Applications, Deployment Challenges, and Research Gaps of the Segment Anything Model Family: A Review

Abstract

Vision foundation models have developed rapidly, and the Segment Anything Model (SAM) family is a representative line of work in promptable segmentation. This paper reviews the development from the original SAM to SAM 2 and SAM 3, and compares two domain-adapted approaches: SAMMed2D for medical imaging and SUGFW for industrial defect detection. Based on the literature reviewed, we summarize practical challenges at the technical, data, and deployment levels, and identify open questions on lightweight deployment, few-shot adaptation, multimodal prompting, and long-horizon video segmentation. The evidence suggests that the SAM family is a flexible general-purpose framework, but using it reliably in specialized fields still needs domain adaptation, representative validation data, and careful engineering for deployment.