This is a labeling interface for some of the validation images of the ILSVRC 2014 classification task. It was written by @karpathy to help evaluate human accuracy on ILSVRC 2014, as desribe in blog entry here. After a lot of training, our best annotators get approximately 5.1% Hit-5 error rate (in other words, all 5 guesses are wrong only 5.1% of the time). See if you can beat Google's GoogLeNet ConvNet that achieves 6.7%! For every image, you have 5 guesses out of the 1000 categories below.

Use normal course (normal distribution, default) Use hard course (images GoogLeNet did not get)