Improving Video Segmentation via Dynamic Anchor Queries

Modern video segmentation methods adopt feature transitions between anchor and target queries to perform cross-frame object association. The smooth feature transitions between anchor and target queries enable these methods to achieve satisfactory performance when tracking continuously appearing objects. However, the emergence and disappearance of objects interrupt the smooth feature transition, and even widen this feature transition gap between anchor and target queries, which causes these methods to all underperform on newly emerging and disappearing objects that are common in the real world. We introduce Dynamic Anchor Queries (DAQ) to shorten the transition gap by dynamically generating anchor queries based on the features of potential newly emerging and disappearing candidates. Furthermore, we introduce a query-level object Emergence and Disappearance Simulation (EDS) strategy, which unleashes DAQ's potential without any additional cost. Finally, we combine our proposed DAQ and EDS with the previous method, DVIS, to obtain DVIS-DAQ. Extensive experiments demonstrate that DVIS-DAQ achieves a new state-of-the-art (SOTA) performance on five mainstream video segmentation benchmarks.