Candidate Talk: A Discriminative Kernel-based Model to Rank Images from Text Queries

February 12, 2008
David Grangier | IDIAP Research Institute, Switzerland

This presentation introduces a discriminative model for the retrieval of pictures from text queries. The core idea of this approach is to minimize a loss directly related to the retrieval performance of the model. For that purpose, we formalize the retrieval task as a ranking problem, and introduce a learning procedure optimizing a loss function related to the ranking performance. This strategy hence addresses the retrieval problem directly and does not rely on an intermediate image annotation task, which contrasts with previous research. Moreover, our learning procedure builds upon recent work on the online learning of kernel- based classifiers. This yields an efficient, scalable algorithm, which can benefit from recent kernels developed for image comparison.
The experiments performed over stock photography data show the advantage of our discriminative ranking approach over state-of-the- art alternatives (e.g. our model yields 26.3% average precision over the Corel dataset, which should be compared to 22.0%, for the best alternative model evaluated). Further analysis of the results shows that our model is especially advantageous over difficult queries such as queries with few relevant pictures or multiple-word queries.

Speaker Details

David Grangier received a master degree from the Eurecom Institute, and from the Ecole Nationale des Telecommunications de Bretagne (2003). He is currently completing his doctoral program at the Ecole Polytechnique Federale de Lausanne, Switzerland. Since 2003, he also works as a research assistant at the IDIAP Research Institute, Switzerland. David Grangier’s research focuses on various aspects of statistical machine learning, including learning to rank, learning distance measures and online learning. He has also strong interrests in pattern recognition algorithms for speech and vision problems.