CWI at the Photo Retrieval Task of ImageCLEF 2009

CWI’s experiments investigate the usefulness of clickthrough data for improving the diversity of image retrieval results. We use the search logs provided to us by Belga to find relevant images; we consider that these correspond to images clicked for queries exactly matching or best matching a topic’s title and cluster titles. To reduce the noise, we also filter these results and only consider those clicked images that are also retrieved by a text-based approach that uses the image captions. To promote diversity, we interleave the images retrieved in the previous step for each of the cluster titles (and also the title). However, given that the clickthrough data available to us cover only a small part of the collection used in the photo retrieval task, our experimental results are inconclusive, although they do provide indications on the reliability of using image search clickthrough data to identify relevant images.