Abstract: Scene graph generation (SGG) aims to understand sophisticated visual information by detecting triplets of subject, object, and their relationship (predicate). Since the predicate labels are ...