WORKSHOP REPORT

Multimedia Information Retrieval

New Challenges in Audio Visual Search
Roelof van Zwol
Yahoo! Research, Spain
[email protected]
Mark Sanderson
University of Sheffield, UK
[email protected]

Stefan R
uger
The Open University, UK
[email protected]
Yosi Mass
IBM Research, Israel
[email protected]

Introduction

With the rising popularity of rich media services such as Flickr, YouTube, and Jumpcut, new
challenges in large scale multimedia information retrieval have emerged that not only rely on
meta-data but on content-based information retrieval combined with the collective knowledge of
users and geo-referenced meta-data that is captured during the creation process. For the future,
it is envisioned that multimedia search in mobile environments or on P2P networks will take off
on a large scale.
This workshop followed four previous SIGIR workshops on multimedia information retrieval
(1998, 1999, 2003, 2005), and aimed to address and explore new challenges in multimedia information retrieval by bringing both researchers and practitioners together. We encouraged submission
and participation in this workshop not only from the core information retrieval community but
also from researchers in databases, multimedia and image processing thus cross-fertilizing to information retrieval research.
In response to the call for papers, 16 submissions were received. Each submission was reviewed
by at least three members of the program committee. At this point, we would like to thank the
members of the Program Committee for their contribution to this workshop. Based on their
recommendations the workshop organizers have selected the 7 best scoring articles for publication
and presentation at the workshop. The articles are grouped around three themes: Querying
multimedia, Content-based multimedia retrieval, and Social media mining and meta-data. For
presenting the article, each author had a time-slot allocated of 25 minutes, of which 5 minutes
were reserved for discussion.
Prior to the regular paper sessions, a keynote speech was given by Dr Wei-Ying Ma on the
topic of The Challenges and Opportunities of Mining Billions of Web Images for Search and
Online Applications.
To further stimulate the interactive character of the workshop we decided to allow each participant of the workshop to propose a subject about which he/she could talk for a maximum of

10 minutes during the Speakers Corner. The output of the Speakers Corner was used to organize the direction of the brainstorm session, which allowed for the discussion of new challenges in
Multimedia Information Retrieval.
In the remainder of this report, we summarize the event following the order of the workshop
program. All articles, and the full proceedings are available online at http://www.yr-bcn.es/
events/mir2007-workshop/.

Keynote Speech by Dr Wei-Ying Ma, Microsoft Research, Beijing

The keynote speech, entitled The Challenges and Opportunities of Mining Billions of Web Images for Search and Online Applications, was delivered by Dr Wei-Ying Ma. He is a principal
researcher at Microsoft Research Asia, and is based in Beijing. As a Research Area Manager, he
leads a team of talented, passionate researchers to advance the state-of-the-art in Web search and
data mining.
In his speech, he argued that although content-based image retrieval has been studied for
decades, most commercial search engines still rely on text information to index Web images.
This is because of the many fundamental limitations in current content-based image retrieval
technologies when applied to Web-scale data and a lack of business incentives to rely on image
content for online advertising. In this talk, he discussed the technical hurdles that exist when
we attempt to build systems to analyze and index billions of Web images based on content. He
presented several important Internet applications that have the potential to take off and make
significant impacts if we find ways to overcome existing technical hurdles.

Paper Session: Querying Multimedia

A Semantic Vector Space for Query by Image Example

Joao Magalhaes, Simon Overell, and Stefan R
uger
Abstract: Content-based image retrieval enables the user to search a database for visually similar
images. In these scenarios, the user submits an example that is compared to the images in
the database by their low-level characteristics such as colour, texture and shape. While visual
similarity is essential for a vast number of applications, there are cases where a user needs to
search for semantically similar images. For example, the user might want to find all images
depicting bears on a river. This might be quite difficult using only low-level features, but using
concept detectors for bear and river will produce results that are semantically closer to what
the user requested. Following this idea, this paper studies a novel paradigm: query by semantic
multimedia example. In this setting the users query is processed at a semantic level: a vector
of concept probabilities is inferred for each image and a similarity metric computes the distance
between the concept vector of the query and of the concept vectors of the images in database.
The system is evaluated with a COREL Stock Photo collection.

