Visual focus of attention in dynamic meeting scenarios

This paper presents our data collection and first evaluations on visual focus of attention during dynamic meeting scenes. We included moving focus targets and unforeseen interruptions in each meeting, by guiding each meeting along a predefined script of events that three participating actors were instructed to follow. Further meeting attendees were not introduced to upcoming actions or the general purpose of the meeting, hence we were able to capture their natural focus changes within this predefined dynamic scenario with an extensive setup of both visual and acoustical sensors throughout our smart room. We present an adaptive approach to estimate visual focus of attention based on head orientation under these unforeseen conditions and show, that our system achieves an overall recognition rate of 59%, compared to 9% less when choosing the best matching focus target directly from the observed head orientation angles