Scholarlink

<p>Existing self-supervised skeleton action recognition methods mainly rely on spatiotemporal structure modeling and often underexploit high-level action semantics. To address this, we propose a semantically guided multi-scale contrastive learning framework (SMCLR), which learns more discriminative and generalizable skeleton action representations by jointly modeling spatial structure, temporal dynamics, and semantic priors. This method introduces body part-level semantic constraints in the spatial dimension to enhance joint topology modeling capabilities, constructs a scale-aware adaptive multi-scale temporal modeling mechanism in the temporal dimension to characterize diverse action dynamics, and utilizes action semantic prototypes generated by an LLM to inject high-level semantic information into the self-supervised contrastive learning process through cross-modal alignment. Experimental results on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD show that SMCLR achieves competitive performance and strong generalization ability under multiple evaluation protocols.</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

Smclr: semantic-guided multi-scale contrastive learning for skeleton-based action recognition

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Min Wei,
  • Hongwei Chen,
  • Xia Li

摘要

Existing self-supervised skeleton action recognition methods mainly rely on spatiotemporal structure modeling and often underexploit high-level action semantics. To address this, we propose a semantically guided multi-scale contrastive learning framework (SMCLR), which learns more discriminative and generalizable skeleton action representations by jointly modeling spatial structure, temporal dynamics, and semantic priors. This method introduces body part-level semantic constraints in the spatial dimension to enhance joint topology modeling capabilities, constructs a scale-aware adaptive multi-scale temporal modeling mechanism in the temporal dimension to characterize diverse action dynamics, and utilizes action semantic prototypes generated by an LLM to inject high-level semantic information into the self-supervised contrastive learning process through cross-modal alignment. Experimental results on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD show that SMCLR achieves competitive performance and strong generalization ability under multiple evaluation protocols.