Towards a Unified Transformer-based Framework for Scene Graph Generation and Human-object Interaction Detection.