To explain what I mean I'll depict the two extremes and something in the middle.
1) Most pragmatic: If you need to just segment a few images for a design project, forget AI. Go into Adobe Photoshop and hand select the outline of the object you need to extract.
2) Middle ground: If you need to build a reasonably accurate app for human aided segmentation of images, use a pre-trained model on a well known architecture.
3) Least pragmatic: If you need to reach unprecedented levels of accuracy on a large volume of images. Do heavily funded research on new and better methods of image segmentation.
So I'm most interested in painting out the spectrum for that middle ground. That is, how much of the wheel needs to be reinvented versus the complexity of the problem.
For example (and this is what lead me here), I need to segment out dogs from several hundred photos that owners have taken. The dog is probably going to be among the main subjects of the photos. Do I need to reinvent the wheel (design an architecture)? Probably not. Do I even need to change the tyres (train my own model)? I'm guessing not. Do I need to code at all? I'm not sure.
While I'm happy to get answers about my use case, it would be awesome if someone could map out the spectrum on my unfinished rainbow.