Image classification on ImageNet

The defined task is top-5 classification among one thousand object categories on a fixed benchmark; the human baseline is a single trained annotator, not a population, so the comparison is narrow. Later networks (ResNet, then vision transformers) pushed the error much lower, and the result has held. It does not carry over to open-world perception: distribution shift, adversarial patches and rare categories still trip these systems, and a fixed benchmark is not visual understanding.