Segment Anything does not need the training to know what is in a photo

Meta AI: As the latest artificial intelligence technologies dealing with automatic object detection teach us, it usually works based on training performed on existing material. This may soon change. Meta is working on a very different approach that could be very useful in specific contexts, which seeks to achieve the same result without needing external help. The social media giant published a post in which it talks about this innovative AI model called Segment Anything, which will be able to detect objects in images and videos even if they are not part of the training sets.

Once perfected, the model will allow us to select elements of a photo or video by clicking on them or simply describing them with a short text. For example, if we choose a cat from a photo or write it in text, the AI model can highlight all the felines in the same or other content. While Segment Anything works autonomously, it can also collaborate with other AI models to further improve accuracy. Both the AI model and the dataset will be downloadable under a non-commercial license, as for now, the project is mainly aimed at researchers and developers who want to collaborate to improve the technology.

A new kind of technology

Right now, Meta uses a similar technology to moderate banned content on social, but in the developer environment, this model is known to be flawed compared to others. Segment Anything has one advantage, handling prompts in real-time, but it could perform better when image processing becomes more demanding. By Meta’s admission, there are far more functional artificial intelligence tools available at the moment, so further development will be needed to refine accuracy and usability further.

One thing is certain: Segment Anything will not immediately find application in operations requiring a certain degree of accuracy, for instance, in robotics or even autonomous driving, but it is much more likely that we will see the first approaches on the web in the near future. With the great success of Bing Chat and Google’s moves in the sector with its chatbot Bard, Meta is feeling the pressure, and it is clear that it does not want to miss the train in its attempt to catch up with the big boys in the industry. We have said it many times, and we will repeat it: 2023 will be the year of artificial intelligence technologies, but the first privacy concerns are already beginning to blossom, and our country was the first to question the data management of one of the most popular LLMs on the market, ChatGPT.