This article is a summary of a YouTube video "How TikTok dances trained an AI to see" by Vox
TLDR Using TikTok videos can improve computer vision and AI accuracy by providing diverse backgrounds for training, such as creating 3D images from 2D and teaching machines to fill in missing parts in photos.
The use of TikTok videos to train AI to see highlights the potential for unconventional data sources to be used in machine learning.
AI relies heavily on ground truth data sets to learn and perform tasks, making the quality and variety of these data sets crucial for success.
Training an AI to make 3D images from 2D ones with just one viewpoint is like a puzzle, requiring ground truth to know if the model is getting things right.
TikTok's diversity of movements and poses actually helped train an AI to see, contrary to the assumption that it was all about the same dance.
Yasamin's program could look at one part, see the depth, make a prediction how the depth would look later, and check her work, like flipping over a flashcard.
Google researchers created "A Dataset of Frozen People" using mannequin challenge videos to teach an AI model to guess what a scene will look like if a camera moves, which shows how creative solutions can be used to train AI models.
AI technology like Virtual Correspondence can match points in 3D scenes from different angles, potentially revolutionizing industries from film to architecture.
The success of AI is dependent on the quality of data and the humans who provide it, as they are the ones teaching the machines.