> However, I think it's a bit odd to treat this type of use case as some sort of AI breakthrough that wasn't possible or wasn't frequently done in the past.
Classic computer vision is an utter PITA - especially when dealing with multiple libraries because everyone insists on using a different bit/byte order, pixel alignment, row/col padding, "where is 0/0 coordinate located and in which directions do the axes grow" and whatnot.
The modern "AI" stuff in contrast can be done by a human in natural language, with no prior experience in coding required.
It's usually the exact opposite for this sort of thing. You can't do this with natural language. Traditional computer vision is well suited to it and works with some tweaks. "Modern" techniques for it require collecting insane amounts of training data for simple things. You can't just throw transfer learning at this because it's a lot different than standard photographs that models are trained on. The old school methods are faster and more reliable for a significant number of problems in the geospatial world. And you still need a lot of deep expertise no matter what.
Classic computer vision is an utter PITA - especially when dealing with multiple libraries because everyone insists on using a different bit/byte order, pixel alignment, row/col padding, "where is 0/0 coordinate located and in which directions do the axes grow" and whatnot.
The modern "AI" stuff in contrast can be done by a human in natural language, with no prior experience in coding required.