MMOOZZGG.com

Glossary

What is ImageNet?

In short

ImageNet is an open dataset of millions of labeled photos that hosted an image recognition contest from 2010 and gave a push to deep learning.

History of AI · ImageNet

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine a huge sticker album where each picture has a label saying what is on it: a cat, a bus, an apple. There are millions of such pictures. Scientists collected this album so that computers could learn to recognize things in photos. Every year teams competed to see whose program made fewer mistakes. One year a program built in a new way won, and since then computers recognize pictures much more accurately.

In plain words

To teach a program to recognize a dog in a photo, you need to show it thousands of labeled dog pictures. In the late 2000s there were almost no collections that large. A team of scientists gathered images from the internet and sorted them into a tree of concepts: animal, below it dog, and below that a specific breed. People checked the labels through a service for small online tasks. Every year a contest ran on a part of the dataset with a thousand classes. In 2012 a deep neural network won it by a wide margin, and after that such networks spread fast.

For business

The main lesson of ImageNet for companies: AI quality often depends on the data, not the algorithm. The network that won in 2012 was possible because someone had labeled more than a million images in advance. In your project, this role goes to photos of defects, document scans or call recordings with labels. Labeling costs money and time, and its volume goes into the budget before you choose a model. The second detail: access to ImageNet itself is given for non-commercial research and education, and it does not own the copyright on the images. Models pretrained on it are used widely, but for a product, check the license of the weights and the data. A pilot shows how many of your own examples you need.

For a specialist

Deng, Dong, Socher, Li, Li and Fei-Fei described the dataset in the paper “ImageNet: A Large-Scale Hierarchical Image Database” (CVPR 2009). It follows the WordNet hierarchy of nouns: each synset has a set of images, and the authors aimed for 500–1000 pictures per concept. The first version held 3.2 million images in 5,247 synsets, and the labels were collected through Amazon Mechanical Turk with quality control. Today the project’s website lists 14,197,122 images and 21,841 synsets. From 2010, the ILSVRC contest (ImageNet Large Scale Visual Recognition Challenge) on object classification and detection ran every year on a subset of 1,000 classes. In 2012 the convolutional network of Krizhevsky, Sutskever and Hinton won it with a top-5 error of 15.3% against 26.2% for second place. After that, pretraining on ImageNet and transfer learning became a standard technique in computer vision.

Example

A warehouse wants to recognize damaged boxes in photos. The team takes a network pretrained on ImageNet and fine-tunes it on two thousand of its own images instead of collecting millions.

Related terms

Questions and answers

How many images are in ImageNet? The project’s website lists 14,197,122 images in 21,841 synsets, that is, concepts from the WordNet dictionary. The ILSVRC contest used a subset of 1,000 classes and about 1.2 million training images.
Why does ImageNet matter for the history of AI? In 2012 a deep convolutional network won its image recognition contest by a wide margin. This result convinced researchers and companies to invest in deep learning and graphics cards.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us