ImageNet: A Large-Scale Hierarchical Image Database
IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings, 2009 · paper · AI and machine learning
Introduces ImageNet, an image database organized by the WordNet hierarchy and built with Amazon Mechanical Turk. In its current state it has 3.2 million images across 5247 categories, and the authors report it is larger, more diverse and more accurate than existing datasets.
Checked: PDF from image-net.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
1 finding
Substance The Fig. 10 caption and Sec. 4.3 say "22 categories" ("Precision and recall of 22 categories"; "We annotated 100 images in 22 different categories"), but the Fig. 10 x-axis shows 24 labelled categories, each with a precision/recall bar pair.
- p. 7 Figure 10 caption and axis labels
Precision and recall of 22 categories from different
- p. 7 Sec. 4.3 text
We annotated 100 images in 22 different categories
- p. 7 Fig. 10 caption and Sec. 4.3
Precision and recall of 22 categories from different levels of the hierarchy
- p. 7 Fig. 10 x-axis labels
dobbin
Only the authors can say what is right here.
Judge's note: Counted the axis labels in the page image on p7: Texas longhorn through spaceshuttle gives 24 labels, and the bar-pair spacing is consistent with 24 pairs. Only the authors can say whether the correct number is 22 or 24.
Second check: another model agreed
Report a problem with this finding
9 minor (typos and the like; most checked by the judge only)
Reference or label Reference [4] lists its third author only as initials: "A. Artale, B. Magnini, and S. C."
- p. 8 Reference [4]
A. Artale, B. Magnini, and S. C. Wordnet for italian
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the true name of the third author (e.g. C. Strapparava) cannot be checked from these files. The page image on p8 does show bare initials with no surname, which is malformed on its face.
Second check: another model agreed
Report a problem with this finding
Wording The abstract names the third application "automatic object clustering", but the Introduction says "image classification and object localization" and Sec. 4.3 is titled "Automatic Object Localization".
- p. 1 Abstract
image classification and automatic object clustering
- p. 1 Introduction
image classification and object localization
- p. 1 Abstract
object recognition, im-age classification and automatic object clustering
- p. 1 Introduction
object recognition, image classification and object localization
Judge's note: Checked the page 1 images (abstract and Introduction) and the page 7 image (Sec. 4.3 heading). The abstract wording is as quoted. The Introduction and the Sec. 4.3 heading both call the third application localization. Sec. 4.3 does run k-means clustering on the detected boxes (Fig. 12), so the abstract is partly defensible. The application is still named two different ways, so this is a minor wording inconsistency.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 5.1 says "The current ImageNet constitutes ∼ 10% of the WordNet synsets", but the paper's own figures (5247 synsets; "around 80, 000 noun synsets in WordNet") give about 6.6%.
- p. 7 Sec. 5.1
The current ImageNet constitutes
- p. 7 Sec. 5.1
The current ImageNet constitutes ∼ 10% of the Word-Net synsets
- p. 1 Introduction
There are around 80, 000 noun synsets in WordNet
- p. 7 Sec 5.1
The current ImageNet constitutes ∼ 10% of the WordNet synsets.
Only the authors can say what is right here.
Judge's note: Checked the page 7 crop (Sec. 5.1) and the page 1 crops (abstract: 5247 synsets and "the 80,000 synsets of WordNet"; Introduction: "around 80, 000 noun synsets"). 5247/80000 is 6.6%, not ~10%. Against the ~50K target synsets named later in Sec. 5.1, the figure is 10.5%, so the authors may have meant that denominator. I found nothing in the text that states a different denominator, so only the authors can say which was intended.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 5.2 says "Fig. 6 has shown that even the largest datasets today have a strong bias in their coverage of different types of objects", but Fig. 6 only compares the depth distribution of "mammal" labels between ImageNet and ESP. Figure labels also write "Imagenet" where the text uses "ImageNet".
- p. 8 Sec. 5.2, A training resource
Fig. 6 has shown that even the largest datasets today have a strong bias in their coverage of different types of objects
- p. 4 Fig. 6 caption
Comparison of the distribution of “mammal” labels over tree depth levels between ImageNet and ESP game
- p. 2 Figure 3 labels
Imagenet Cattle Subtree
- p. 4 Figure 6 legend
Imagenet
Judge's note: Checked Sec. 5.2 on p8, the Fig. 6 caption and legend on p4, and the Fig. 3 labels on p2. Fig. 6 shows bias over hierarchy depth for one subtree and two datasets, not coverage across object types or across the largest datasets. The capitalisation difference is cosmetic.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 4.2 says "Fig. 9 illustrates the result of our experiment on the mammal subtree", yet interprets the plot with non-mammal examples: "minivan" and "(e.g. vehicles, mammal, artifact, etc.)".
- p. 6 Sec. 4.2
Fig. 9 illustrates the result of our experiment on the mammal subtree.
- p. 6 Sec. 4.2
minivan, polar bear) as opposed to the top of the tree (e.g. vehicles, mammal, artifact, etc.)
Only the authors can say what is right here.
Judge's note: Checked Sec. 4.2 on p6. The experiment is described as run on the mammal subtree only, but minivan, vehicles and artifact are not mammal synsets. Either the experiment scope or the examples are misdescribed; only the authors can say which.
Second check: another model agreed
Report a problem with this finding
Typo Redundant wording: "An average of 99.7% precision is achieved on average."
- p. 2 Accuracy paragraph
An average of 99.7% precision is achieved on average.
- p. 2 Accuracy paragraph
An average of 99.7% preci-sion is achieved on average
Judge's note: The page image on p2 shows the sentence exactly as quoted, with "average" repeated.
Second check: none (judge only)
Report a problem with this finding
Typo The Fig. 5 caption reads "(c) Examples images from Caltech101"; it should read "Example images", as in part (b).
- p. 4 Figure 5 caption
(c) Examples images from Caltech101
- p. 4 Figure 5 caption, part (c)
(c) Examples images from Caltech101 and average images.
Judge's note: The caption on p4 shows "Examples images" in part (c), while part (b) of the same caption uses "Example images".
Second check: none (judge only)
Report a problem with this finding
Typo Missing space before the parenthesis: "our experiments and results(Fig. 8)".
- p. 5 Sec. 4.1
our experiments and results(Fig. 8)
Judge's note: The page image on p5 confirms there is no space before the parenthesis, so this is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Garbled phrase "a broad of range of vision related research"; it should read "a broad range of".
- p. 8 Sec. 5.2
broad of range of vision related research
- p. 8 Section 5.2, first paragraph
a broad of range of vision related research
Judge's note: The page image on p8 shows the same wording, so this is not an extraction artefact.
Second check: none (judge only)