ImageNet
Stanford University and Princeton University (ImageNet project)
The 14-million-photo set whose 2012 contest kicked off the deep-learning boom.
Photos gathered from the web and labeled by crowdworkers against the WordNet noun hierarchy. The best-known slice, ILSVRC-2012, has about 1.28 million training images in 1,000 classes and was the benchmark on which neural networks first beat other methods decisively. Its 'person' categories were later found to include offensive labels and were pruned.
Access
- Cost
- FreeFree after creating an account and agreeing to the terms of access.
- License
- Open, non-commercial onlyTerms of access limit use to non-commercial research and education; image copyrights stay with owners
- Sign-up
- Free account
- How you get it
- Bulk downloadJPEG, XML (boxes)
What's in it
- One row is
- One image labeled with a WordNet synset (and bounding boxes for a subset)
- Covers
- Web images, skewed toward North America and Europe
- Years
- Collected 2007-2010; ILSVRC contests 2010-2017
- Updated
- One-time release
- Size
- About 14 million images in about 21,000 categories; ILSVRC-2012 has about 1.28 million training images
Quality
Historically pivotal and still a standard pretraining and evaluation set, with well-documented label noise and representational problems.
- Person categories contained offensive labels and were removed in 2019
- Several percent of validation labels are wrong or ambiguous
- Images include identifiable people photographed without consent
Connects to
- COCO (Common Objects in Context): Adds object locations and captions for scenes, where ImageNet has one label per image
- Open Images V7: A larger, openly licensed alternative with boxes and masks
- LAION-5B: Web-scale image-text pairs that replaced labeled sets like ImageNet for pretraining
Questions it could answer
- Which classes are most often confused, and are they genuinely ambiguous?
- How geographically diverse are the photos in household-object categories?