Scholarlink

<p>We present a Weight-Allocated Clustering-Based Lightweight Network (WACLN) for 3D human pose estimation, aiming to strike an effective balance between detection accuracy and computational efficiency. Our approach introduces a Multi-Branch Hypothesis Generation module that utilizes a parallel multi-branch architecture to efficiently extract features, with deep branches focusing on individual joint locations and shallow branches emphasizing inter-joint correlations. Furthermore, we develop a Frames Pruning module leveraging a Weight-Allocated Clustering Algorithm to compress long video sequences into a limited set of key frames, eliminating redundancy and significantly reducing computational load. Experimental results on the Human3.6M and MPI-INF-3DHP datasets demonstrate that WACLN achieves a 45% reduction in computational cost compared to the baseline MHFormer, with a 0.4mm decrease in MPJPE, while achieving the shortest training time among current state-of-the-art models. Here, we show that our method effectively balances accuracy and efficiency, making it suitable for real-time applications. The code is available at <a href="https://github.com/wangshicahng/WACLN.">https://github.com/wangshicahng/WACLN.</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

Optimized lightweight 3D human pose estimation via multi-branch hypothesis generation and frame pruning

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • GaoXiang Ma,
  • Ji Zhao,
  • HongHan Li

摘要

We present a Weight-Allocated Clustering-Based Lightweight Network (WACLN) for 3D human pose estimation, aiming to strike an effective balance between detection accuracy and computational efficiency. Our approach introduces a Multi-Branch Hypothesis Generation module that utilizes a parallel multi-branch architecture to efficiently extract features, with deep branches focusing on individual joint locations and shallow branches emphasizing inter-joint correlations. Furthermore, we develop a Frames Pruning module leveraging a Weight-Allocated Clustering Algorithm to compress long video sequences into a limited set of key frames, eliminating redundancy and significantly reducing computational load. Experimental results on the Human3.6M and MPI-INF-3DHP datasets demonstrate that WACLN achieves a 45% reduction in computational cost compared to the baseline MHFormer, with a 0.4mm decrease in MPJPE, while achieving the shortest training time among current state-of-the-art models. Here, we show that our method effectively balances accuracy and efficiency, making it suitable for real-time applications. The code is available at https://github.com/wangshicahng/WACLN.