Generalizing Universal Image Segmentation to Online Video Segmentation via a Bridging Matcher

Video segmentation and image segmentation share inherent similarities, yet existing video segmentation methods often rely on dedicated architectural redesign. This results in redundant computation and underutilization of the strong capabilities of pre-trained image segmenters. To address this issue, we propose Uni-Matcher, a bridging framework enabling reuse of mature universal image segmentation backbones for video segmentation. Uni-Matcher takes a mask-class-prototype-temporal quadruplet as input, integrates heterogeneous features via a Feature Assembler, and adopts a hybrid hard-soft matching strategy: a Hard Matcher establishes initial cross-frame associations using multi-dimensional similarity metrics, a Soft Modulator dynamically adjusts for feature deviations in complex scenarios, and a Temporal Updater incorporates historical context to ensure robust long-term tracking stability. Extensive experiments on VIPSeg, YouTube-VIS 2019, and VSPW datasets demonstrate that Uni-Matcher achieves state-of-the-art performance on video panoptic, instance, and semantic segmentation tasks when adapting various image backbones, verifying its universality and robustness and providing a straightforward paradigm for bridging image and video segmentation.