Video object segmentation plays a vital role to many robotic tasks, beyond the satisfied
accuracy, quickly adapt to the new scenario with very limited annotations and conduct a
quick inference are also important. In this paper, we are specifically concerned with the task
of fast segmenting all pixels of a target object in all frames, given the annotation mask in the
first frame. Even when such annotation is available, this remains a challenging problem
because of the changing appearance and shape of the object over time. In this paper, we …