Comprehensive review of segment anything model across multiple domains

Abstract

The Segment Anything Model (SAM) has demonstrated exceptional zero-shot and few-shot generalization capabilities, enabling its effective transfer to novel image processing tasks and supporting diverse downstream applications. As a foundational component in large-scale systems or when integrated with complementary techniques for co-optimization, SAM has significantly advanced the development of image segmentation across multiple domains. However, inherent complexities within domain-specific datasets present critical challenges in boundary refinement, real-time processing, and computational efficiency. This review systematically summarizes current SAM applications in diverse scenarios, evaluates its strengths and limitations, and identifies recurrent challenges in representative datasets. Key optimization strategies for enhancing SAM's performance, generalizability, and efficiency are highlighted. The insights provided aim to guide future dataset construction and interdisciplinary applications, facilitating technological advancements in image segmentation.