SAM 2

The original Segment Anything Model (SAM), released by Meta in 2023, was designed for segmenting objects in static images. SAM 2, released in 2024, extends this capability to video by incorporating a memory mechanism that tracks and segments objects as they move, change, or become occluded across multiple frames.

SAM 2 was developed by researchers at Meta as part of the company's broader artificial intelligence research program in computer vision. Meta released the model with openly available weights and code, continuing the open-research approach taken with the original SAM.

SAM 2 uses a streaming memory architecture that maintains a memory bank of features from previously processed frames, allowing the model to track objects across frames even through occlusions and appearance changes. This memory-based approach enables the model to segment objects throughout a video clip without needing to reprocess the entire sequence from scratch.

Meta released SAM 2 under an open license, making the model weights and code publicly available for research and development through their GitHub repository. The release includes pretrained model variants and a demo application that allows users to test video segmentation on their own footage.