I'm serious. If you have to rely on mono, single image inputs then yeah ImageNet is going to do better. But it will also mistake every picture of a coke can as

I'm serious. If you have to rely on mono, single image inputs then yeah ImageNet is going to do better. But it will also mistake every picture of a coke can as the real thing. It will be horrifically sensitive to malicious inputs. Much better would be to use 2 calibrated lenses and do 3D reconstruction. Even if you're just doing the reconstruction as a sanity check for a NN to weed out the false positives.