Scholarlink

<p>Gaussian splatting (GS) enables high-quality real-time novel view synthesis by representing scenes with explicit Gaussian primitives and leveraging efficient differentiable rendering. It has expanded from static reconstruction to dynamic modeling, editability, and semantic understanding, showing value in simultaneous localization and mapping (SLAM), autonomous driving, digital humans, and medical imaging. To address the rapid growth of the literature, increasingly diverse technical routes, and inconsistent evaluation protocols, this survey first reviews the representation and rendering fundamentals of Gaussian splatting and summarizes commonly used datasets and metrics to establish a comparative baseline. It then structures key methods around two threads, covering efficiency and quality optimization for static scenes and spatiotemporal modeling for dynamic scenes. The survey further summarizes the needs for AI-driven semantic understanding and editability, artificial intelligence generated content (AIGC), and contextualized representation and reasoning in embodied intelligent interaction. Finally, it discusses capability boundaries through representative applications, outlines major challenges, and highlights future research directions to inform method design and deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A review of recent advances in Gaussian splatting

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Shuaitao Fan,
  • Lu Zeng,
  • Wei He,
  • Lunning Zhang,
  • Wenhe Chen,
  • Zhenhua Pan

摘要

Gaussian splatting (GS) enables high-quality real-time novel view synthesis by representing scenes with explicit Gaussian primitives and leveraging efficient differentiable rendering. It has expanded from static reconstruction to dynamic modeling, editability, and semantic understanding, showing value in simultaneous localization and mapping (SLAM), autonomous driving, digital humans, and medical imaging. To address the rapid growth of the literature, increasingly diverse technical routes, and inconsistent evaluation protocols, this survey first reviews the representation and rendering fundamentals of Gaussian splatting and summarizes commonly used datasets and metrics to establish a comparative baseline. It then structures key methods around two threads, covering efficiency and quality optimization for static scenes and spatiotemporal modeling for dynamic scenes. The survey further summarizes the needs for AI-driven semantic understanding and editability, artificial intelligence generated content (AIGC), and contextualized representation and reasoning in embodied intelligent interaction. Finally, it discusses capability boundaries through representative applications, outlines major challenges, and highlights future research directions to inform method design and deployment.