Segment Anything Model

Segment Anything Model (SAM) is Meta's open-source image segmentation foundation model, widely regarded as SOTA in its task — as Professor Yuan Li put it in an Oasis Capital interview, SAM "pushed segmentation to a certain extreme" and can serve as a base model feeding features to downstream tasks like pose estimation and object-relation recognition, though it still struggles with low-resolution images, small objects, and dense scenes . In July 2024 Meta released SAM 2, the first unified model for real-time, promptable segmentation across both images and video: it improves image accuracy over the original, cuts required interaction time to a third, uses a streaming-memory design for frame-by-frame video processing, and shipped alongside the SA-V dataset of ~51,000 annotated videos and 600,000+ masklets — open source and free to use . Its influence has extended into the research community, with teams like Zhejiang University's ReLER lab building derivatives such as SAM-Track to bring its capabilities to video .