Scholarlink

<p>Human pose estimation, which involves predicting keypoint locations in images to model the human body structure, is widely used in tasks such as human-computer interaction and action recognition. Despite the strong performance of current high-accuracy models, their high computational complexity and large parameter sizes limit their real-time deployment on edge devices. To address this, we propose EfficientGLS-Pose, an efficient and lightweight global–local collaborative pose estimation framework designed to optimize multi-scale feature modeling. The framework includes a lightweight multi-scale feature extraction module (CSP-PMSFA), a global–local selective fusion pyramid (GLS-FPN), and a lightweight decoupled detection head (LADH-PoseHead). On the MSCOCO2017 and CrowdPose datasets, EfficientGLS-Pose-t achieves 59.5% COCO AP with only 1.9M parameters and 4.5 GFLOPs, while EfficientGLS-Pose-b achieves 62.8% and 68.3% on these datasets, respectively. These results demonstrate the potential of EfficientGLS-Pose for edge deployment, offering a balance between accuracy 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

Efficientgls-pose: enhancing human pose estimation efficiency and accuracy via global–local collaborative modeling

  • 期刊论文

  • 发表日期：

    2026年3月11日

查看全文
  • Yongfeng Qi,
  • Yuanzhe Lin,
  • Mingsen Hu,
  • Heng Zhang,
  • Gaoyang Dai

摘要

Human pose estimation, which involves predicting keypoint locations in images to model the human body structure, is widely used in tasks such as human-computer interaction and action recognition. Despite the strong performance of current high-accuracy models, their high computational complexity and large parameter sizes limit their real-time deployment on edge devices. To address this, we propose EfficientGLS-Pose, an efficient and lightweight global–local collaborative pose estimation framework designed to optimize multi-scale feature modeling. The framework includes a lightweight multi-scale feature extraction module (CSP-PMSFA), a global–local selective fusion pyramid (GLS-FPN), and a lightweight decoupled detection head (LADH-PoseHead). On the MSCOCO2017 and CrowdPose datasets, EfficientGLS-Pose-t achieves 59.5% COCO AP with only 1.9M parameters and 4.5 GFLOPs, while EfficientGLS-Pose-b achieves 62.8% and 68.3% on these datasets, respectively. These results demonstrate the potential of EfficientGLS-Pose for edge deployment, offering a balance between accuracy and efficiency.