Learning, detection and representation of multi-agent events in videos