Scholarlink

<p>Image-based 3D reconstruction has become a crucial research focus in computer vision owing to its high precision and real-time performance. However, existing methods still exhibit inherent limitations in low-texture scenarios, including unstable feature matching and incomplete geometric structures in reconstructed point clouds. To address these challenges, this paper proposes a 3D reconstruction system capable of generating highly realistic point cloud models for low-texture objects. The system comprises an image acquisition device, a Dual-loop Pose Estimation module, and an improved depth map generation model TL-PatchmatchNet. The image acquisition device captures multi-view images of low-texture objects to construct a dedicated dataset. The Dual-loop Pose Estimation module achieves initial pose estimation through bidirectional incremental reconstruction and refines results by integrating heatmaps. The TL-PatchmatchNet algorithm enhances feature representation capabilities in weakly textured regions by incorporating attention mechanisms and texture constraints into its baseline architecture, enabling more robust point cloud reconstruction. Experimental results demonstrate the framework’s superior robustness and reconstruction accuracy across both conventional and low-texture datasets.</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

Multi-view 3D reconstruction system based on algorithm for low-texture small objects

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Qinghao Zhou,
  • Xiao Liang,
  • Jie Qian,
  • Wenjing Yang,
  • Feng Cao

摘要

Image-based 3D reconstruction has become a crucial research focus in computer vision owing to its high precision and real-time performance. However, existing methods still exhibit inherent limitations in low-texture scenarios, including unstable feature matching and incomplete geometric structures in reconstructed point clouds. To address these challenges, this paper proposes a 3D reconstruction system capable of generating highly realistic point cloud models for low-texture objects. The system comprises an image acquisition device, a Dual-loop Pose Estimation module, and an improved depth map generation model TL-PatchmatchNet. The image acquisition device captures multi-view images of low-texture objects to construct a dedicated dataset. The Dual-loop Pose Estimation module achieves initial pose estimation through bidirectional incremental reconstruction and refines results by integrating heatmaps. The TL-PatchmatchNet algorithm enhances feature representation capabilities in weakly textured regions by incorporating attention mechanisms and texture constraints into its baseline architecture, enabling more robust point cloud reconstruction. Experimental results demonstrate the framework’s superior robustness and reconstruction accuracy across both conventional and low-texture datasets.