unsupervised multi-label ood detection

[sailfish009]

Is it possible to apply this to a multi-label dataset (COCO2017) instead of a single-label dataset?
I'm seeing the positive unlabeled way of learning as more realistic, any feedback would be appreciated.
Positive Label Is All You Need for Multi-Label Classification