Compact representations of videos through dominant and multiple motion estimation