Scholarlink

<p>3D human pose estimation from monocular video is a core task in computer vision and virtual reality. Recovering poses with absolute metric scale remains challenging due to depth and scale ambiguities. Existing methods often rely on depth sensors or dense annotations, limiting real-world deployment. Here, we present a multi-frame framework for absolute-scale 3D human pose estimation using only relative depth priors and sparse supervision. The model integrates a depth-scale recovery module and temporal feature fusion to resolve scale ambiguity without absolute depth annotations. We show the method achieves 48.3&#xa0;mm MPJPE and 36.7&#xa0;mm PA-MPJPE on Human3.6M, outperforming state-of-the-art approaches. This lightweight, annotation-efficient solution supports applications in smart sports, human–computer interaction, and virtual reality, broadening the deployment scope of monocular 3D pose systems. The code for our model implementation can be found at: <a href="https://github.com/gitcodepapers/MFM3DposeNet">https://github.com/gitcodepapers/MFM3DposeNet</a>.</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

Absolute-scale 3D human pose estimation from monocular videos via relative depth priors and multi-frame fusion

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • Huagang He,
  • Yinghao Li,
  • Beiyang Liu,
  • Qingxia Wu

摘要

3D human pose estimation from monocular video is a core task in computer vision and virtual reality. Recovering poses with absolute metric scale remains challenging due to depth and scale ambiguities. Existing methods often rely on depth sensors or dense annotations, limiting real-world deployment. Here, we present a multi-frame framework for absolute-scale 3D human pose estimation using only relative depth priors and sparse supervision. The model integrates a depth-scale recovery module and temporal feature fusion to resolve scale ambiguity without absolute depth annotations. We show the method achieves 48.3 mm MPJPE and 36.7 mm PA-MPJPE on Human3.6M, outperforming state-of-the-art approaches. This lightweight, annotation-efficient solution supports applications in smart sports, human–computer interaction, and virtual reality, broadening the deployment scope of monocular 3D pose systems. The code for our model implementation can be found at: https://github.com/gitcodepapers/MFM3DposeNet.