Towards Flexible Visual Relationship Segmentation