Scholarlink

<p>Establishing dense correspondences is fundamental for building accurate three-dimensional statistical models (3DSMs) of the human head and face, supporting applications in computer vision, medicine, and virtual reality. Traditional correspondence methods depend heavily on manually annotated landmarks, which are labor-intensive, prone to operator bias, and sensitive to dataset scale and diversity. To address these limitations, we propose a fully automated, markerless dense correspondence framework that integrates global rigid alignment and local residual refinement using a lightweight multilayer perceptron (MLP). Specifically, our approach employs region-aware sampling to efficiently preserve key facial features while reducing redundancy in non-critical areas, and applies a hybrid optimization scheme that decouples global rigid transformation (via SVD) from non-rigid local corrections (via residual MLP). Experimental results on a large-scale 3D head scan dataset demonstrate that our method reduces the Mean Corresponding Point Distance (MCPD) by 58% (3.75 ± 0.25&#xa0;mm vs. 9.0 ± 0.56&#xa0;mm) and shortens processing time by 30% (45.61 ± 0.08&#xa0;s vs. 64.98 ± 0.36&#xa0;s) compared to traditional global fitting, with all improvements statistically significant (<i>p &lt; </i>0.001). Additionally, our pipeline achieves sub-millimeter landmark detection error (0.68 ± 0.36&#xa0;mm) and demonstrates robust performance against classical and learning-based baselines. These results validate the effectiveness, efficiency, and semantic accuracy of our approach, highlighting its practical potential for large-scale medical imaging, virtual reality, and personalized product design. Future work will focus on extending robustness to more challenging real-world scenarios and integrating learning-based strategies for further enhancement.</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

Dense correspondence relationships of 3D facial models under a global and local fitting framework

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Wenyu Ma,
  • Xiaobo Bai,
  • Baojun Liu,
  • Yaqing Yu,
  • Yu Zhu

摘要

Establishing dense correspondences is fundamental for building accurate three-dimensional statistical models (3DSMs) of the human head and face, supporting applications in computer vision, medicine, and virtual reality. Traditional correspondence methods depend heavily on manually annotated landmarks, which are labor-intensive, prone to operator bias, and sensitive to dataset scale and diversity. To address these limitations, we propose a fully automated, markerless dense correspondence framework that integrates global rigid alignment and local residual refinement using a lightweight multilayer perceptron (MLP). Specifically, our approach employs region-aware sampling to efficiently preserve key facial features while reducing redundancy in non-critical areas, and applies a hybrid optimization scheme that decouples global rigid transformation (via SVD) from non-rigid local corrections (via residual MLP). Experimental results on a large-scale 3D head scan dataset demonstrate that our method reduces the Mean Corresponding Point Distance (MCPD) by 58% (3.75 ± 0.25 mm vs. 9.0 ± 0.56 mm) and shortens processing time by 30% (45.61 ± 0.08 s vs. 64.98 ± 0.36 s) compared to traditional global fitting, with all improvements statistically significant (p < 0.001). Additionally, our pipeline achieves sub-millimeter landmark detection error (0.68 ± 0.36 mm) and demonstrates robust performance against classical and learning-based baselines. These results validate the effectiveness, efficiency, and semantic accuracy of our approach, highlighting its practical potential for large-scale medical imaging, virtual reality, and personalized product design. Future work will focus on extending robustness to more challenging real-world scenarios and integrating learning-based strategies for further enhancement.