Don't Guess, Just Ask: Resolving Ambiguity in Referring Segmentation via Multi-turn Clarification

Researchers introduced IC-Seg, a segmentation system that asks clarifying questions when a referring prompt is ambiguous.

The framework uses multi-turn dialogue to resolve user intent before making the visual selection.

They also propose Hi-GRPO, a hierarchical training method that improves both clarification quality and segmentation behavior.

A new benchmark, Ambi-RVOS, was released to evaluate ambiguous referring video object segmentation.