Scholarlink

<p>Concrete defects usually diminish the structural aesthetics, and more importantly, they may significantly impact the safety and durability of structures. While conventional inspection methods are often subjective and inefficient, modern computer vision techniques offer a more automated and accurate alternative. To capture the spatial information of defects, a framework combining Segment Anything Model 2 (SAM 2) and 3D Gaussian Splatting (3DGS) was proposed for high-quality defect segmentation and 3D reconstruction based on monocular video. The proposed framework demonstrates advantages in surface detail, defect segmentation, and rendering speed compared to existing approaches. Key contributions of this work include the effective integration of SAM 2 and 3DGS, improved surface feature representation, and an effective solution for 3D reconstruction using monocular video. Experimental results highlight the framework's improvements in accuracy, detail, and scalability over existing methods.</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

Defect segmentation and 3D reconstruction in concrete structures using SAM 2 and 3D Gaussian splatting

  • 期刊论文

  • 发表日期：

    2025年7月17日

查看全文
  • Dayou Duan,
  • Zuocai Wang,
  • Yu Xin,
  • Yajie Ding

摘要

Concrete defects usually diminish the structural aesthetics, and more importantly, they may significantly impact the safety and durability of structures. While conventional inspection methods are often subjective and inefficient, modern computer vision techniques offer a more automated and accurate alternative. To capture the spatial information of defects, a framework combining Segment Anything Model 2 (SAM 2) and 3D Gaussian Splatting (3DGS) was proposed for high-quality defect segmentation and 3D reconstruction based on monocular video. The proposed framework demonstrates advantages in surface detail, defect segmentation, and rendering speed compared to existing approaches. Key contributions of this work include the effective integration of SAM 2 and 3DGS, improved surface feature representation, and an effective solution for 3D reconstruction using monocular video. Experimental results highlight the framework's improvements in accuracy, detail, and scalability over existing methods.