Speaker: Matthias Bernhard (ICG)
Stimuli obtained from highly dynamic 3D virtual environments and synchronous eye-tracking data are commonly used by algorithms that strive to correlate gaze to scene objects, a process referred to as Gaze-To-Object Mapping (GTOM). We propose to address this problem with a probabilistic approach using Bayesian inference. The desired result of the inference is a predicted probability density function (PDF) specifying for each object in the scene a probability to be attended by the user. To evaluate the quality of a predicted attention PDF, we present a methodology to assess the information value (i.e., likelihood) in the predictions of di erent approaches that can be used to infer object attention. To this end, we propose an experiment based on a visual search task which allows us to determine the object of attention at a certain point in time under controlled conditions.
We perform this experiment with a wide range of static and dynamic visual scenes to obtain a ground-truth evaluation data set, allowing us to assess GTOM techniques in a set of 30 particularly challenging cases.