Hey bigThrum,
As this is an interest of mine also, I’m glad you brought it up. Perhaps you’ve already seen it, but i’ll refer you to the thread from 2017 on this subject where I shared my experience tackling this problem.
A nice sanity check that we added was utilizing instance-based learning with histograms of the images, which was relevant for my use-case (images from pretty much the same angle of the same object)