media not just for the author that built the collection but also by these relevant people.
Until recently, the \album" concept was one of the main metaphor for helping users to organise their personal collection, thus staying close to how photo prints 1 http://www.flickr.com 2 http://picasaweb.google.com 3 http://www.facebook.com where organised previously. However, new metaphors of organisation are now emerging to leverage more complex indexing and search. Flickr for instance has introduced a very loose organisation system, focusing on tags to group photos, and with the availability of GPS technology, media management services have introduced the possibility to \geotag" media and to browse and search them with location based services. Some have also introduced search and navigation services based on who is in the photo and when it was taken, using the metadata provided by the camera.
This use of media metadata is moving away from the physical photo album metaphor. However, there is still a semantic gap between the low level metadata, the high level information of who is in the photo or where it was taken and how people group their media for personal archiving and sharing. In fact, most popular media management services still provide the \album" metaphor4 as people still have a need to group media together in ways that are more meaningful to them than just a location or time grouping. Researchers are thus focusing around the event metaphor to combine metadata seems to represent part of the higher level intent of the users when they group their media.
While this metaphor is backed by some early user studies, these were led before the large adoption of social media sharing services and there has been little recent research on how users actually use events digitally to organise and share their media. Discovering if this is the case is not an easy task, and in this paper we discuss an initial study of the sharing behaviour of users on Flickr and Picasa to see if they are currently using events when sharing their photos online. We rst introduce the current work on event representation for media management and the current model of events (Section 2). In the following sections we discuss how we have collected data on Flickr and Picasa (Section 3), how people use event metadata when organising in the album metaphor of these sites (Section 4) and how, even when they do not explicitly use such metadata, they often share media by using an event-centric vocabulary (Section 5). 2
It seems that people take photos to archive important events and share within
their close community. [
This organisation of photos \chronologically by event" eases the search and
retrieval of speci c photos in personal collections as it aligns with the way
memory is structured. According to [
Last.fm and Upcoming.org are services that already try to link media and
event. They do so for public events such as concerts or conferences, but still
do not allow users to share their personal events (e.g. weddings, birthdays). [
[
Practical models for events can be found in the IPTC G2 family of news
exchange standards are provided. EventML5 is one of these standards oriented
at describing public events in a journalistic fashion, although support for media
is limited, and this model is close to Chalfen's idea that media are only used to
support a story. A set of requirements for a base model of events is presented
in [
Event-centric services such as Upcoming.org or Last.fm are focused on public
events such as concerts or conferences. While datasets [
These websites do not provide a way to organise photos around events but provide a way to group photos in albums.These albums can only have a very small amount of metadata and are not presented as events to the user. On Picasa, albums can have a title and a description, and optionally a date and a location; on Flickr, sets can only have a title and a description.
We are interested in seeing how the users describe albums they share on Picasa and Flickr by using the title and description elds. Our hypothesis is that if they share media related to events, they will provide the event metadata in the elds that are available to them and we will nd event references in the titles and descriptions of the albums. We are focusing on these two social sharing sites as they are some of the more popular available at the time of writing; while Facebook is also very popular, it provides very similar features (album based organisation of photos) and does not allow data collection.
We have thus collected a dataset of digital albums shared on Flickr and Picasa. To select users, we use the \explore" pages of each website that feature randomly selected photos; from these photos, we nd a set of random users and collect all public albums that are shared by these users. For each album shared on Flickr we retrieve: (a) the title of the set, (b) the user identi cation, (c) the URL of the set and (d) the number of photos and videos within the set. For each album from Picasa we collect: (a) its URL, (b) the date speci ed for this album, (c) the number of photos, (d) the title, (e) the description and (f) the user identi cation.
Because both websites are international, many entries are not written in
English. In this paper, we are only able to process metadata provided in English and
thus want to lter out the other languages. The perl Lingua::Identify7 module
was used to identify the language of the title and description (when available)
in each album entry. The algorithm provided by this module was trained on the
EuroParl corpus [
We are interested to see if users refer to locations when they describe albums and have thus automatically processed the dataset to nd references to geo7 http://search.cpan.org/~ambs/Lingua-Identify/ graphic locations. The Yahoo! Placemaker8 service is used to perform this task. This freely available geoparsing service can identify place references in unstructured text. While Yahoo! does not provide information on the accuracy of their algorithm, from our manually annotated sample, we found that Placemaker is able to detect if there is at least one location reference in an English title with 81.2% accuracy.
References to dates are also of interest to us as time is a main attribute of an event. To detect such references, we analysed each title with a custom date parsing algorithm that detects full dates but also partial dates (e.g. \Paris'08") and periods (e.g. \40.5 miler in Sespe Wilderness April 2nd - 5th 2010"). On our manually annotated sample, this algorithm performed with 88.1% accuracy.
We have collected 32 168 sets from Flickr and 88 593 albums from Picasa over the month of July 20119. We have kept only English albums, resulting in 5 339 (16.6%) sets from Flickr, and 11 355 (12.8%) albums from Picasa. 4
According to the de nition that we introduced in Section 2, the two main attributes de ning an event are its location and when it happened. Thus, if users are to describe events using albums when sharing their photo, they will probably specify some of these attributes within the available attributes. We found that in the Picasa dataset, only 31% of the albums have a description and thus, in this paper, we focus on the title attribute of the albums as we do not have enough data to draw conclusions from the descriptions. 8 http://developer.yahoo.com/geo/placemaker/ 9 the random crawling collected albums posted between 2006 and 2011. on Picasa is lled automatically with the album creation date if the user does not specify any value explicitly. When there is a date in the title on Picasa, it is often not consistent with the album date eld. It seems that while the users are ready and interested to share their photos around dates, they are not motivated to ll in an extra metadata eld. The reason behind this might be a limitation in Picasa's interface or it can simply be because the users do not see the gain in lling this extra eld.
