Scholarlink

<p>Skeleton-based action recognition has gained significant attention due to its efficiency and robustness. However, traditional methods often struggle to extract structured information from human skeleton sequences at multiple levels: spatial, temporal, and sequence levels. To address this, we propose a Multi-level Fusion Token (MFT) framework that leverages a novel mask strategy to simultaneously mask skeleton sequences in both spatial and temporal domains. By incorporating three-level fusion tokens, our model effectively extracts informative features from the temporal, spatial, and sequence levels during reconstruction learning. Furthermore, we enhance the expressiveness of these tokens through knowledge distillation and contrastive learning tasks. Extensive experiments on benchmark datasets demonstrate the superiority of our MFT framework over state-of-the-art methods, underscoring its potential to significantly advance skeleton-based action recognition. The code will be available at: <a href="https://github.com/jie7410/MFT">https://github.com/jie7410/MFT</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

Multi-level fusion tokens for enhanced self-supervised skeleton-based action recognition

  • 期刊论文

  • 发表日期：

    2025年12月10日

查看全文
  • Jie Zhao,
  • Kaida Ning,
  • Feng Zhou,
  • Junjun Pan,
  • Hongwen Xu,
  • Ju Dai

摘要

Skeleton-based action recognition has gained significant attention due to its efficiency and robustness. However, traditional methods often struggle to extract structured information from human skeleton sequences at multiple levels: spatial, temporal, and sequence levels. To address this, we propose a Multi-level Fusion Token (MFT) framework that leverages a novel mask strategy to simultaneously mask skeleton sequences in both spatial and temporal domains. By incorporating three-level fusion tokens, our model effectively extracts informative features from the temporal, spatial, and sequence levels during reconstruction learning. Furthermore, we enhance the expressiveness of these tokens through knowledge distillation and contrastive learning tasks. Extensive experiments on benchmark datasets demonstrate the superiority of our MFT framework over state-of-the-art methods, underscoring its potential to significantly advance skeleton-based action recognition. The code will be available at: https://github.com/jie7410/MFT.