Scholarlink

<p>Recent research on accurately generating human bodies, such as SMPL-X, for human motion and expression has gained momentum. However, traditional 2D image-based methods for generating human body models inevitably face limitations, such as depth ambiguity. This paper proposes a method that accurately generates human bodies, like SMPL or SMPL-X, using 3D volumetric data. Since 3D volumetric data contains precise spatial information, we leverage it to estimate SMPL-X joints accurately and reconstruct the SMPL-X model from them. The proposed method extracts 2D key points from multi-view images projected from 3D volumetric data and generates an accurate human body model through confidence-based 3D reconstruction. By applying this method, we can generate body models with a mean per-joint position error (MPJPE) of about 50&#xa0;mm, resolving the depth ambiguity issue. Additionally, by fitting 3D volumetric data to human mesh models, we effectively utilize the rich depth information to overcome challenges such as depth ambiguity.</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

High accurate SMPL-X generation based on volumetric reconstruction

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Jung-Woo Kim,
  • Hak-Bum Lee,
  • Seung-Hwan Yoon,
  • Seung-Jun Yang,
  • Gi-Mun Um,
  • Young-Ho Seo

摘要

Recent research on accurately generating human bodies, such as SMPL-X, for human motion and expression has gained momentum. However, traditional 2D image-based methods for generating human body models inevitably face limitations, such as depth ambiguity. This paper proposes a method that accurately generates human bodies, like SMPL or SMPL-X, using 3D volumetric data. Since 3D volumetric data contains precise spatial information, we leverage it to estimate SMPL-X joints accurately and reconstruct the SMPL-X model from them. The proposed method extracts 2D key points from multi-view images projected from 3D volumetric data and generates an accurate human body model through confidence-based 3D reconstruction. By applying this method, we can generate body models with a mean per-joint position error (MPJPE) of about 50 mm, resolving the depth ambiguity issue. Additionally, by fitting 3D volumetric data to human mesh models, we effectively utilize the rich depth information to overcome challenges such as depth ambiguity.