A Query Language for Multimedia Content

Jonathan Mamou, Yosi Mass, Michal Shmueli-Scheuer and Benjamin Sznajder
Abstract: The growing amount of digital multimedia data available today and the de-facto MPEG7 standard for multimedia content description has lead to the requirement of a query language for
multimedia content. MPEG-7 is expressed in XML and it defines descriptors of the multimedia
content such as audio-visual descriptors, location and time attributes as well as other metadata
such as media author, media URI and more. While most search solutions for multimedia today are
based on text annotations, having the MPEG-7 standard opens an opportunity for real multimedia
content based retrieval. In this paper we propose an IR-style query language for such multimedia
content based retrieval that exploits the XML representation of MPEG-7. The query language is
an extension of the XML Fragments query language that was originally designed as a QueryBy-Example for text-only XML collections. We mainly focus on the unique characteristics of
Multimedia content which needs to support similarity search query (range search and K-nearest
neighbors) and queries on spatio-temporal attributes.

Paper Session: Content-based Multimedia Retrieval

Sharing Personal Experiences while Navigating in Physical Spaces

Rui Jesus, Ricardo Dias, Rute Frias, Arnaldo Abrantes, Nuno Correia
Abstract: Social network popularity has been increasing on the Web over the last years. These web
sites provide an easy way to share digital data between people with common interests. With this
way of exchanging memories of personal experiences, large repositories of data are being created.
Indeed personal media, like photos and videos, are excellent vehicles to exchange experiences,
however to manage large multimedia databases suitable tools are needed. This paper describes a
multimedia retrieval system to access the personal memories shared by the community that visits
a point of interest. The multimedia retrieval system uses multimodal information: visual content,
GPS data and audio information annotated at capture time. This system can be used via a Web
site or when people are visiting the place. The multimedia retrieval system was evaluated using a
mobile interface in a cultural heritage site where the personal media can be shared by visitors and
can be used to guide the visit. Experimental results are presented to illustrate the effectiveness of
the system.
An Optimal Set of Indices for Dynamic Combinations of Metric Spaces
Benjamin Bustos, Nelson Morales
Abstract: A recent trend to improve the effectiveness of similarity queries in multimedia databases
is based on dynamic combinations of metric spaces. The efficiency issue when using these dynamic
combinations is still an open problem, especially in the case of binary weights. Our solution resorts
to the use of a set of indices. We describe a binary linear program that finds the optimal set of
indices given space constraints. Because binary linear programming is NP-hard in general, we
also develop greedy algorithms that find good set of indices quickly. The solutions returned by
the approximation algorithms are very close to the optimal value for the instances where these
can be calculated.

Building Detectors to Support Searches on Combined Semantic Concepts

Robin Aly, Djoerd Hiemstra, Roeland Ordelman
Abstract: Bridging the semantic gap is one of the big challenges in multimedia information retrieval. It exists between the extraction of low-level features of a video and its conceptual contents.
In order to understand the conceptual content of a video a common approach is building concept
detectors. A problem of this approach is that the number of detectors is impossible to determine.
This paper presents a set of 8 methods on how to combine two existing concepts into a new one,
which occurs when both concepts appear at the same time. The scores for each shot of a video
for the combined concept are computed from the output of the underlying detectors. The findings
are evaluated on basis of the output of the 101 detectors including a comparison to the theoretical
possibility to train a classifier on each combined concept. The precision gains are significant,
specially for methods which also consider the chronological surrounding of a shot promising.

Paper Session: Social Media Mining and Meta-data

Learning Taxonomies in Large Image Databases

Lokesh Setia, Hans Burkhardt
Abstract: Growing image collections have created a need for effective retrieval mechanisms. Although content-based image retrieval systems have made huge strides in the last decade, they
often are not sufficient by themselves. Many databases, such as those at Flickr are augmented by
keywords supplied by its users. A big stumbling block however lies in the fact that many keywords are actually similar or occur in common combinations which is not captured by the linear
metadata system employed in the databases. This paper proposes a novel algorithm to learn a
visual taxonomy for an image database, given only a set of labels and a set of extracted feature
vectors for each image. The taxonomy tree could be used to enhance the user search experience
in several ways. Encouraging results are reported with experiments performed on a subset of the
well known Corel Database.
Multimodal Document Alignment: towards a Fully-indexed Multimedia Archive
Dalila Mekhaldi
Abstract: This paper presents a multimodal document alignment framework, which aims to a
cross-indexing of the various documents within multimodal applications (eg, meetings and lectures). These documents might be either present during the event (eg, static documents) or
generated after the event (eg, speech transcript). In the current study, three documents types
have been considered, static documents, speech transcript and slideshows. The framework presented in this paper is an extension of a previous bimodal alignment framework, between static
documents and speech transcript of meetings. Furthermore, several features have been considered and studied in our enhanced alignment framework, such as effect of the TF.IDF metric, the
WordNet thesaurus and the noisy speech on our alignment process. The obtained results in this
enhanced framework prove that the slideshows are a considerable alternative for speech recordings
to temporally index static documents. Moreover, it is shown that the integration of slideshows
has improved and reinforced the bimodal alignment of static documents with speech transcript.

