The first ILSVRC
The ImageNet dataset became an annual competition: a thousand classes, a million and a half images, one metric for everyone.
Why it matters
Computer vision acquired an annual measure, and it is where the advantage of deep networks became visible two years later.
The format was inherited from the PASCAL VOC challenge but scaled up two orders of magnitude in both images and classes. The first two years were won by methods with hand-designed features, with error around 26 percent. In 2012 AlexNet gave 15 percent, a gap nobody could dispute. Without this annual ruler the 2012 result would have remained one paper among many.