KAIST adapts open-vocabulary segmentation with prompt disagreement instead of pixel annotations

KAIST · hf · 2026-09-30

KAIST proposes a preference-guided adaptation framework for open-vocabulary semantic segmentation that replaces dense mask supervision with binary preferences mined from 'prompt disagreement' — systematic differences across prompt templates. Localized preference queries are optimized via Region-Localized Preference Optimization with consistency regularization. On MESS, it consistently improves diverse OVSS backbones without pixel annotations and stays robust under noisy preferences. Code released.

Original post →

More from Research

Research channel →