Speakers Corner

In response to the call for contributions for the speakers corner 6 participants stood up and took
the opportunity to share their thoughts in front of the workshop participants.
Jussi Karlgren - Use cases in multi-media information access
Jussi Karlgren presented the CHORUS coordination action and argumented for the informed creation of USE CASES as a focus for service design, system development, and algorithm evaluation.
He stated that evaluation is dear to our hearts, all of ours, really, if we come to a SIGIR conference. We know about relevance assessment, precision, and recall... but how these arguably
bloodless target metrics are translated to not-only-reliable-but-valid-system evaluation exercises
is non-trivial. Especially so for systems in new areas such as multimedia retrieval, where the usage
scenarios may be less obviously patternable to previous tradition.
Mor Naaman - Social Media: Changing the image of multimedia
Mor Naaman pleaded that the advent of media-sharing sites like Flickr and YouTube has drastically increased the volume of community-contributed multimedia resources available on the web.
These collections have a previously unimagined depth and breadth, and have generated new opportunities and new challenges to multimedia research. How do we analyze, understand and
extract patterns from these new collections? How do we use this analysis to improve current
applications and introduce new ones? These questions were discussed and a demo was shown.
Laura Hollink - Bringing semantics into multimedia retrieval
As stated by Laura Hollink, two very different research fields are now slowly growing towards each
other: semantic web and (content-based) image retrieval. This combination has the potential
to improve retrieval performance and open up new ways of searching. Large bodies of structured knowledge are available on the web. Annotations of images, eg, tags, manual annotations,
automatically detected concepts etc, can be linked to these ontologies and take advantage of
its semantics: queries can be answered even though no directly matching annotation was found;
relations between tags can be used for browsing.
Yosi Mass - Search in audio-visual content using P2P information retrieval
As the coordinator of the SAPIR european project, Yosi Mass discussed that today, Web searches
are dominated by search giants such as Google, Yahoo, or MSN that deploy a centralized approach to indexing and utilize text-only indexes enriched by page rank algorithms. Consequently,
while it is possible to search for audio-visual content, the search is limited to associated text and
metadata annotations. Supporting real content-based, audio-visual search requires media-specific
understanding and extremely high CPU utilization, which would not scale in todays centralized
solutions. He claimed that large-scale, distributed P2P architectures will make it possible to search
audio-visual content using the query-by-example paradigm.

Stefan R
uger - Advertisement
Stefan announced that the Knowledge Media Institute of The Open University is hiring talented
researchers to address new challenges in audiovisual search.
Mark Sanderson - External meta-data to enhance multimedia search
Mark Sanderson pointed out that the problems of multimedia search can be mitigated by use of
externally gathered metadata. Images geo-referenced with GPS data for example can lead to a
number of additional improvements. For example, researchers at Berkeley have queried weather
databases to determine the conditions at the time and place where a photograph was taken to
guess quite accurately what weather conditions are pictured in the photograph.

Brainstorm Session

Following the trend that was set during the speakers corner, the brainstorm session was used
to clarify many of the open issues and controversial statements that were brought in by the
participants. The brainstorm was highly interactive and according to the participants, a good
way to wrap up the workshop, which was perceived by the participants as a highly successful
event.

Acknowledgments

We would like to thank the members of the program committee for their contributions to the
material. We also extend our sincere thanks to SIGIR, to the keynote speaker, all the paper
presenters, to the speakers of the Speakers Corner and all 36 participants, who jointly made this
workshop an outstanding workshop.
This workshop was supported by the European Community under the Information Society
Technologies (IST) priority of the 6th Framework Programme for R&D projects: SEMEDIA
(IST-FP6-045032), PHAROS (IST-FP6-045035), Tripod (IST-FP6-045335), and SAPIR (ISTFP6-045128).