MinVIS: A Minimal Video Instance Segmentation Framework without Video-based Training

By only training a query-based image instance segmentation model, MinVIS outperforms the previous best result on the challenging Occluded VIS dataset by over 10% AP .