In social scenes, people form multiple groups and their behaviors are governed by mutual interactions. Their behaviors are often strongly correlated, e.g., they pay attention to the same point of interest, walk together in a crowd, or make reciprocal gestures during dyadic interactions. In this tutorial, we will review the existing group behavior analyses in computer vision. This will provide a complete overview of the area through three fundamentals:
1) computational representations of social signals in terms of gaze, poses, and gestures (social statics)
2) predictive models that encode the relationship between the social signals (social dynamics)
3) various applications of such social scene understanding to vision, graphics, and robotics
The tutorial will also review social interaction datasets that allow us to empirically rediscover theories of social signals in psychology and sociology such as joint attention, F-formation, and proxemics.