Atlas record: https://www.aitimeline.dev/en/events/evt-0006/
Milestone · June 2009
Placed by the contemporary primary publication. The exact event date is not known; its documented interval appears below.
ImageNet provided a large dataset for visual recognition.
A shared benchmark made progress in image recognition easier to measure and compare.
A dataset’s labels and collection choices can also shape what a benchmark measures.
ImageNet · Published June 2009
arXiv; International Journal of Computer Vision (2015) · Published September 2014
The same thesis, embodied in a dataset rather than a paper.
ImageNet sets itself against this collection: here a synset averages 1000 images, 10-25% of them clean; in ImageNet about 99%, at full resolution.
The 2010 paper on crowdworkers lists ImageNet among the work done on this marketplace: labelling millions of images became possible because labour could be ordered through an API.
The ImageNet paper calls LabelMe complementary to itself: about 200 categories and 30,000 images, but with the outlines and locations of objects.
ImageNet in 2009 compares itself with the ESP dataset: the game's labels sit at the "basic level", do not disambiguate word senses and are public for only 60,000 images.
Web scale as a source of data is the precondition later datasets depended on.
The same move ImageNet later repeated: a shared labelled set as the condition for progress.
Both records are about an open contest over a dataset as a way to move a field.
The dataset was turned into an annual competition with a fixed metric.
The AlexNet paper trains and evaluates its network on ImageNet images: direct use of the earlier dataset.
The Inception v3 network was first trained on about 1.28 million images of 1,000 ImageNet classes from 2014 and only then fine-tuned on skin images.
The authors name ImageNet as the model: the dataset is meant to give audio event recognition what ImageNet gave object recognition in images.
A dataset of a billion masks plays the role ImageNet played for classification.