# Evaluating SAM2's Role in Camouflaged Object Detection: From SAM to SAM2

**Authors:** Lv Tang, Bo Li  
**arXiv:** [2407.21596](https://arxiv.org/abs/2407.21596)  
**Date:** 2024-07-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.21596v1/

## Key points

- Evolution of Segment Anything Model (SAM) by Meta AI Research
- Introduction of Segment Anything Model 2 (SAM2) for video and image segmentation tasks
- Advancements in SAM2 over SAM: improvements in applicable domains, segmentation accuracy, and running speed
- Concerning decline in SAM2's ability to discern distinct objects without prompts compared to SAM
- Focus on camouflaged object detection to evaluate SAM2's performance decrease
- Aim to stimulate further research within the SAM model family
- Findings provide insights for improving object segmentation models

## Summary

In their paper titled "Evaluating SAM2's Role in Camouflaged Object Detection: From SAM to SAM2," authors Lv Tang and Bo Li delve into the evolution of the Segment Anything Model (SAM) introduced by Meta AI Research. Initially designed as a generic object segmentation model, SAM quickly gained traction within the academic community. To enhance its capabilities for video applications, Meta developed Segment Anything Model 2 (SAM2), a unified model proficient in both video and image segmentation tasks. SAM2 exhibits significant advancements over its predecessor, boasting improvements in applicable domains, promptable segmentation accuracy, and running speed. However, the study highlights a concerning decline in SAM2's ability to discern distinct objects within images without prompts when operating in auto mode compared to SAM. To evaluate this performance decrease, the authors focus on the challenging task of camouflaged object detection. By scrutinizing SAM2's efficacy in identifying concealed objects, Tang and Li aim to stimulate further research and exploration within the SAM model family. Their findings shed light on potential areas for improvement and inspire researchers to delve deeper into refining object segmentation models. The results of their investigation can be accessed through the provided link: \url{https://github.com/luckybird1994/SAMCOD}.