Why does Faster R-CNN ResNet-50 FPN have 91 classes insteand of 80?

As far as I know, COCO only provides annotations for 80 classes. My question is how was Faster R-CNN ResNet-50 FPN form torchvision trained with 91 classes?

I am guessing this is due to the 91 stuff categories the COCO dataset has.
Screenshot from 2020-10-21 11-26-21
Screenshot from the COCO website

Not 100% sure tho…