Scholarlink

To address the challenge of enhancing the performance of human pose estimation algorithms while reducing floating-point computation, this paper proposes an efficient top-down lightweight human pose estimation model called Efficient RTMPose, which is built on RTMPose. Firstly, we propose MBLConv and apply it to EfficientNet as the backbone network of RTMPose to reduce the number of parameters and computational load. MBLConv comprises a series of convolutions and the Large Selective Kernel (LSK) mechanism, allowing for the dynamic adjustment of the receptive field of the feature extraction backbone. This allows the model to have a larger receptive field, enhancing its capability for human keypoint detection tasks. Next, we propose the tAPE-GAU, which integrates time Absolute Position Encoding (tAPE) into the Gate Attention Unit (GAU). By leveraging both global and local spatial information, tAPE-GAU maintains low parameter and computational complexity while enhancing the network’s performance and robustness. This paper conducts experimental validation using MBLConv and tAPE-GAU on the MPII and COCO2017 datasets. The results demonstrate significant improvements in accuracy, parameter efficiency, and computational load compared to traditional models.

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

Lightweight and Efficient Top-Down Human Pose Estimation Algorithm Research

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xiaofang Mu,
  • Minghui Song,
  • Hong Shi,
  • Mingxing Hou,
  • Shuxian Guo,
  • Wu Xiaotong

摘要

To address the challenge of enhancing the performance of human pose estimation algorithms while reducing floating-point computation, this paper proposes an efficient top-down lightweight human pose estimation model called Efficient RTMPose, which is built on RTMPose. Firstly, we propose MBLConv and apply it to EfficientNet as the backbone network of RTMPose to reduce the number of parameters and computational load. MBLConv comprises a series of convolutions and the Large Selective Kernel (LSK) mechanism, allowing for the dynamic adjustment of the receptive field of the feature extraction backbone. This allows the model to have a larger receptive field, enhancing its capability for human keypoint detection tasks. Next, we propose the tAPE-GAU, which integrates time Absolute Position Encoding (tAPE) into the Gate Attention Unit (GAU). By leveraging both global and local spatial information, tAPE-GAU maintains low parameter and computational complexity while enhancing the network’s performance and robustness. This paper conducts experimental validation using MBLConv and tAPE-GAU on the MPII and COCO2017 datasets. The results demonstrate significant improvements in accuracy, parameter efficiency, and computational load compared to traditional models.