Meta has announced its endeavor to democratize Segmentation – the task of identifying which image pixels belong to an object – by setting up the “Segment Anything” (SA) project.
This, said Meta, was a new task, model, & dataset for image segmentation introduced by Meta Research. Using an efficient model in a data collection loop, they built the largest segmentation dataset to date with over 1 billion masks on 11 million licensed and privacy-respecting images.
The tool could be useful not just in photo editing & Content creation, Meta says, but in fields as diverse as augmented reality & Earth studies.
The model is designed & trained to be “promptable”, so it can transfer zero-shot to new image distributions & tasks. Its zero-shot performance is impressive – often competitive with or even superior to prior fully supervised results. The Segment Anything Model (SAM) & corresponding dataset (SA-1B) of 1B masks & 11M images have been released to foster research into foundation models for computer vision.
