I wonder if, for this field, the larger solution is scene detection. Actors, what they are looking at, who they are talking to, (and what they are saying).