Abstract: Current prevailing Video Object Segmentation (VOS) methods follow the pipeline of extraction-then-matching, which first extracts features on current and reference frames independently, and ...