Decrypt’s Art, Fashion, and Entertainment Hub.
With its Metaverse ambitions in shambles, Meta is now looking to AI to drive its next stage of development. One of Meta’s latest projects, the social media giant announced on Wednesday, is called the Segment Anything Model.
Segment Anything helps users identify specific items in an image with a few clicks. While still in demo mode, the company says Segment Anything can already take a photo and individually identify the pixels comprising everything in the picture so that one or more items can be separated from the rest of the image.
Meta coming in hot with SAM
Segment Anything Model (SAM) is a promptable segmentation system. It can "cut out" any object, in any image, with a single click.
Masks can be tracked in videos, enable image editing apps, and even be lifted to 3D
🧵Quick tour and test pic.twitter.com/YC0JSWYy9X
— Nick St. Pierre (@nickfloats) April 5, 2023
“Segmentation—identifying which image pixels belong to an object—is a core task in computer vision and is used in a broad array of applications, from analyzing scientific imagery to editing photos,” Meta wrote in a post announcing the new model.
Meta said creating an accurate segmentation model for specific tasks requires highly specialized work by technical experts with access to AI training infrastructure and large volumes of carefully annotated in-domain data.
“We achieve greater generalization than previous approaches by collecting a new dataset of an unprecedented size.” Ross Girshick, a research scientist at Meta, told Decrypt in an email. “Crucially, in this dataset, we did not restrict the types of objects we annotated.
“Thanks to the scale of the data and its generality, our resulting model shows impressive capabilities to handle types of images that were not seen during training, like ego-centric images, microscopy, or underwater photos,” Girshick added.
Generative artificial intelligence is an A.I. system that generates text, images, or other media in response to prompts. Some of the most prominent examples of this technology are OpenAI’s ChatGPT and the digital art platform Midjourney.
Meta says the Segment Anything AI system was trained on over 11 million images. As Girshick explained, Meta is making Segment Anything available for the research community under a permissive open license, Apache 2.0, that can be accessed through the Segment Anything Github.
"A key aspect of privacy laws is that data collection must be done transparently and with the individual's full consent,” Lyle Solomon, Principal attorney at Oak View Law Group, told Decrypt. “Using AI for facial recognition without express consent raises questions about potential privacy law violations. Additionally, companies should avoid sharing facial data with third parties unless the individual has consented, and any sharing must adhere to privacy law provisions."
Girshick says Segment Anything is in its research phase with no plans to use it in production. Still, there are concerns related to privacy in the potential uses of artificial intelligence.
In February, Meta pivoted from its plans to launch a metaverse to focus on other products, including artificial intelligence, announcing the creation of a new product group focused on generative A.I. This shift occurred after the company laid off over 10,000 workers after ending its Instagram NFT project.
Global leaders, having grown weary of the advance of artificial intelligence, have expressed concerns and open investigations into the technology and what it means for user privacy and safety after the launch of OpenAI’s ChatGPT. Italy has already banned the popular chatbot.
“Many users do not understand how this process works or what the consequences of this can be long term if their face is used to train a machine learning model without their consent,” Kristen Ruby, president of social media and A.I. consultant firm Ruby Media Group, told Decrypt.
“The biggest challenge many companies have is obtaining access to large-scale training data, and there is no better source of training data than what people provide on social media networks,” she said.
Ruby suggests checking if a company has included a machine learning clause that informs users how their data is being used and if they can opt out of future training models. She notes that many companies currently have an opt-in default setting, but that may change to opt-out in the future.
“We have employed various privacy-preserving techniques, such as blurring faces and other personally identifying information (e.g. license plates),” Girshick said. “Users can report offensive content to us by sending an email to email@example.com with the id of the image, and we will remove it from the dataset.”