Object Recognition as Machine Translation: Learning a Lexicon for a Fixed Image Vocabulary