Scholarlink

<p>In recent years, Transformer-based 3D human pose estimation has achieved remarkable progress. However, existing methods primarily focus on interactions between joint pairs, neglecting the holistic coordination among multiple joints and the higher-level structural semantics, which leads to poor performance in handling challenging poses such as self-occlusion, complex, or rare postures. To address this issue, we propose HSA-DMT, a unified framework that integrates a Hierarchical Structure-Aware Embedding (HSAE) with a Dual-Mode Temporal Encoder (DMTE). Specifically, HSAE incorporates prior knowledge of human body structure and establishes a three-level interaction framework("joint <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\leftrightarrow \)</EquationSource> </InlineEquation> adjacent joints", "joint <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\leftrightarrow \)</EquationSource> </InlineEquation> limb region" and "joint <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\leftrightarrow \)</EquationSource> </InlineEquation> pose domain"), thereby enhancing structural modeling. Meanwhile, DMTE dynamically fuses short-term and long-term temporal information by processing both continuous and strided sequences in parallel, effectively capturing subtle positional variations and global motion trends. This integrated approach to spatial-structural and temporal-dynamic modeling facilitates more robust pose estimation in complex scenarios. Experiments on the Human3.6M and MPI-INF-3DHP datasets demonstrate that HSA-DMT achieves performance comparable to state-of-the-art approaches in 3D human pose estimation.</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

Hsa-dmt: hierarchical structure-aware embedding and dual-mode temporal modeling for challenging 3D human pose estimation

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Junfen Chen,
  • Zhaoyang Zhou,
  • Bojun Xie,
  • Jie Zhu,
  • Yan Li

摘要

In recent years, Transformer-based 3D human pose estimation has achieved remarkable progress. However, existing methods primarily focus on interactions between joint pairs, neglecting the holistic coordination among multiple joints and the higher-level structural semantics, which leads to poor performance in handling challenging poses such as self-occlusion, complex, or rare postures. To address this issue, we propose HSA-DMT, a unified framework that integrates a Hierarchical Structure-Aware Embedding (HSAE) with a Dual-Mode Temporal Encoder (DMTE). Specifically, HSAE incorporates prior knowledge of human body structure and establishes a three-level interaction framework("joint \(\leftrightarrow \) adjacent joints", "joint \(\leftrightarrow \) limb region" and "joint \(\leftrightarrow \) pose domain"), thereby enhancing structural modeling. Meanwhile, DMTE dynamically fuses short-term and long-term temporal information by processing both continuous and strided sequences in parallel, effectively capturing subtle positional variations and global motion trends. This integrated approach to spatial-structural and temporal-dynamic modeling facilitates more robust pose estimation in complex scenarios. Experiments on the Human3.6M and MPI-INF-3DHP datasets demonstrate that HSA-DMT achieves performance comparable to state-of-the-art approaches in 3D human pose estimation.