How many high-level concepts will fill the semantic gap in news video retrieval?