SAM2

SAM2 is a unified foundation model for real-time, promptable object segmentation in both images and video.

Meta AI's Segment Anything Model 2 (SAM2) extends the original SAM architecture to the temporal domain, enabling frame-to-frame object tracking with zero-shot generalization. It processes video at 44 frames per second and utilizes a memory bank to maintain mask consistency across occlusions. Trained on the SA-V dataset (51,000 videos and 600,000 masks), it allows users to select objects via clicks, boxes, or masks. This is the go-to engine for high-fidelity rotoscoping, scientific imaging, and autonomous navigation pipelines.

https://ai.meta.com/sam2