On Wednesday, Meta, the parent company of Facebook, unveiled an advanced artificial intelligence model capable of identifying individual objects within images. Alongside this model, the company also released a dataset of image annotations, which it claims is the largest dataset of its type ever created.
The research division of Meta highlighted in a blog post that the Segment Anything Model, abbreviated as SAM, can recognize objects in both images and videos, even if those specific items were not included in its training data.
Users can utilize SAM to select objects by either clicking on them or providing text prompts. For instance, when the prompt “cat” was entered, the model successfully drew boxes around multiple cats featured in a photograph.
Amidst a surge of interest in artificial intelligence following the popularization of OpenAI’s ChatGPT, tech giants have been eager to showcase their latest breakthroughs in this field. This trend has led to increased investments and a competitive push for leadership in AI technology.
While Meta has hinted at various features leveraging generative AI—the same technology that makes tools like ChatGPT notable—it has yet to introduce a finalized product. Proposed tools include one that generates surreal videos based on text instructions and another that creates illustrations for children’s books from written narratives.
Mark Zuckerberg, the CEO of Meta, indicated that enhancing the company’s apps with generative AI “creative aids” will be a key focus for this year.
Meta already employs similar technologies to SAM for tasks such as photo tagging, content moderation, and curating user recommendations on platforms like Facebook and Instagram.
The release of SAM is intended to make this technology more accessible to a wider audience.
The SAM model and its associated dataset will be available for download under a non-commercial license. Users submitting their own images to an accompanying prototype must also agree to use the platform exclusively for research purposes.
© Thomson Reuters 2023
