Jia Deng · 2009 IEEE Conference on Computer Vision and Pattern Recognition 2009 · Dataset creation and validation study · n=3.2 million images (5,247 synsets)

ImageNet: A large-scale hierarchical image database

Cited 62989 times in the scientific literature.

Level 5 - mechanism / opinion, no new human data

Level 5 by design analogy; technical computer science dataset development and validation without human clinical trial design.

OpenAlex W2108598243 · doi:10.1109/cvpr.2009.5206848 · record verified 2026-08-26

What was done

The authors developed ImageNet, a large-scale hierarchical image database organized according to the WordNet semantic hierarchy. Images were collected from the internet and verified using crowdsourced human annotation via Amazon Mechanical Turk. The authors analyzed the dataset's scale, accuracy, and diversity, and demonstrated utility across object recognition, image classification, and automatic object clustering tasks.

What was found

At the time of publication, ImageNet comprised 12 subtrees spanning 5,247 WordNet synsets and 3.2 million verified full-resolution images, working toward a broader target of populating ~80,000 synsets with 500–1,000 images each. Specific benchmark performance metrics for the demonstration tasks were not reported in the abstract.

Why it matters

ImageNet provided a massive increase in scale, semantic organization, and annotation accuracy compared to existing computer vision benchmarks, establishing a foundation for modern visual recognition models.

Limits

The abstract reports only an intermediate state (5,247 of ~80,000 target synsets) and does not quantify classification accuracy rates, annotation error margins, or specific benchmark performance gains over prior datasets.

Cited by