Michael Dorr, Eleonora Vig, Karl Gegenfurtner, and Erhardt Barth

A major limitation of our visual communication capabilities is that we can attend to only a very limited number of features and events at any one time. Therefore, we are developing gaze-contingent systems that guide the user's gaze by changing the saliency distribution in real time.

We present subjects with high-resolution videos of natural scenes while recording their eye movements. Based on the current gaze position and a measure of visual saliency, we first predict candidate locations that are likely to be attended in the near future. We then decrease saliency (a simple modification would be e.g. a reduction of local contrast) at all such locations but one where saliency is increased.

We currently derive the transformations to decrease and increase saliency using Machine Learning techniques. From a large data set of eye movements on dynamic natural scenes, we obtain information on the structural differences of attended and unattended movie regions.

We will present results from our first attempts at implementing the above strategy.

Poster in pdf format.