There are many parallels with seismic interpretation here. Many companies/etc keep promising to "revolutionize" interpretation and remove the need for the "tedious" work of a geologist/geophysicist. This is very appealing to management for a wide variety of reasons, so it gets a lot of funding.
What folks miss is that an interpreter _isn't_ just drawing lines / picking reflectors. That's less than 1% of the time spent, if you're doing it right.
Instead, the interpreter's role is to incorporate all of the information from _outside_ the image. E.g. "sure we see this here, but it can't be X because we see Y in this other area", or "there must be a fault in this unimaged area because we see a fold 10 km away".
By definition, you're in a data poor environment. The features you're interested in are almost never what's clearly imaged -- instead, you're predicting what's in that unimaged, "mushy" area over there through fundamental laws of physics like conservation of mass and understanding of the larger regional context. Those are deeply difficult to incorporate in machine learning in practice.
Put a different way, the role is not to come up with a reasonable realization from an image or detect feature X in an image. It's to outline the entire space of physically valid solutions and, most importantly, reject the non-physically valid solutions.
What folks miss is that an interpreter _isn't_ just drawing lines / picking reflectors. That's less than 1% of the time spent, if you're doing it right.
Instead, the interpreter's role is to incorporate all of the information from _outside_ the image. E.g. "sure we see this here, but it can't be X because we see Y in this other area", or "there must be a fault in this unimaged area because we see a fold 10 km away".
By definition, you're in a data poor environment. The features you're interested in are almost never what's clearly imaged -- instead, you're predicting what's in that unimaged, "mushy" area over there through fundamental laws of physics like conservation of mass and understanding of the larger regional context. Those are deeply difficult to incorporate in machine learning in practice.
Put a different way, the role is not to come up with a reasonable realization from an image or detect feature X in an image. It's to outline the entire space of physically valid solutions and, most importantly, reject the non-physically valid solutions.