I was having a look in the docs for image segmentation here: https://dev.fast.ai/tutorial.transfer_learning and saw that the camvid results look pretty bad:
What are those yellow-orange things all across the images?
Yes, you need to run a longer training to have something decent. The training is kept short for testing purposes.
I suggest you’ll do that, this looks pretty shabby on the docs
The main thing is was missing was a smaller batch size. I’ve fixed that in the nb now and ran 8 epochs, so it’ll look fine next time the docs are exported.