# Can SAM Count Anything? An Empirical Study on SAM Counting

**Authors:** Zhiheng Ma, Xiaopeng Hong, Qinnan Shangguan  
**arXiv:** [2304.10817](https://arxiv.org/abs/2304.10817)  
**Date:** 2023-04-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.10817v1/

## Key points

- Study titled "Can SAM Count Anything? An Empirical Study on SAM Counting"
- Investigates the capabilities of the Segment Anything model (SAM) developed by Meta AI
- SAM known for impressive performance in class-agnostic segmenting
- Specifically explores SAM's potential for few-shot object counting
- Compares SAM's performance with other existing few-shot counting methods
- Findings reveal unsatisfactory performance of SAM without further fine-tuning, especially for small and crowded objects
- Additional adjustments and refinements needed to improve accuracy in challenging scenarios
- Code related to research can be accessed at \url{https://github.com/Vision-Intelligence-and-Robots-Group/count-anything}
- Study highlights limitations of SAM for few-shot object counting and emphasizes need for further development to handle various object types and densities.

## Summary

In a recent study titled "Can SAM Count Anything? An Empirical Study on SAM Counting," authors Zhiheng Ma, Xiaopeng Hong, and Qinnan Shangguan explore the capabilities of the Segment Anything model (SAM) developed by Meta AI. SAM has gained attention for its impressive performance in class-agnostic segmenting. The researchers specifically investigate SAM's potential for few-shot object counting, which involves counting objects of an unseen category by providing a few bounding boxes as examples. The study compares SAM's performance with other existing few-shot counting methods. The findings reveal that SAM's current performance is unsatisfactory without further fine-tuning, especially when it comes to small and crowded objects. This suggests that additional adjustments and refinements are necessary to improve its accuracy in these challenging scenarios. For those interested in exploring the code related to this research, it can be accessed at \url{https://github.com/Vision-Intelligence-and-Robots-Group/count-anything}. Overall, this empirical study sheds light on the limitations of SAM for few-shot object counting and highlights the need for further development in order to enhance its effectiveness in handling various object types and densities.