ImageNet

ImageNet is a foundational large-scale image dataset that organizes photographs according to the WordNet noun hierarchy. The full database contains more than 14 million human-verified images labeled across tens of thousands of concepts (synsets), collected from web image search and annotated at scale via crowdsourcing on Amazon Mechanical Turk. It was created by researchers at Princeton and Stanford, led by Fei-Fei Li, and first presented in 2009.

What’s in it

Each synset groups hundreds to thousands of example images of a concept, from broad categories to fine-grained species and object types. The most influential slice is the ILSVRC-2012 subset: 1,000 mutually exclusive classes with about 1.28 million training images plus 50,000 validation and 100,000 test images, which became the field’s default classification benchmark and is often referred to simply as ImageNet-1k. Beyond it, researchers pretrain on larger cuts such as ImageNet-21k, while diagnostic variants like ImageNet-V2, ImageNet-C, and ImageNet-R probe robustness and generalization.

Tasks, license & access

ImageNet is used chiefly for image classification and as a pretraining source whose learned features transfer broadly to detection, segmentation, and other vision tasks; the 2012 ILSVRC results helped ignite the deep-learning era. Access is free for non-commercial research after agreeing to the terms of use; images are provided as URLs and downloads with annotations, and users are responsible for image copyright.

• Modality: labeled natural images
• Scale: 14M+ images, ~21K synsets; ILSVRC subset 1,000 classes, ~1.28M train
• Tasks: classification, transfer-learning pretraining
• License: non-commercial research terms of use
• Creator: Princeton/Stanford (Fei-Fei Li et al.)

As the benchmark that catalyzed modern deep learning, ImageNet remains a reference point for pretraining and evaluating visual representations.