# Fast Segment Anything

**Authors:** Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, Jinqiao Wang  
**arXiv:** [2306.12156](https://arxiv.org/abs/2306.12156)  
**Date:** 2023-06-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.12156v1/

## Key points

- SAM model gaining popularity in computer vision for tasks such as image segmentation, captioning, and editing
- High computation costs limit its application in industry scenarios
- Researchers propose an alternative method that achieves comparable performance while significantly reducing computation time
- Approach involves reformulating the task as segments-generation and prompting and using a regular CNN detector with an instance segmentation branch to convert it into an instance segmentation problem
- Training existing instance segmentation methods using only 1/50 of the SA-1B dataset published by SAM authors achieved comparable performance with SAM method at 50 times higher run-time speed
- Researchers provide sufficient experimental results to demonstrate their method's effectiveness and plan to release codes and demos on GitHub
- Work offers a promising solution to reduce computation costs without sacrificing performance in computer vision tasks involving high-resolution inputs.

## Summary

In the field of computer vision, the segment anything model (SAM) has been gaining popularity due to its significant impact on various tasks such as image segmentation, captioning, and editing. However, its high computation costs have limited its application in industry scenarios. To address this issue, a team of researchers proposed an alternative method that can achieve comparable performance while significantly reducing computation time. Their approach involves reformulating the task as segments-generation and prompting and using a regular CNN detector with an instance segmentation branch to convert it into an instance segmentation problem. By training existing instance segmentation methods using only 1/50 of the SA-1B dataset published by SAM authors, they achieved comparable performance with the SAM method at 50 times higher run-time speed. The researchers provide sufficient experimental results to demonstrate their method's effectiveness and plan to release codes and demos on GitHub. Their work offers a promising solution to reduce computation costs without sacrificing performance in computer vision tasks involving high-resolution inputs.