Accessibility setting

Section: New Results

Linking, navigation and analytics

NLP-driven hyperlink construction in broadcast videos

Participants : Rémi Bois, Vincent Claveau, Guillaume Gravier, Pascale Sébillot, Anca-Roxana Şimon.

In collaboration with Sien Moens, Katholieke Universiteit Leuven, Éric Jamet and Martin Ragot, Univ. Rennes 2.

The hyperlinking sub-task of the MediaEval Search and Hyperlinking task aims at creating hyperlinks between predefined anchor segments, i.e., fragments of videos, and short related video segments, called targets, that have to be automatically extracted from videos of a large collection. Capitalizing on the experience acquired in previous participations [54] , we proposed a two step approach exploiting speech material: Potential target segments are first generated relying on a topic segmentation technique; For each anchor, the best 20 target segments are then selected according to two distinct strategies. The first strategy focuses on the identification of very similar targets using n-grams and named entities, while the second one makes use of an intermediate structure built from topic models, which offers the possibility to control serendipity and to explain the links created [53] .

In 2014, we also initiated the CominLabs project “Linking media in acceptable hypergraphs” dedicated to the creation of explicit and meaningful links between multimedia documents or fragments of documents. Two main issues were adressed: The construction of a corpus, composed of audio and video news, reports and debates, newspapers and blog websites, as well as social networks; A preliminary study of the perceived usefulness of various types of links by end-users.

Analytics in collections of art critics

Participant : Vincent Claveau.

In collaboration with Fabienne Moreau and Nicolas Thély, Univ. Rennes 2.

We aim at exploiting text mining techniques in the service of digital humanities, and more precisely in the field of art criticism. It relies on a collaboration between our team, linguists and art and aesthetics specialists. In preliminary work [56] , we adapted term extraction, named entity recognition and information retrieval techniques to this field to extract multiple linguistic clues from art review articles. Future work will make the most of these clues and clustering approaches to build a navigable and structured collection of the articles.

Data models for navigation

Participant : Laurent Amsaleg.

In collaboration with Björn Þór Jónsson, Grímur Tómasson, Hlynur Sigurþórsson, Áslaug Eríksdóttir and Marta Kristin Larusdottir, School of Computer Science, Reykjavík University.

Digital photo collections—personal, professional, or social—have been growing ever larger, leaving users overwhelmed. It is therefore increasingly important to provide effective browsing tools for photo collections. Learning from the resounding success of multi-dimensional analysis (MDA) in the business intelligence community for On-Line Analytical Processing (OLAP) applications, we propose a multi-dimensional model for media browsing, called $M^3$, that combines MDA concepts with concepts from faceted browsing. We present the data model and describe preliminary evaluations, made using server and client prototypes, which indicate that users find the model useful and easy to use [38] . A photo navigation prototype was demonstrated at the Intl. Conf. on Multimedia Modeling [37] .

Exploiting k-nn graphs for image retrieval

Participants : Laurent Amsaleg, Hervé Jégou, Giorgos Tolias.

We have proposed two techniques exploiting the relationship between the images with a collection. In [29] , we revisit how to exploit the $k$-reciprocal nearest neighbors to produce, for a given query, a neighborhood that improves over the one obtained with the original metric. This strategy is simpler than concurrent prior work, yet it is both effective and less sensitive to parameters. Second, we propose to employ measures defined on sets of shared nearest neighbors in order to re-rank the shortlist. Both methods are simple, yet they significantly improve the accuracy of image search engines on standard benchmarks. We also introduced a query expansion technique [18] for image search that is faster and more precise than the existing ones. The expansion generates an enriched representation which refines the initial local descriptors individually by aggregating those of the database, while new descriptors are produced from the images that are deemed relevant. The technique has two computational advantages over other query expansion techniques. First, the size of the enriched representation is comparable to that of the initial query. Second, the technique is effective even without using any geometry, in which case searching a database comprising 105k images typically takes 80 ms on a desktop machine. Overall, our technique significantly outperforms the visual query expansion state of the art on popular benchmarks.