What is ImageNet? A huge set of over 14 million labeled photos

A huge set of over 14 million labeled photos, and the yearly contest built around it.

ImageNet is a giant photo dataset released in 2009 by a team led by Fei-Fei Li. It has since grown to more than 14 million photos, each labeled by people with what's in it: "cat", "bicycle" and so on. Labeling that many pictures took the work of a huge number of people recruited online.

From 2010, a yearly contest asked AI systems to sort these photos into 1,000 categories. In 2012 a University of Toronto team using a deep neural network won with a top-5 error of 15.3%, far ahead of second place at 26.2%. That result is often called the start of the deep learning era.

Think of it like this

Think of ImageNet as a gigantic deck of picture flashcards: photo on the front, name on the back. The contest was like a class quiz to see who could name the most cards correctly.

For example

The labels are detailed: even dogs are split into separate breeds.

Common mix-ups

Mix-upImageNet is an AI program that appeared in 2012.
ActuallyImageNet isn't a program from 2012 but a photo collection presented in 2009. Its categories borrow the structure of WordNet, an English word database, branching like a tree — "animal → dog → breed" — so you can work down from broad groups to narrow ones.
Mix-upComputers labelled ImageNet automatically.
ActuallySeveral people looked at each photo separately, and the answers they agreed on were kept. Still, being human work, it contains wrong labels, and inappropriate labels in the people categories were found later and removed.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: Why is the 2012 competition result so often called the start of the deep learning era?
Activity: Draw a bar chart of the figures in record E0035 and this entry (the winner's 15.3% against 26.2% for second place) and, in groups, explain in plain words what "the share of photos with no right answer among the top five guesses" means. Then label twenty photos of classroom objects yourselves and count how many caused arguments (cup or mug?) — and talk about why labelling is hard.