Scholarlink

<p>Graph Convolutional Networks (GCNs) have become the mainstream paradigm for skeleton-based action recognition, but existing methods face three core bottlenecks: (i) rigid discrete temporal convolution fails to adaptively model human actions with diverse rhythms and scales; (ii) multi-scale feature fusion causes severe channel redundancy and noise, masking key semantic features without adaptive channel calibration; (iii) insufficient discriminative constraints of cross-entropy loss lead to misclassification of visually similar actions. To address these issues, this paper proposes an Adaptive Temporal Alignment and Contrastive Refinement Network (ATACR-Net), following the design principle of "alignment-refinement-discrimination": a learnable continuous temporal sampling mechanism is introduced through a Multi-scale Adaptive Temporal Convolution Module (MATCM) to reconstruct subframe-level temporal continuity and align key dynamic features; a Squeeze-and-Excitation (SE) module is used to dynamically calibrate fused features, suppressing redundancy and highlighting discriminative channels; and Smooth Spatio-Temporal Contrastive Learning (SSTCL) is used as a regularization term to increase inter-class distance, reduce intra-class variance, and enhance the model’s discriminative ability. Experiments on the NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets demonstrate that the proposed method achieves state-of-the-art performance, validating the effectiveness of the integrated framework. The code is publicly available at <a href="https://github.com/zxqcsj/ATACR-Net.">https://github.com/zxqcsj/ATACR-Net.</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

Atacr-net: adaptive temporal alignment and contrastive refinement network for skeleton-based action recognition

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Liqiong Chen,
  • Xu Zeng,
  • Ming Zong

摘要

Graph Convolutional Networks (GCNs) have become the mainstream paradigm for skeleton-based action recognition, but existing methods face three core bottlenecks: (i) rigid discrete temporal convolution fails to adaptively model human actions with diverse rhythms and scales; (ii) multi-scale feature fusion causes severe channel redundancy and noise, masking key semantic features without adaptive channel calibration; (iii) insufficient discriminative constraints of cross-entropy loss lead to misclassification of visually similar actions. To address these issues, this paper proposes an Adaptive Temporal Alignment and Contrastive Refinement Network (ATACR-Net), following the design principle of "alignment-refinement-discrimination": a learnable continuous temporal sampling mechanism is introduced through a Multi-scale Adaptive Temporal Convolution Module (MATCM) to reconstruct subframe-level temporal continuity and align key dynamic features; a Squeeze-and-Excitation (SE) module is used to dynamically calibrate fused features, suppressing redundancy and highlighting discriminative channels; and Smooth Spatio-Temporal Contrastive Learning (SSTCL) is used as a regularization term to increase inter-class distance, reduce intra-class variance, and enhance the model’s discriminative ability. Experiments on the NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets demonstrate that the proposed method achieves state-of-the-art performance, validating the effectiveness of the integrated framework. The code is publicly available at https://github.com/zxqcsj/ATACR-Net.