Section: New Results

Competitions and international evaluation campaign

Mediaeval's affect task: Violent scenes detection task

Participants : Guillaume Gravier, Patrick Gros, Cédric Penet.

The project-team participated in the Affect Task of the MediaEval 2012 benchmark, both as part of the organizing team and as competitor [64] , [67] .

Mediaeval's placing task: Geo-localization of videos

Participants : Jonathan Delhumeau, Guillaume Gravier, Hervé Jégou, Michele Trevisiol.

This work was partly done in the context of the Quaero project.

We developed an efficient and effective system to identify the geographic location of videos using a multimodal cascade of techniques exploiting all available sources of information, from user assigned tags to user data and image content. We also proposed a novel hierarchical strategy to exploit tags using information retrieval techniques. A coarse geographic area is first identified before refining the search to find exact geo-coordinates. Area and coordinates are obtained from a vector space representation of the tags using appropriate weighting and normalization [68] .

We participated in the Placing Task of the MediaEval 2012 benchmark, where we ranked first on one of the mandatory runs (no gazeteers, no dictionary).

Mediaeval: Search & hyperlinking

Participants : Guillaume Gravier, Camille Guinaudeau, Pascale Sébillot.

We participated in the Search and Hyperlinking task proposed in the framework of the MediaEval benchmark initiative in 2012. We developed a solution for the hyperlinking subtask in which participants were required to return a ranked list of video segments potentially relevant to the answer provided for an initial query, thus creating links between video segments.

Our solution, based on information retrieval techniques, implements two separate module: The retrieval of relevant videos, followed by the selection of short segments specifically corresponding to the information need. First, the hyperlinking module computes the similarity between a video segment query and the collection of videos and returns a ranked list of relevant videos. We investigated several parameterization and ranking strategies. In the second step, we extract from each video the segment that is the closest, from a meaning point of view, to the video segment query, using topic segmentation methods [42] .

Our system ranked either first or second depending on the evaluation conditions.

ETAPE named entities evaluation campaign

Participant : Christian Raymond.

Christian Raymond participated to the ETAPE Named Entities evaluation campaign. The goal was to propose a system able to tag NE following the new tree-stuctured NE definition given in the Quaero project. The evaluation has been done on manual and 5 automatic transcriptions of french TV and Radio shows produced by 5 different automatic speech recognition systems. The system was ranked first with results far better than those of the other participating systems.

DEFT evaluation campaign participation

Participants : Vincent Claveau, Christian Raymond.

Christian Raymond and Vincent Claveau participated to DEFT . The task proposed was to work on a corpus of scientific papers, by focusing the work on the issue of indexing the scientific papers: identifying the keywords chosen by the authors to index their paper, considering both abstract and whole article. Two tasks were proposed which led them to test two different strategies . For the first task, a list of keywords was provided. Based on that, our first strategy is to consider that as an Information Retrieval problem in which the keywords are the queries that are attributed to the best ranked documents. This approach yielded very good results. For the second task, only the articles were known. For this task, our approach is mainly based on a term extraction system whose results are reordered by a machine learning [27] technique.

Trecvid: Multimedia Indexing task

Participants : Jonathan Delhumeau, Philippe-Henri Gosselin, Hervé Jégou.

This work was partly done in the context of the Quaero project.

Texmex has taking part to the Quaero [50] and IRIM [21] submissions of Trecvid in the Multimedia indexing task, by providing some state-of-the-art image descriptors and collaborating with the LIG to set up the dimensionality reduction tool for high-dimensional vectors. The Quaero Rank was ranked 3rd in the full task (1st amongst European submissions).