# ImageNet-21K

ImageNet-21K is the full ImageNet dataset with over 21,000 visual categories (synsets) and 14 million hand-annotated images, used for training and evaluating visual object recognition models in machine learning.

ImageNet-21K refers to the complete ImageNet dataset, a large visual database designed for visual object recognition software research. It contains more than 14 million hand-annotated images organized into over 21,000 categories, known as synsets, derived from the WordNet lexical database. The dataset is freely available as annotations of third-party image URLs, though the actual images are not owned by ImageNet. Since 2010, ImageNet has hosted an annual contest, the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), which uses a trimmed subset of 1,000 non-overlapping classes, commonly called ImageNet-1K.

## History

AI researcher Fei-Fei Li began developing the idea for ImageNet in 2006, aiming to expand the data available for training AI algorithms. In 2007, she met Princeton professor Christiane Fellbaum, a creator of WordNet, and subsequently built ImageNet starting from roughly 22,000 nouns in WordNet. Li was also inspired by a 1987 estimate that the average person recognizes about 30,000 kinds of objects.

As an assistant professor at Princeton, Li assembled a team and used Amazon Mechanical Turk for image classification. Labeling started in July 2008 and ended in April 2010, involving 49,000 workers from 167 countries who filtered and labeled over 160 million candidate images. Each of the 14 million images was labeled three times. The original plan called for 10,000 images per category across 40,000 categories, totaling 400 million images, but this was not achieved. The first public presentation was a poster at the 2009 Conference on Computer Vision and Pattern Recognition (CVPR) in Florida.

In 2009, Alex Berg suggested adding object localization, leading to a collaboration with the PASCAL Visual Object Classes contest and the launch of ILSVRC in 2010, which featured 1,000 classes and object localization, compared to PASCAL VOC's 20 classes.

## Significance for Deep Learning

On 30 September 2012, a convolutional neural network (CNN) called AlexNet achieved a top-5 error of 15.3% in the ImageNet 2012 Challenge, more than 10.8 percentage points lower than the runner-up. This success was enabled by using graphics processing units (GPUs) during training, a key factor in the [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) revolution. According to The Economist, "Suddenly people started to pay attention, not just within the AI community but across the technology industry as a whole."

In 2015, AlexNet was outperformed by Microsoft's very deep CNN with over 100 layers, which won the ImageNet 2015 contest with a 3.57% error rate. Andrej Karpathy estimated in 2014 that with concentrated effort, he could reach 5.1% error, and about 10 people from his lab reached roughly 12-13% with less effort. It was estimated that with maximal effort, a human could achieve 2.4% error.

## Dataset

ImageNet crowdsources its annotation process. Image-level annotations indicate the presence or absence of an object class, while object-level annotations provide bounding boxes around visible objects. The dataset uses a variant of the WordNet schema, augmented with 120 dog breed categories for fine-grained classification. In 2012, ImageNet was the world's largest academic user of Mechanical Turk, with workers identifying an average of 50 images per minute.

As of April 30, 2010, the dataset had 21,841 non-empty synsets, 14,197,122 images, 1,034,908 images with bounding box annotations, and 1.2 million images with SIFT features (for 1,000 synsets).

### Categories

Categories were filtered from WordNet concepts, each called a "synset" (synonym set). WordNet 3.0 contains over 100,000 synsets, mostly nouns (80,000+), and ImageNet filtered these to 21,841 countable nouns that can be visually illustrated. Each synset has a WordNet ID (wnid), starting with "n" because only nouns are included; for example, the wnid for "dog, domestic dog, Canis familiaris" is "n02084071". Categories span 9 levels, from level 1 (e.g., "mammal") to level 9 (e.g., "German shepherd").

### Image Format

Images were scraped from online search engines (Google, Picsearch, MSN, Yahoo, Flickr) using synonyms in multiple languages. They are in RGB format with varying resolutions; for example, in the 2012 version, the "fish" category ranges from 4288 x 2848 to 75 x 56 pixels. In [machine-learning](https://www.wikiprompt.org/wiki/machine-learning), images are typically preprocessed to a constant resolution and whitened. In PyTorch, ImageNet images are normalized by dividing pixel values to [0,1], subtracting [0.485, 0.456, 0.406], and dividing by [0.229, 0.224, 0.225], which are the dataset's mean and standard deviations.

### Labels and Annotations

Each image is labeled with exactly one wnid. Dense SIFT features (raw descriptors, quantized codewords, and coordinates) are available for ImageNet-1K, designed for bag-of-visual-words models. Bounding boxes are available for about 3,000 popular synsets, with an average of 150 images per synset. Some images also have attribute annotations.

## Impact and Legacy

ImageNet-21K has been instrumental in advancing [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) and [computer-vision](https://www.wikiprompt.org/wiki/computer-vision) research. The ILSVRC challenge, particularly the 2012 AlexNet result, catalyzed the widespread adoption of [neural-network](https://www.wikiprompt.org/wiki/neural-network) and [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) methods. The dataset's scale and hierarchical structure have enabled training of large models, influencing subsequent developments in [generative-ai](https://www.wikiprompt.org/wiki/generative-ai) and other fields. Although the full 21K dataset is less commonly used than the 1K subset for benchmarking, it remains a valuable resource for pretraining and research on fine-grained classification.

---
Source: https://www.wikiprompt.org/wiki/imagenet-21k
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-12T16:26:32.776742+00:00