ImageNet: a large-scale hierarchical image database for object recognition

Researchers led by Li Fei-Fei introduced ImageNet, a dataset organised around more than 100,000 concept categories and aimed at providing roughly 1,000 human-annotated images per category, on which the ILSVRC challenge was later built.

Background

By the late 2000s, computer vision researchers were spread across many tasks: stereo vision, image retrieval, 3D reconstruction, segmentation. Object categorisation was widely seen as one of the most fundamental capabilities of both human and machine vision, and researchers recognised a need for a high-quality benchmark with clearly established evaluation metrics. Researchers also faced a data problem. Machine learning methods that could generalise well needed to be trained and tested at a scale that existing datasets simply did not reach.

The people who built ImageNet were explicit about both pressures. As they described it, there was “a growing demand for a high quality object categorization benchmark with clearly established evaluation metrics” alongside “a critical need for more data to enable more generalizable machine learning methods.”

What happened

Jia Deng, W. Dong, R. Socher, L.-J. Li, Kai Li and Li Fei-Fei presented ImageNet at CVPR in 2009.

The dataset was organised according to the WordNet hierarchy, a large lexical database that groups words into sets of synonyms called synsets. WordNet contains more than 100,000 synsets, most of them nouns. ImageNet aimed to provide on average 1,000 images per synset, each batch quality-controlled and human-annotated. The ambition was tens of millions of labelled images covering most of the concepts in the hierarchy.

Because no single institution could annotate at that scale alone, the team used crowdsourcing to collect and verify labels. The images themselves were gathered from the web; ImageNet did not own their copyright but compiled an accurate list of web images for each synset.

Built on top of the dataset was the ImageNet Large Scale Visual Recognition Challenge, which ran annually from 2010 onward. The challenge covered object category classification and detection across hundreds of categories and millions of images, and drew in more than fifty institutions over the first five years documented in the 2014 challenge survey (2010–2014). That breadth made it a practical measure of where the field stood and where it was falling short. In 2019, the project received the PAMI Longuet-Higgins Prize, awarded retrospectively to the most impactful paper from CVPR 2009.

Why it mattered

Before ImageNet, researchers recognised a need for a high-quality object categorisation benchmark with clearly established evaluation metrics alongside more data to enable more generalisable machine learning methods. The annual challenge built on the dataset drew in more than fifty institutions over the first five years documented in the 2014 challenge survey (2010–2014) and coincided with key breakthroughs in categorical object recognition and is frequently cited alongside those advances.

Organisations

Stanford University, Princeton University

Sources

• Paper of record. doi.org. Primary source
• ImageNet. image-net.org. Official
• ImageNet Large Scale Visual Recognition Challenge. arxiv.org. Primary source

Cite this page

AI Achievements. (2009). ImageNet: a large-scale hierarchical image database for object recognition. Retrieved 2026-08-29, from https://achievements.ai/milestone/imagenet-a-large-scale-hierarchical-image-database-for

@misc{achievements_imagenet_a_large_scale_hierarchical_image_database_for,
  title  = {ImageNet: a large-scale hierarchical image database for object recognition},
  author = {{AI Achievements}},
  year   = {2009},
  url    = {https://achievements.ai/milestone/imagenet-a-large-scale-hierarchical-image-database-for}
}