Video Audience Analysis using Bayesian Networks and Face Demographics

In this paper, we propose an approach to study and to model audience attention to videos within digital signage scenarios. An experimental setup was conceived to simultaneously display videos of various categories and to capture videos from the audience and surrounding environment. Face analysis via a deep neural network is performed to estimate gender and age groups. In the proposed approach, a Bayesian Network is built to model possible relationships between the audience's age, gender and face size (which is indicative of the distance to the display) and the video content types. A publicly available video dataset of 152 videos was created for displaying purposes. An experimental evaluation indicated varying degrees of attention to different videos, depending on age and gender. The area under the ROC curve of the built Bayesian Network was 0.82. The proposed approach allows to better understand the possible relationships between audience demographics and video contents, which may, in turn, be useful for displaying the most appropriate content to a particular audience, help with the automatic insertion of ads (based on audience categories), among other applications.