It’s demonstrated the first notebook (under the heading “Deep learning is not just for image classification”). Extending this is suggested as “further research” in the 15_arch_details notebook.
The code is extremely simple because of Fastai’s formulaic construction of a DataBlock and creation of a u-net:
Yes, Icevision is pretty good. Comes with features useful for out-of-the-box parsing of various data formats (object detection, instance segmentation). Using it with fastai is very straightforward (Icevision provides the model and data loaders, you do the rest with fastai).