Abstract

In this talk, I will give an overview of our studies of eye movement behaviour in both healthy and clinical populations. Because the synthetic stimuli typically used in the psychophysical laboratory are only poor representations of the highly complex everyday visual input, we use videos of natural scenes as a better proxy for the real world. This greater ecological validity, however, comes at a cost, and I will discuss some of the technical challenges involved in maintaining precise experimental control. In recent work, we also developed new ways to robustly classify dynamic eye movements, which enabled the first quantitative characterization of smooth pursuit eye movements under natural viewing conditions.
Finally, I will present ongoing work on the prediction of eye movements in videos. Using convolutional neural networks that were modified to deal with spatio-temporal input, our approach outperforms all state-of-the-art saliency algorithms on several large-scale data sets.