Towards optimal bag-of-features for object categorization and semantic video retrieval