# SAM 2: Segment Anything in Images and Videos

**Authors:** Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, Christoph Feichtenhofer  
**arXiv:** [2408.00714](https://arxiv.org/abs/2408.00714)  
**Date:** 2024-08-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.00714v1/

## Key points

- The paper introduces SAM 2, a foundational model for promptable visual segmentation in images and videos.
- SAM 2 leverages user interaction to enhance the model and dataset collection process, resulting in the largest video segmentation dataset to date.
- Built on a simple transformer architecture with streaming memory capabilities, SAM 2 demonstrates robust performance across various tasks through training on an extensive dataset.
- In video segmentation tasks, SAM 2 showcases improved accuracy while requiring fewer interactions compared to previous methodologies.
- In image segmentation tasks, SAM 2 outperforms its predecessor (SAM) by being more accurate and faster.
- The authors believe that SAM 2 will advance video segmentation and related perception tasks in computer vision.
- A version of the model along with the dataset and an interactive demo are being released at https://ai.meta.com/sam2.

## Summary

The paper "SAM 2: Segment Anything in Images and Videos" introduces the , a foundational model for promptable visual segmentation in both images and videos. Developed by authors Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu,
Ross Girshick,Piotr Dollár,and Christoph Feichtenhofer,the  leverages user interaction to enhance both the model and dataset collection process,resulting in the largest video segmentation dataset to date. Built on a simple transformer architecture with streaming memory capabilities,the  demonstrates robust performance across various tasks through training on their extensive dataset. In video segmentation specifically,it showcases improved accuracy while requiring fewer interactions compared to previous methodologies. Additionally,in image segmentation tasks, outperforms its predecessor - the Segment Anything Model (SAM) - by being more accurate and faster. The authors believe that their innovative approach will advance video segmentation and related perception tasks in computer vision. To facilitate further research and application of ,a version of the model along with the dataset and an interactive demo are being released at https://ai.meta.com/sam2.