ImageNet

What ImageNet is

ImageNet is a large image database organized as a hierarchy of object categories and built on the WordNet conceptual model ^[2]. The China Computer Federation describes it as structured and hierarchical, with 21,841 object categories and more than 14 million images ^[2]. China Agricultural University gives a comparable account: more than 14 million image URLs, manually annotated to indicate which objects appear in the pictures, with bounding boxes drawn for at least one million images ^[3].

The word "hierarchical" matters here. The categories are not a flat list of unrelated labels; they are arranged along the same kind of relations that WordNet uses to connect words, so that one category can sit under another as a more specific kind of thing ^[2].

Where it came from

ImageNet was created and released in 2009 at Princeton University by Fei-Fei Li, Kai Li and collaborators ^[2]. The Spring 2010 Release announcement records the size of the collection at that point: more than 11 million images spread over 15,500 visual categories ^[1]. Descriptions of the database as it is usually cited give the larger figures of 21,841 categories and more than 14 million images ^[2].

Browsing and annotation

Users can browse the categories in two ways: through their WordNet structure, and by popularity ranking ^[1]. The second option lets someone start from the categories that people look at most, rather than from the top of a semantic tree.

The 2010 release also carried bounding box annotations for nearly 1,000 popular visual categories, with 200 images annotated per category ^[1]. A bounding box is simply a rectangle drawn around the object in the picture, giving both its position and its extent. At the scale of the whole database, at least one million images carry such boxes ^[3].

What is actually distributed

One detail shapes how ImageNet is used. What the project makes freely available is an annotation database over third-party image URLs; the images themselves do not belong to ImageNet ^[3]. In other words, the resource records which picture at which address contains which object, and where that object sits, while the pictures stay with whoever hosts them. Anyone working with the database therefore depends on those external addresses continuing to resolve.

The 1,000-class subset and ILSVRC

The ILSVRC challenge built on ImageNet used a pruned list of 1,000 non-overlapping classes ^[3]. Going from tens of thousands of categories to a thousand classes that do not overlap each other makes the label space fixed and well defined: every system in the challenge is asked about the same thousand categories, and no category is a mere sub-case of another. This subset is the form of ImageNet most often encountered in descriptions of visual recognition work.

Why it matters

ImageNet's importance comes from three things taken together. First, it connects images to a semantic system: because categories follow WordNet's structure, a label is not an isolated string but a position in a network of meanings ^[2]. Second, the scale is unusual — more than 14 million images and at least a million bounding boxes ^[2][3]. Third, the annotation database is freely available even though the images are not owned by the project ^[3], which makes the labels themselves easy to obtain and reuse.

Through the 1,000-class pruned list used by ILSVRC ^[3], the database also supplied a common set of categories against which object-recognition systems could be posed the same questions. For a field that needs large numbers of labelled pictures, that combination of breadth, structure and open annotation is what made ImageNet a standing reference point.