SAM (Segment Anything Model): Definition and Examples

SAM, short for Segment Anything Model, is a foundation model for computer vision published by Meta AI Research in April 2023. Its goal is to solve image segmentation in a universal way: rather than being trained to recognize specific object categories, SAM can segment any visual element in any image, including objects it has never seen during training.

The model works through a three-component architecture: an image encoder (based on a Vision Transformer), a prompt encoder that interprets user input (point, box, mask, or text), and a lightweight mask decoder that produces the final segmentation. This architecture enables real-time interaction: the image is encoded once, then the user can quickly iterate by providing different prompts.

SAM was trained on the SA-1B dataset, one of the largest segmentation datasets ever created, containing over one billion masks on 11 million images. This volume of data, combined with the massive pre-training approach, gives SAM its remarkable zero-shot generalization capabilities.

Since its release, SAM has been adapted into several versions (SAM 2 for video, HQ-SAM for better accuracy, FastSAM for speed) and has become a fundamental building block in many computer vision pipelines, from automatic data annotation to photo retouching, robotics, and medical imaging.