Robust validation of Visual Focus of Attention using adaptive fusion of head and eye gaze patterns

We propose a framework for inferring the focus of attention of a person, utilizing information coming both from head rotation and eye gaze estimation. To this aim, we use fuzzy logic to estimate confidence on the gaze of a person towards a specific point, and results are compared to human annotation. For head pose we propose Bayesian modality fusion of both local and holistic information, while for eye gaze we propose a methodology that calculates eye gaze directionality, removing the influence of head rotation, using a simple camera. For local information, feature positions are used, while holistic information makes use of face region. Holistic information uses Convolutional Neural Networks which have been shown to be immune to small translations and distortions of test data. This is vital for an application in an unpretending environment, where background noise should be expected. The ability of the system to estimate focus of attention towards specific areas, for unknown users, is grounded at the end of the paper.