Zhixuan Liu, Jiaqing Fan, Zhengtong Zhu, Fanzhang Li · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3834973
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Referring Video Object Segmentation (RVOS) segments the object described by a natural language expression throughout a video. Despite steady progress, prevailing approaches often rely on excessive task-specific training and cross-modal fusion, yet remain vulnerable to language-vision mismatch: CLIP hallucination, background co-occurrence, and tracking bias cause identity switches and temporal drift under distractors, occlusions, or large motions. We introduce POLAR (Polarized Text Differentials), a training-free, plug-and-play versatile vision-language bridge that provides a robust language-consistency signal to reliably identify the referred target and guide its propagation across frames. Without fine-tuning any backbone, POLAR establishes a precise mask-aware, facet-conditioned language-vision relation by contrasting facet-consistent and facet-inconsistent evidence; incorporates Top M evidence aggregation, logistic calibration, and facet gating with adaptive weighting; and thereby suppresses hallucination, discounts co-occurrence, and mitigates tracking bias to stabilize segmentation across long videos and challenging scenes. POLAR achieves a new state-of-the-art performance on Ref-YouTube-VOS, Ref-DAVIS17 and MeViS, all while remaining simple, efficient, and easy to deploy. Code is available at https://github.com/Louis-zxLiu/POLAR.
ACM International Conference on Multimedia (ACM MM) 2026 · 2 citations
No comments yet — start the discussion below.