Whereas the earliest systems used mostly colour, texture and shape features for the retrieval, in recent systems the users are put into a loop to provide the so called user-relevance feedback to improve the retrieval performance and to overcome the semantic gap and subjectivity of human perception. Whereas the CBIR system uses low level features (like colour, shape and texture) for image retrieval, humans interpret image content semantically. The mapping of those semantic descriptions to low level features is very challenging for humans, not to say impossible for complicated images. Furthermore different persons (or the same persons under different circumstances) may perceive the same visual content differently.

As an interesting source for user-relevance feedback an eye tracker can be used, a device that measures the fixations and saccades when the user looks at images. By recording the areas of the images the user looks at when searching for images, we can guide the CBIR system to put more attention on those areas to improve the retrieval performance. As a first step into this new research area, we first have to find out if the recording of eye movements really leads to a better retrieval performance. In an experiment a query image and six other images are provided. The users' task is to click on that image out of the six ones that he or she assumes to be most similar to the query image. Then the six most similar images to the clicked one are calculated and presented to user. The procedure is repeated until the query image is found. This experiment is performed twice: As a Java Web-Interface where we only record on which image the user clicks and also as an eye-tracking experiment where we additionally record the users' eye movements. If the eye tracking experiment leads to faster retrieval performances we will have proven the usefulness of considering visual data for image retrieval.