Scholarlink

<p>Face parsing is a fundamental component of many advanced face technologies, which assigns labels to each pixel on the face data. Although three-dimensional (3D) face parsing has the potential to outperform its two-dimensional (2D) counterpart, it remains challenging due to the high cost of processing 3D mesh data. Recent works have introduced various methods for 3D surface segmentation, but their performance is still limited and they consume large amounts of memory and computation. In this paper, we propose a “3D–2D–3D” strategy for 3D face parsing. First, we transform 3D face data into a topological disk-like 2D face image containing spatial and textural information via conformal parameterization. Subsequently, we use a specific 2D deep learning network called CPFNet to achieve 2D face image semantic segmentation with multiscale technology and feature aggregation. Finally, the 2D semantic result is inversely remapped to the 3D face data to achieve 3D face parsing. Experimental results show that both CPFNet and our “3D–2D–3D” strategy accomplish high-quality 3D face parsing and outperform some 2D networks and 3D methods in both qualitative and quantitative comparisons.</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

3D face parsing based on 2D CPFNet: conformal parameterized face parsing network

  • 期刊论文

  • 发表日期：

    2025年2月11日

查看全文
  • M. Yang,
  • W. Sun,
  • Y. Wang,
  • G. Zhou,
  • J. Tong,
  • P. Zhou

摘要

Face parsing is a fundamental component of many advanced face technologies, which assigns labels to each pixel on the face data. Although three-dimensional (3D) face parsing has the potential to outperform its two-dimensional (2D) counterpart, it remains challenging due to the high cost of processing 3D mesh data. Recent works have introduced various methods for 3D surface segmentation, but their performance is still limited and they consume large amounts of memory and computation. In this paper, we propose a “3D–2D–3D” strategy for 3D face parsing. First, we transform 3D face data into a topological disk-like 2D face image containing spatial and textural information via conformal parameterization. Subsequently, we use a specific 2D deep learning network called CPFNet to achieve 2D face image semantic segmentation with multiscale technology and feature aggregation. Finally, the 2D semantic result is inversely remapped to the 3D face data to achieve 3D face parsing. Experimental results show that both CPFNet and our “3D–2D–3D” strategy accomplish high-quality 3D face parsing and outperform some 2D networks and 3D methods in both qualitative and quantitative comparisons.