Cast your mind back to the internet in 2016. Do you have hazy memories of the Mannequin Challenge? Well, the viral YouTube trend has now been used to train a neural network in understanding 3D scenes.
The context: We are naturally good at interpreting 2D videos as 3D scenes, but machines need to be taught how to do it. It’s a useful skill to have: the ability to reconstruct the depth and arrangement of freely moving objects can help robots maneuver in unfamiliar surroundings. That’s why the challenge has long captivated computer-vision researchers, especially in the context of self-driving cars.
The data: To approach this problem, a team at Google AI turned to an unexpected data set: thousands of YouTube videos of people performing the Mannequin Challenge. (If it happened to pass you by at the time, this involved standing as still as possible while someone moved around you, filming the pose from all angles.) These videos also happen to be a novel source of data for understanding the depth of a 2D image.
The method: The researchers converted 2,000 of the videos into 2D images with high-resolution depth data and used them to train a neural network. It was then able to predict the depth of moving objects in a video at much higher accuracy than was possible with previous state-of-the-art methods. Last week, the researchers were awarded a best paper honorable mention at a major computer vision conference.
Unknowing participants: The researchers also released their data set to support future research, meaning that thousands of people who participated in the Mannequin Challenge will unknowingly continue to contribute to the advancement of computer vision and robotics research. While that may come as an uncomfortable surprise to some, this is the rule in AI research rather than the exception.
Many of the field’s most foundational data sets, including Fei-Fei Li’s ImageNet, which kicked off the deep-learning revolution, were compiled from publicly available data scraped from Twitter, Wikipedia, Flickr, and other sources. The practice is motivated by the immense amount of data required to train deep-learning algorithms and has only been exacerbated in recent years as researchers produce ever bigger models to achieve breakthrough results.
Data privacy: As we have written before, this data-scraping practice is neither obviously good nor bad but calls into question the norms around consent in the industry. As data becomes increasingly commoditized and monetized, technologists should think about whether the way they’re using someone’s data aligns with the spirit of why it was originally generated and shared.
Why Meta’s latest large language model survived only three days online
Galactica was supposed to help scientists. Instead, it mindlessly spat out biased and incorrect nonsense.
DeepMind’s game-playing AI has beaten a 50-year-old record in computer science
The new version of AlphaZero discovered a faster way to do matrix multiplication, a core problem in computing that affects thousands of everyday computer tasks.
A bot that watched 70,000 hours of Minecraft could unlock AI’s next big thing
Online videos are a vast and untapped source of training data—and OpenAI says it has a new way to use it.
The White House just unveiled a new AI Bill of Rights
It's the first big step to hold AI to account.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.