Eye movement modelling and gaze guidance

Michael Dorr
Institute for Neuro- and Bioinformatics
University of Luebeck

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.

Our strategy is to 1) predict a number of candidate points where a subject
might look, based on the current gaze position and a measure of visual
saliency; 2) increase saliency at the optimal target for the next saccade;
and 3) decrease saliency at all other candidate points. For 1), we employ
machine learning techniques on a large data set of eye movements on
high-resolution natural videos and achieve a very favourable prediction
accuracy of 0.8 ROC score; for 2) and 3), we have implemented a fast
gaze-contingent filtering algorithm based on a spatio-temporal Laplacian
pyramid.

In this talk, we will present results from our first attempts at implementing
the above strategy and discuss the remaining technical and perceptual
challenges.