You are on page 1of 4

ImageNet

A Large-Scale Hierarchical Image Database

Réalisé par : Medjili mohamed naime & Bounab mounaam


IMAGNET A LARGE-SCALE HIERARCHICAL IMAGE DATABASE [Date de publication]

Introduction :
The digital era's data explosion inspires ImageNet, a large-scale
image ontology with 3.2 million images, leveraging WordNet's
hierarchy and Amazon Mechanical Turk for construction, offering a
vital resource for advanced image applications.
2. Properties of ImageNet :
Structured hierarchically from WordNet, aims for 50 millions
labeled high-resolution images, with the current focus on 12 subtrees,
notably mammal and vehicle categories.
– Scale : ImageNet's scale is evident with 3.2 million annotated
images across 5247 categories, making it the largest clean image
dataset in vision research.
- Hierarchy : ImageNet employs a densely populated semantic
hierarchy akin to WordNet, utilizing interlinked synsets through
relations like "IS-A," resulting in an unmatched density,
exemplified by 147 dog categories not found in other
vision datasets.
- Accuracy : ImageNet aims for high precision throughout the
WordNet hierarchy, exemplified by an average of 99.7%,
acknowledging challenges in distinguishing finer categories
within the hierarchy.
- Diversity : ImageNet prioritizes diversity, quantifying it through
the average image's JPG file size, with the expectation that more
diverse synsets yield blurrier average images, demonstrated in
comparisons with Caltech101.
- TinyImage : 32x32, with 80 million low-resolution images,
contrasts with ImageNet's high-quality synsets (approx. 99%
precision) and full-resolution images 400x350, making ImageNet
more suitable for robust algorithm development and evaluation.
IMAGNET A LARGE-SCALE HIERARCHICAL IMAGE DATABASE [Date de publication]

- ESP Dataset : The ESP dataset, obtained through an online


game, exhibits a biased distribution at the "basic level" and sense
disambiguation challenges, with limited public availability, while
ImageNet offers a more balanced hierarchy distribution and avoids such
issues, providing a larger and publicly accessible dataset.
-LabelMe and Lotus Hill datasets : They complement ImageNet
with detailed object outlines, yet ImageNet's broader scope, larger
category and image counts, sourced from the entire Internet, set it apart.
The Lotus Hill dataset is purchasable.
3. Constructing ImageNet :
ImageNet is an ambitious project. Our goal is to complete the
construction of around 50 million images in the next two years. We
describe here the method we use to construct ImageNet, shedding light
on how properties of Sec. To can be ensured in this process.
3.1. Collecting Candidate Images :
In ImageNet's inception, despite a 10% internet search accuracy, it
aims for 500-1000 clean images per synset. Utilizing WordNet
synonyms and multilingual translations, ImageNet meticulously
compiles a diverse pool of over 10,000 images per synset, laying a
strong foundation for computer vision research.
3.2 Cleaning Candidate Images :
Human evaluators on Amazon Mechanical Turk ensure the
accuracy of the dataset through meticulous verification of each
candidate image.
Users verify synset presence in candidate images, prioritizing
diversity by overlooking occlusions and scene complexities
in labeling tasks.
IMAGNET A LARGE-SCALE HIERARCHICAL IMAGE DATABASE [Date de publication]

To overcome challenges, multiple users independently label


images, requiring a convincing majority for positivity; an algorithm
dynamically adjusts consensus levels based on semantic difficulty,
successfully filtering candidate images and ensuring a high percentage
of cleanliness per synset.

You might also like