Scholarlink

<p>Human pose estimation, also known as human keypoint detection, aims to accurately localize human keypoints and understand spatial relationships among body parts from image or video data. Despite the strong feature extraction capability of convolutional neural networks, capturing global relationships among keypoints remains challenging. To address this, we propose BR-Pose, a hybrid model that combines convolutional and attention mechanisms. BR-Pose utilizes HRNet to extract multi-scale features and employs a multi-level weight fusion module to adaptively fuse these features. Furthermore, BiFormer, equipped with a bi-level routing attention mechanism, is introduced to encode the fused features, enabling the model to learn dependencies among keypoints with low computational overhead. Experimental results on the COCO and MPII datasets demonstrate the effectiveness and superiority of BR-Pose, achieving competitive performance while maintaining low computational costs. Visualization analyses show that BR-Pose can accurately localize keypoints even in complex occlusion environments, highlighting its robustness and practical applicability. The source code is publicly available at <a href="https://github.com/ll8657/BR-Pose.">https://github.com/ll8657/BR-Pose.</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

BR-Pose: enhancing human pose estimation through Bi-level routing attention and multi-level weight fusion

  • 期刊论文

  • 发表日期：

    2025年4月23日

查看全文
  • Zhi Liu,
  • Lei Liu,
  • Shengzhao Hao

摘要

Human pose estimation, also known as human keypoint detection, aims to accurately localize human keypoints and understand spatial relationships among body parts from image or video data. Despite the strong feature extraction capability of convolutional neural networks, capturing global relationships among keypoints remains challenging. To address this, we propose BR-Pose, a hybrid model that combines convolutional and attention mechanisms. BR-Pose utilizes HRNet to extract multi-scale features and employs a multi-level weight fusion module to adaptively fuse these features. Furthermore, BiFormer, equipped with a bi-level routing attention mechanism, is introduced to encode the fused features, enabling the model to learn dependencies among keypoints with low computational overhead. Experimental results on the COCO and MPII datasets demonstrate the effectiveness and superiority of BR-Pose, achieving competitive performance while maintaining low computational costs. Visualization analyses show that BR-Pose can accurately localize keypoints even in complex occlusion environments, highlighting its robustness and practical applicability. The source code is publicly available at https://github.com/ll8657/BR-Pose.