Scholarlink

<p>Conventional 2D human pose estimation methods commonly formulate pose inference as a global regression or dense heatmap prediction problem under implicit assumptions of standard human body proportions. Such assumptions limit robustness in virtual avatar scenarios, where exaggerated body shapes, personalized clothing, self-occlusion, and fast motion are common. In this paper, we reformulate 2D pose estimation as a joint-centric, physics-inspired spatiotemporal inference problem. A lightweight joint-centric spatiotemporal feature extraction framework models each anatomical joint through an independent joint-wise branch to capture localized spatial evidence and temporal motion dynamics, while efficient spatiotemporal coordination preserves motion continuity without relying on heavy global attention architectures. In addition, physics-inspired kinematic and dynamic priors are introduced as soft regularization terms to encourage physically plausible and temporally stable pose predictions. Experiments on COCO, MPII and a VR avatar dataset demonstrate that the proposed method improves robustness across both real-human and virtual avatar scenarios while maintaining real-time performance. The code is available at <a href="https://github.com/quanzhong0429/JC-STNet">https://github.com/quanzhong0429/JC-STNet</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

JC-STNet: a physics-inspired joint-centric spatiotemporal network for real-time 2D pose estimation

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Xingquan Cai,
  • Kaijie Qu,
  • Chuansheng Liu,
  • Xinzhu Pu,
  • Liang Shi

摘要

Conventional 2D human pose estimation methods commonly formulate pose inference as a global regression or dense heatmap prediction problem under implicit assumptions of standard human body proportions. Such assumptions limit robustness in virtual avatar scenarios, where exaggerated body shapes, personalized clothing, self-occlusion, and fast motion are common. In this paper, we reformulate 2D pose estimation as a joint-centric, physics-inspired spatiotemporal inference problem. A lightweight joint-centric spatiotemporal feature extraction framework models each anatomical joint through an independent joint-wise branch to capture localized spatial evidence and temporal motion dynamics, while efficient spatiotemporal coordination preserves motion continuity without relying on heavy global attention architectures. In addition, physics-inspired kinematic and dynamic priors are introduced as soft regularization terms to encourage physically plausible and temporally stable pose predictions. Experiments on COCO, MPII and a VR avatar dataset demonstrate that the proposed method improves robustness across both real-human and virtual avatar scenarios while maintaining real-time performance. The code is available at https://github.com/quanzhong0429/JC-STNet.