Scholarlink

<p>Human pose estimation is a critical challenge in computer vision, with existing methods struggling to balance model efficiency and accuracy. Overemphasis on accuracy often results in high computational costs, while focusing on lightweight design can compromise performance. To address these challenges, we propose SD-HRNet, a lightweight human pose estimation network built upon the HRNet architecture. The core innovation of SD-HRNet is the introduction of SDShuffleBlock, which integrates shared-weight downsampling and partial convolution techniques. These innovations achieve a trade-off between computational efficiency and predictive accuracy. When evaluated on COCO2017 and MPII benchmarks, SD-HRNet outperforms existing lightweight models, achieving 66.6% AP on COCO2017 with only 205&#xa0;M FLOPs and 0.98M parameters and achieving inference speeds of 132+ FPS on NVIDIA A100 GPUs and 1.5+ FPS on CPU. Compared to current methods, SD-HRNet improves AP by 1.5% to 2.3% while reducing parameters by 18% to 30%. Ablation studies confirm the effectiveness of the lightweight backbone, shared-weight convolution, and partial convolution blocks, demonstrating SD-HRNet as a high-performance, lightweight solution for 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

SD-HRNet: a lightweight high-resolution network for human pose estimation based on spatial decoupling

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • Zhixin Li,
  • Yunfeng Dong,
  • Xiaoming Wu,
  • Xiangzhi Liu

摘要

Human pose estimation is a critical challenge in computer vision, with existing methods struggling to balance model efficiency and accuracy. Overemphasis on accuracy often results in high computational costs, while focusing on lightweight design can compromise performance. To address these challenges, we propose SD-HRNet, a lightweight human pose estimation network built upon the HRNet architecture. The core innovation of SD-HRNet is the introduction of SDShuffleBlock, which integrates shared-weight downsampling and partial convolution techniques. These innovations achieve a trade-off between computational efficiency and predictive accuracy. When evaluated on COCO2017 and MPII benchmarks, SD-HRNet outperforms existing lightweight models, achieving 66.6% AP on COCO2017 with only 205 M FLOPs and 0.98M parameters and achieving inference speeds of 132+ FPS on NVIDIA A100 GPUs and 1.5+ FPS on CPU. Compared to current methods, SD-HRNet improves AP by 1.5% to 2.3% while reducing parameters by 18% to 30%. Ablation studies confirm the effectiveness of the lightweight backbone, shared-weight convolution, and partial convolution blocks, demonstrating SD-HRNet as a high-performance, lightweight solution for pose estimation.