These are normal because you’re predicting on out of distribution images, particularly close-up views of objects, which are not part of the MSCOCO training dataset.
