AMR 2011- PROGRAM [.pdf document]


1. Multimedia Retrieval Evaluation - Dr. Martha Larson

Short description: Multimedia retrieval has long benefited from benchmarking initiatives that offer tasks to the research community. A task consists of a problem definition, a data set and ground truth against which solutions to the task are evaluated. Benchmarking benefits the research community by bringing scientists together in a productive mix of cooperation and competition that fights fragmentation and promotes efficient use of resources. This talk discusses multimedia retrieval benchmarks and the contribution that they make to multimedia research. In particular, the MediaEval benchmark ( is introduced and typical tasks are described. MediaEval tasks emphasize social and language aspects to multimedia retrieval and are designed based on specific use scenarios. The tasks covered include: geo-coordinate prediction for video, violent scenes detection for movies and low-resource spoken term detection for the Spoken Web.

2. Growing lifelong musical soulmates: a playground for research on (inter)active, adaptive music-aware systems - Prof. P. Herrera

Short description: In this presentation I will examine different achievements that, by means of their convergence, make possible to transcend traditional and limited concepts of music information retrieval. As MIR shifts its focus from “information” to “interaction” we are aimed towards applications that: i) actively learn from behaviour patterns, sensed data, multimodal sources, and from explicit interactive training episodes, ii) take care for the overall musical wellness of the user (by providing news about preferred artists, recommending unknown but likeable music, facilitating musical epiphanies, etc.), and iii) accompany users beyond the usual life-cycle of their physical devices and become supportive tools in critical stages of their development (e.g, adolescence or senescence), demanding and providing attention and care in order to grow and evolve their music awareness as users do.

3. Large-scale multimedia retrieval: distributing multimodal interactive learning - Prof. S. Marchand-Maillet (Viper group - University of Geneva)

Short description: Multimedia retrieval could largely be achieved thanks to user feedback interpretation. Machine learning strategies such as Boosting can be designed to help in performing information fusion to gather and exploit every piece of knowledge the user is providing to the system.
As a natural extension of these working mechanisms, we have more recently attacked the problem of scalability in multimedia retrieval. We look at how computation and information access may be scheduled over a network of computers with distributed storage to preserve the usability and usefulness of our tools when applied over large colections of items bearing multimodal information.
Here, we therefore emphasise and review both aspects of retrieval performance and robustness against the increase in the scale of the dataset and in the complexity of the data. We summarise our achievements that have already resulted into concrete developments.