Figure 2: Overview of HIVE. A frozen visual encoder processes a global view and multiple overlap-
ping local views. HVR refines local features through local-to-global-to-local interaction, producing
holistic local representations {FH }R . IVE iteratively evolves an image-specific vocabulary set
Lr r=1
V∗ under visual guidance. Finally, the local predictions derived from {ZV∗}R are refined and
Lr r=1
fused to produce the final segmentation Y˜︁ V∗.