Scholarlink

<p>Accurate 2D human pose estimation relies on the efficient fusion of multi-scale and multi-level features. However, existing methods often suffer from pixel misalignment and feature heterogeneity across resolutions, leading to information loss and reduced performance. To address these issues, we propose the Feature-Aligned High-Resolution Network (FA-HRNet), which incorporates a Multi-Scale Feature-Aligned Hourglass Module. This module leverages Hourglass-UP and Hourglass-Down components to explicitly align multi-scale features and model cross-scale semantic relationships. Furthermore, a coordinate attention module is integrated to enhance heatmap quality. Extensive experiments on the COCO and MPII datasets demonstrate that FA-HRNet achieves state-of-the-art performance, outperforming existing methods in both accuracy and robustness, especially in complex pose estimation scenarios. Evaluations on both LSP and VISAD datasets validate FA-HRNet’s strong feature generalization power and excellent cross-domain transferability. Our work highlights the importance of fine-grained feature alignment in high-resolution networks for improving 2D human pose estimation. Code and models will be available at <a href="https://github.com/ZungWin/FA-HRNet">https://github.com/ZungWin/FA-HRNet</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

Enhanced 2D human pose estimation via feature-aligned high-resolution network

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Yuhe Zhu,
  • Zhangwen Lyu,
  • Rong Liu,
  • Yinwei Zhan

摘要

Accurate 2D human pose estimation relies on the efficient fusion of multi-scale and multi-level features. However, existing methods often suffer from pixel misalignment and feature heterogeneity across resolutions, leading to information loss and reduced performance. To address these issues, we propose the Feature-Aligned High-Resolution Network (FA-HRNet), which incorporates a Multi-Scale Feature-Aligned Hourglass Module. This module leverages Hourglass-UP and Hourglass-Down components to explicitly align multi-scale features and model cross-scale semantic relationships. Furthermore, a coordinate attention module is integrated to enhance heatmap quality. Extensive experiments on the COCO and MPII datasets demonstrate that FA-HRNet achieves state-of-the-art performance, outperforming existing methods in both accuracy and robustness, especially in complex pose estimation scenarios. Evaluations on both LSP and VISAD datasets validate FA-HRNet’s strong feature generalization power and excellent cross-domain transferability. Our work highlights the importance of fine-grained feature alignment in high-resolution networks for improving 2D human pose estimation. Code and models will be available at https://github.com/ZungWin/FA-HRNet.