Meta’s new AI model focuses on making computer vision tools and applications more accessible to the enterprise.
On April 5th, the AI Lab, part of Meta Platforms, introduced the Segment Anything Model (SAM). Part of the Lab’s Segment Anything project, SAM can remove objects from any image with a single click. This project also includes a new segmentation dataset.
According to Meta, SAM can not only identify untrained images, but it can also automatically segment images within seconds. Additionally, the model can estimate segments that have not yet been trained.
Challenges in computer vision
Many in the computer vision market find it difficult to access a set of well-labeled images that can be used to train a model.
Forrester Research analyst Rowan Curran said: Curran says it’s a much better solution than manually tagging things in images or having algorithms identify images with a limited number of classifications.
“The ability to segment any arbitrary fragment of any image and do it in a meaningful way is very powerful,” he added. “This makes it easier to create labeled data for training traditional models, further democratizing many computer vision applications.”
SAM is in line with Meta AI’s purpose of publishing research and advancing the market.
RPA2AI Research analyst Kashyap Kompella said: In addition, Meta needs AI for content moderation, translation and fake news flagging, targeted advertising, image processing and Metaverse strategies, Kompella said.
“The Segment Anything Model is a building block that can be used in multiple applications. [augmented reality] application and fits well with their stated metaverse strategy,” he said.
SAM has many uses, including image and video editing, medical imaging, and scientific research, but a narrower model could be better, Kompella said. For example, AI trained on many medical images may outperform AI trained on large datasets such as SAM.
“For high-risk applications, we still need domain-specific segmentation models,” he continued.
According to Meta, SAM was tested across genders, skin tones, and age-neutral performance.
Models, datasets and papers are now available under the Apache open source license. Users can quickly experiment and build on top of them.