Report on the ImageCLEF Experiment: How to Visually Retrieve Images from the St. Andrews Collection using GIFT.
Henning Müller, Antoine Geissbühler, Patrick Ruch · CLEF (Working Notes) · 2004
The imageCLEF task of the Cross Language Evaluation forum has as its main goal the retrieval of images from multi–lingual test collections, or retrieval of images where the query is in a different language than the collection itself. The 2003 imageCLEF task saw no group using the visual information of the images that is inherently language independent. In 2004, this changed and a few groups among them the university hospitals of Geneva are submitting visual runs for the queries. The query topics are definitely defined in a way that makes visual retrieval extremely hard as pure visual similarity plays a marginal role whereas semantics and background knowledge are extremely important, that can only be obtained from textual captions. This article describes the submission of an entirely visual result set to the task. This article will also define possible improvements for visual retrieval systems with the current data. Most important is Section 4 that explains possible ways to make this query task more appealing to visual retrieval research groups, explaining problems of content–based retrieval and what such a task could do to help overcome the present problems. A benchmarking event is needed for visual information retrieval to lower current barriers in retrieval performance. ImageCLEF can help to define such an event and identify areas where visual retrieval might be better than textual and vice–versa. The combination of visual and textual features together is another important field where research is needed.