visual relationship detection can bridge the gap between computer vision and
natural language for scene understanding of images. Different from pure object
recognition tasks, the relation triplets of subject-predicate-object lie on an
extreme diversity space, such as \textit{person-beh