What is Segment Anything Model (SAM)?

Computers often need to work out which pixels of a picture belong to which object, a task called image segmentation, so that an image becomes easier to analyze, for example in medical scans. A long-standing difficulty is that models trained on one dataset, with a fixed list of object categories, generally failed to carry over to another. The Segment Anything Model (SAM) is an AI model built for this task: given a prompt such as a point, a box or a rough mask marking an object, it outputs an outline of that object, and it is designed to transfer to new kinds of images without retraining. The paper introducing it appeared at ICCV 2023, a computer vision conference and describes a project with a model and a dataset of over 1 billion object outlines (masks) on 11 million images.

Written from primary sources and reference pages, and checked against them by an independent AI fact-checker. Drafted with AI assistance.