First of all having millions of images is of itself a good way to combat overfitting. But yes, you touch upon a very important point, that is the dataset must

First of all having millions of images is of itself a good way to combat overfitting. But yes, you touch upon a very important point, that is the dataset must be sufficiently diverse. To be fair, humans have this problem as well; if I showed you an octagonal red sign that had the words "GO" inscribed in the middle you may still mistake it for a stop sign at first.