Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling (CVPR 2022)