ImageEval: XRCE Approaches
This document describes XRCE’s participation to Imageval, more specifically to the mixed Text-Image search. After reviewing stateof-the-art methods to exploit the correlations between texts and images in multimedia retrieval, we will examine the single-media search components and describe how we have combined them in the framework of ImagEval. It appeared that, with our current settings and the Imageval corpus, no “early fusion” approach gave significantly better results than a “late fusion” method, so that this paper is mainly dedicated to the latter approach. In this track, exploiting textual information with the Language Modelling approach alone already offered very satisfying performance, much larger than purely visual search. Still, late fusion was able to increase monomedia results by more than 10% (relative), showing the usefulness of combining both types of information, even if the purely visual retrieval component gives relatively poor results.