Meta releases Segment Anything 2 (SAM 2)
Released with the SA-V dataset of roughly 51,000 videos and 600,000+ masklets, more than four times the video count of the largest prior public segmentation dataset.
- Open weights & ecosystem
- Models & capabilities
- Minor
Meta released Segment Anything Model 2 (SAM 2), extending its 2023 Segment Anything Model from static images to video. Meta described it as the first unified model able to identify which pixels belong to a target object in either an image or a video, tracking that object across frames in real time even as it moved quickly, changed appearance or was briefly obscured — a substantially harder problem than the single-frame segmentation the original SAM handled.
The model and its weights were released openly under an Apache 2.0 licence, alongside the SA-V dataset: roughly 51,000 real-world videos annotated with more than 600,000 “masklets” (spatio-temporal object masks), which Meta said was several times larger in both video count and annotation volume than the largest existing public video-segmentation dataset. Meta pointed to applications spanning video editing, mixed reality, faster annotation of training data for autonomous vehicles, and scientific uses such as analysing marine biology or medical imaging footage.
As with the original Segment Anything release, Meta’s decision to open both the model and the underlying dataset — rather than keeping either proprietary — was aimed at seeding a wider ecosystem of tools built on top of its segmentation research, continuing the company’s strategy of using open releases in computer vision to establish de facto industry standards.