It makes sense that we'd occasionally find things like this, if human perception does work somewhat similarly to convolutional neural nets. Everyone trains their own feature detectors, and almost everyone ends up with something that works pretty well. (People who don't end up getting called things like "faceblind" or "tonedeaf" when their features don't work for a particular thing.) So it'd make sense that there are edge cases where two common approaches get different results. You could even argue these count as "adversarial examples".
It makes sense that we'd occasionally find things like this, if human perception does work somewhat similarly to convolutional neural nets. Everyone trains their own feature detectors, and almost everyone ends up with something that works pretty well. (People who don't end up getting called things like "faceblind" or "tonedeaf" when their features don't work for a particular thing.) So it'd make sense that there are edge cases where two common approaches get different results. You could even argue these count as "adversarial examples".