TL;DR although all researchers are overfitting the visual dataset, the relative rank of those classfiers are stable. In some sense, what happens about overfitting dataset in computer vision research is not too wrong.