While the date is an important attribute of events, albums with only a date reference are not always events according to our previous de nition. In fact, from a preliminary manual annotation of the Picasa dataset, we can see that 27.3% of the albums with a date reference but no location reference are not really events. This is because there are catch-all albums for entire years or months, where users put photos of many di erent events in the album (e.g. \Misc. Apr. 2009"). The album is thus only a way to aggregate photos in a time range and not used to represent a speci c event. This happens also when people share photos of their newborn child for milestone periods (e.g. \Jake - 9 months: March").
There are less albums with an explicit location reference, but it still makes for a fourth of the dataset. From the manual annotation, we can see that 78.7% of the albums with only a location reference are actually events. In the same way as with the dates, users use locations for catch-all albums where they put photos of a location they visited multiple times but not for any speci c event (for instance for photos of their home-town).
In these two cases, we can see that the dates and locations are sometimes used only as aggregators for media that could be replaced by automatic metadata based services. However, it seems that the users are not aware of, or willing to use, these services on the studied websites.
96.8% of the albums with a date and a location together were annotated as being events by the manual validators. While these albums represent a small amount of the dataset, we can already see that when space and time are speci ed in the title, the users wanted to share an important event. 5
In the previous section, we have looked at how users might use album attributes to describe explicitly an event location or date. However, there are many events represented on Picasa and Flickr that do not include explicit dates or locations. For instance \Janet and Ian's wedding", \father's day", \Michelle's shower" or \Christmas Eve" are all titles of albums from our dataset that do represent important personal events with no explicit dates or locations. Thus, there might be more albums in this dataset that represent events than the previous section's analysis hinted.
In fact, if we look at the most popular words used in the titles (see Table 2), many of them are references to events (e.g. \party, \wedding", \trip") or time periods, without having explicit dates. Note that, while not shown in Table 2, the most popular words in the vocabulary are years, in fact on Flickr, 11.0% of the vocabulary are numerals while on Picasa 17.5% of the words used are numbers. Table 2 reports gures in per-thousand, while the distribution of the vocabulary follows a very steep long-tail curve, the most popular words still do not cover a large part of the album vocabulary.
While it is easy to see that in the most used words in the dataset there
are concepts representing events, it is not an exhaustive view of the dataset
and it would be interesting to see how many albums refer to events by using
such vocabulary. However, it is not easy to exhaustively list manually the whole
vocabulary that could be used to refer to events. We take a semi-automatic
approach, using WordNet [
We found that around half of the albums (Flickr: 49.4%; Picasa: 50.9%) have
a title with at least one word that represents an event according to WordNet. Of
these albums, only 29.6% have a date or a location (or both) in the title. There
are indeed many albums that describe events without providing either an
explicit date or a location reference (e.g. \Katie's Swiss trip", \Field trip - Farm",
10 Note that because of homography, the same word can appear under di erent
concepts.
\Lily fathers Day"). From a preliminary analysis at these albums, it seems that
many of them either refer to the third important attribute of an event: the
participants; or to relative dates (e.g. \Father's Day", \My Birthday") or locations
(e.g. \Trip Home"). In fact, we can see in Tables 3 and 4 that the day#n#3
and calendar day#n#1 are among the most used concepts. This is in line with
Jain's [
WordNet is a very detailed vocabulary and many terms that it declares as relating to the event concepts might not be used by the users to refer to events. Indeed, there is ambiguity in the vocabulary and we have taken a naive approach where we count the occurrence of all possible words without applying disambiguation. For instance, Michigan#n#3 appears as one of the most popular leaf concepts for Flickr; however, this concept represents a card game called \Michigan" but might have been used by users in their album title as the location. The other top concepts however represent less ambiguously event references.
This con rms Chalfen's conclusions that people like to take photos around
personal events that they then share with a community made of close relations
([
The results we found, while preliminary, show that there is a tendency for users
to share photos around places and location. While this is not a guarantee that
they are sharing albums about speci c personal events, it seems to align with
the previous observations of Zacks et al and Kurby et al [
While most event models discussed in the state of the art (for instance [
In addition, in accordance to Chalfen's [
As was pointed out in [
In this paper we present a preliminary study of a dataset of albums shared on Flickr and Picasa. As we show, while these two services have di erent interfaces and features, users tend to have the same behavior on both sites and we believe that this demonstrates some general intent of the users more than site-speci c behaviors. While this is a raw analysis of the data and a more extensive manual annotation is required, we have found that a signi cant amount of users share media online illustrating personal events, and use time-location metadata to describe them. In fact, we have found that more than a third of the albums shared reference a date in their title and more than a fourth refer explicitly to a location. Users also seem to group their photos around important personal events (e.g. birthdays, wedding, festivities) without always specifying explicitly a location or date.
We are planning future work, in particular in analysing the user needs and habits directly with the users, it shows that they already try to use events when sharing media, even when the applicative work ow does not allow it explicitly. We are also planning to extend this work to study the current use of geo-tagging when sharing media. Therefore providing users with new interfaces and services using the event metaphor should improve their experience and the searchability of the media they share online. 11 http://www.iptc.org The research leading to these results has received funding from the European Community's Seventh Framework Programme (FP7/2007-2013) under grant agreement no248984 GLOCAL and no247758 EternalS.