Meta has launched Segment Anything 2 (SA2), an advanced model for real-time promptable object segmentation that extends capabilities from still images to video, achieving state-of-the-art performance.

The model was introduced by CEO Mark Zuckerberg at the SIGGRAPH conference, showcasing significant advancements in video processing efficiency.

SA2 improves segmentation efficiency and accuracy, reducing the interaction time required for human input by three times compared to previous methods.

The model demonstrates high performance across multiple zero-shot video datasets and is approximately six times faster than its predecessor, SAM.

Accompanying SAM 2 is the SA-V dataset, which consists of around 51,000 videos and over 600,000 masklets, significantly enhancing video segmentation training resources.

The GitHub repository for SA2 is available for users to access and experiment with the model, which is open-sourced under the permissive Apache 2.0 license.

To facilitate training, Meta is releasing a large annotated database of 50,000 videos specifically created for SA2, while another 100,000 videos used for training will remain private.

Meta has positioned itself as a leader in open AI, with prior models like LLaMa and Segment Anything contributing to its reputation.

Potential applications for SAM 2 include creating new video effects, enhancing annotation tools for visual data, and aiding in scientific research and medical imaging.

Zuckerberg highlighted the model's potential applications in scientific research, such as studying coral reefs and natural habitats.

Meta invites the AI community to explore SAM 2, utilize the dataset, and contribute to ongoing research in universal video and image segmentation.