Facebook's parent company, Meta, has released a new artificial intelligence (AI) model called the Segment Anything Model (SAM), which can identify objects within images and videos. SAM can even recognize items that it has not encountered during its training. The model is designed and trained to be promptable, meaning it can transfer zero-shot to new image distributions and tasks. SAM's performance capabilities were evaluated on numerous tasks, and its zero-shot performance was found to be impressive, often competitive or even superior to prior fully supervised results. Meta already uses similar technology internally for activities like tagging photos, moderating prohibited content, and recommending posts to users of Facebook and Instagram. The release of SAM will broaden access to this type of technology.
Users can select objects using SAM by clicking on them or writing text prompts. For instance, writing the word "cat" prompted the tool to draw boxes around each of the several cats in a photo. The SAM model and dataset will be available for download under a non-commercial license, and users uploading their images to an accompanying prototype must agree to use it for research purposes only.
The rise of AI has raised ethical and societal concerns about technology that can generate convincing prose or imagery that looks like it's the work of humans. Meta's release of SAM has come amidst the national and global conversation about AI after the release of ChatGPT, an AI chatbot that helped spark a race among tech giants to unveil similar tools.
Meta's research division has released a paper detailing its latest AI model to foster research into foundation models for computer vision. SAM is one of the latest developments in AI, which are rapidly changing the way we interact with technology. The release of SAM is expected to broaden access to technology similar to what Meta already uses internally and help advance research in this field.