Predicting which objects will be named in an image