Using Semantic Features for Scene Classification: how Good do they Need to Be?