Scholarlink

<p>Photorealistic 3D human reconstruction remains a persistent challenge in computer vision. Clothed Human Reconstruction (CHR) aims to recover 3D humans from various inputs while preserving clothing details. Leveraging the ubiquitous RGB photographs and videos in daily scenarios, existing RGB-based methods enable user-friendly reconstruction and demonstrate promising results. Nevertheless, inherent limitations persist due to intricate garment details, occlusions, and individual diversity, which frequently lead to inaccurate geometry and poor appearance, particularly for occluded regions. The recent emergence of diffusion models provides a generative paradigm to mitigate these challenges, particularly in occluded view synthesis and texture inpainting. We start with 3D human representations and then elaborate on the non-diffusion and diffusion-based reconstruction methods by analyzing their methodological innovations and limitations. In addition, we summarize the related datasets as well as the commonly adopted evaluation metrics, and we also conduct a brief analysis of benchmark results. Finally, we draw conclusions through discussing the open issues and sharing our insights on future directions.</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

Clothed human reconstruction from RGB data: A survey of traditional and diffusion-based methods

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月20日

查看全文
  • Dong-Lin Chen,
  • Mohd Shafry Mohd Rahim,
  • Hiew Moi Sim,
  • Bin Wang,
  • Min-Song Li,
  • Asniyani Nur Haidar Abdullah

摘要

Photorealistic 3D human reconstruction remains a persistent challenge in computer vision. Clothed Human Reconstruction (CHR) aims to recover 3D humans from various inputs while preserving clothing details. Leveraging the ubiquitous RGB photographs and videos in daily scenarios, existing RGB-based methods enable user-friendly reconstruction and demonstrate promising results. Nevertheless, inherent limitations persist due to intricate garment details, occlusions, and individual diversity, which frequently lead to inaccurate geometry and poor appearance, particularly for occluded regions. The recent emergence of diffusion models provides a generative paradigm to mitigate these challenges, particularly in occluded view synthesis and texture inpainting. We start with 3D human representations and then elaborate on the non-diffusion and diffusion-based reconstruction methods by analyzing their methodological innovations and limitations. In addition, we summarize the related datasets as well as the commonly adopted evaluation metrics, and we also conduct a brief analysis of benchmark results. Finally, we draw conclusions through discussing the open issues and sharing our insights on future directions.