Tube-embodied gradient vector flow fields for unsupervised video object plane (VOP) segmentation