Automatic concept detector refinement for large-scale video semantic annotation