Scholarlink

<p>Estimating 3D human poses from monocular images is a pivotal challenge in computer vision, offering comprehensive spatial geometric information crucial for action analysis and semantic understanding. Despite recent advances, existing methods still struggle to balance global long-range dependencies with local spatiotemporal details, often overlooking temporal autocorrelation within pose sequences. This paper introduces FLTGFormer, a dual-stream Transformer-GCN network enhanced by frequency loss, to address these challenges. The Transformer stream incorporates a polarity-aware attention mechanism (PolaFormer) to refine global spatial representation, while the GCN stream employs a dynamic adaptive fusion graph convolution module (DAF-GCNFormer) to model local spatiotemporal dependencies. An adaptive gating mechanism integrates features from both streams, and a frequency loss term helps recover autocorrelation, reducing temporal regression bias. Experimental results on the Human3.6&#xa0;M and MPI-INF-3DHP datasets demonstrate significant enhancements in estimation accuracy (MPJPE scores of 37.2&#xa0;mm and 14.8&#xa0;mm, respectively) with low computational complexity, achieving an effective trade-off between performance and efficiency.</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

FLTGFormer: a dual-stream transformer-GCN network enhanced by frequency loss for monocular 3D human pose estimation

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Xuejing Gu,
  • Zibo Zhao,
  • Congzhe Zhang

摘要

Estimating 3D human poses from monocular images is a pivotal challenge in computer vision, offering comprehensive spatial geometric information crucial for action analysis and semantic understanding. Despite recent advances, existing methods still struggle to balance global long-range dependencies with local spatiotemporal details, often overlooking temporal autocorrelation within pose sequences. This paper introduces FLTGFormer, a dual-stream Transformer-GCN network enhanced by frequency loss, to address these challenges. The Transformer stream incorporates a polarity-aware attention mechanism (PolaFormer) to refine global spatial representation, while the GCN stream employs a dynamic adaptive fusion graph convolution module (DAF-GCNFormer) to model local spatiotemporal dependencies. An adaptive gating mechanism integrates features from both streams, and a frequency loss term helps recover autocorrelation, reducing temporal regression bias. Experimental results on the Human3.6 M and MPI-INF-3DHP datasets demonstrate significant enhancements in estimation accuracy (MPJPE scores of 37.2 mm and 14.8 mm, respectively) with low computational complexity, achieving an effective trade-off between performance and efficiency.