Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling (Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022)