Coarse-to-Fine Active Segmentation of Interactable Parts in Real Scene Images