The research division of tech giant Meta has just presented an artificial intelligence capable of identifying and extracting elements from an image or a video, even if it has not been trained on some of these elements. . An announcement that places the company a little more in the race for AI.
It’s Meta’s turn to shine in the race to experiment with artificial intelligence. The parent company of Facebook, Instagram and WhatsApp has just revealed an artificial intelligence model capable of recognizing elements in an image or video and then identifying them. Still in development, this Segment Anything Model (SAM)was featured in a pre-published scientific article on Meta’s research division blog, Wednesday, April 5.
If similar models already exist, the particularity of SAM would be to be able to find a requested object “ even in cases where he was not confronted with these elements during his learning“, explains the British press agencyReuters.
Recognize, identify, extract
A detail that is not one, since the artificial intelligences responsible for processing images (such as Dall-E, Midjourney or Scribble Diffusion) are first trained on huge databases in order to learn little by little to recognize an apple, a bicycle or a human. Clearly, it takes time for an AI to associate a word with an object.
Segment Anything was trained with the same approach. As demonstrated by a demo available online, the AI has specialized in analyzing a given content and then detecting the outlines of an element, whether by clicking on it or scanning the entire image, in order to extract from the photo.
YouTube linkSubscribe to Frandroid
On its presentation site, Meta also shows that the tool can identify an element requested by text search: type “catin the search bar, and SAM will draw a rectangle around each of the cats in the photo.
Similarly, a demonstration of a video analyzed by this AI testifies to its ability to name objects filmed, even briefly.
Guess what he doesn’t know
However, this learning modellearned a general notion of what objects are”, according to Meta, allowing him to categorize “unfamiliar objects and images without requiring additional training.Similarly, Segment Anything appears to be able to reproduce a 3D image of an object from a simple photo, according to another video from the company’s site.
YouTube linkSubscribe to Frandroid
The AI is therefore also able to guess the missing elements. Various artificial intelligence features that Meta could add to its applications “This year“, accordingReutersciting a recent statement from CEO Mark Zuckerberg on integrating AI as a business priority.
Meta puts on his bib for the race
As the British press agency recalls, Meta already uses “in-house SAM-like technology for activities such as tagging photos, moderating prohibited content, and determining which posts to recommend to Facebook and Instagram users“.
But beyond being useful internally, this new technology is also an opportunity for Meta to maintain its place in the race for AI experimentation, which began with the release of ChatGPT by OpenAI in November 2022. .
Since then, Microsoft and then Google have launched their own service using artificial intelligence (Microsoft New Bing and Google Bard). If the Meta service is more intended for research, it remains a way for the company to put an extra toe on this market without betting too big… as was the case with the metaverse.
Our colleagues from Numerama are launching Watt Else, their newsletter dedicated to the mobility of the future. Sign up here to make sure you receive the next issue!