Comment by cess11

That looks like a pretty good starting point, thanks. I've been dabbling in vision models but need a much higher degree of accuracy than they seem able to provide, opting instead for more traditional techniques and handling errors manually.