Images and movies from the web, and others collected for the purpose, have required human labeling for machine learning.

Judge as of 2024-09-29. True: the evidence shows that image and video datasets drawn from the web, and those collected for the purpose, used for training machine learning (for example ImageNet, COCO, video action datasets), have relied on human labelling or annotation. Mostly true: this is generally so, but some notable use of web images and video, such as self-supervised or caption-paired training (for example CLIP using alt-text), needed no explicit human labelling. Treat this as a nuance and not a refutation, since alt-text is human-written. Mixed: a large share of significant image or video ML by 2024 used no human labelling at all, so the claim holds only for part of the field. Mostly false: human labelling was the exception rather than the norm. False: the evidence shows that web images and video used in ML have not required human labelling.

Set before the research began.

1. E1 Microsoft COCO: Common Objects in Context arXiv (ar5iv rendering)

   “the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation.”

2. E2 The Kinetics Human Action Video Dataset arXiv (ar5iv rendering)

   “and then using Amazon Mechanical Turkers (AMT) to decide if the clip contains the action or not.”

3. E3 CLIP: Connecting text and images OpenAI · 5 January 2021

   “The ImageNet dataset, one of the largest efforts in this space, required over 25,000 workers to annotate 14 million images for 22,000 object categories.”

4. E4 CurriculumNet: Weakly Supervised Learning from Large-Scale Web Images ECCV 2018 (Malong Technologies authors)

   “The training data contains 2,439,574 images in total, but without any human annotation.”

The claim holds for the major image and video datasets cited. COCO was built with extensive crowd-worker annotation (E1), Kinetics clips were labelled by Amazon Mechanical Turk workers (E2), and ImageNet needed over 25,000 workers to annotate 14 million images (E3). However, CurriculumNet trained on about 2.4 million web images without any human annotation (E4). This makes the claim generally right, with the notable exceptions the criteria treat as nuance rather than refutation.