Scholarlink

<p>Skeleton-based human action recognition (HAR) identifies actions by capturing joint motions and posture changes, and it has broad applications. Despite promising results, current methods still face challenges. Over-smoothing and high-frequency loss in the CNN and GCN methods limit the ability to differentiate similar actions due to a lack of fine-grained features, leading to hard samples. To address the above issues, we propose DPConv3D, a dual-pathway relearning framework based on 3D CNNs. The benchmark pathway learns general action representations and establishes a discriminator for hard samples using latent knowledge. In contrast, the relearning pathway incorporates multi-loss and joint spatiotemporal attention (JSTA) to enhance fine-grained feature learning. DPConv3D combines general and fine-grained features to enhance HAR performance and further improves it by fusing joint and limb streams. Notably, fine-tuning the relearning pathway alone enables the baseline to enhance fine-grained HAR. We evaluate DPConv3D on four public datasets: FineGYM, HMDB51, NTU RGB+D, and NTU RGB+D 120, demonstrating its effectiveness in enhancing fine-grained HAR while achieving state-of-the-art results on the first three benchmarks. The code is publicly available at <a href="https://github.com/Xie-Hao0818/arskl">https://github.com/Xie-Hao0818/arskl</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

Enhanced fine-grained relearning for skeleton-based action recognition

  • 期刊论文

  • 发表日期：

    2025年3月22日

查看全文
  • Zhenghao Xie,
  • Junfen Chen,
  • Yingying Wang,
  • Bojun Xie

摘要

Skeleton-based human action recognition (HAR) identifies actions by capturing joint motions and posture changes, and it has broad applications. Despite promising results, current methods still face challenges. Over-smoothing and high-frequency loss in the CNN and GCN methods limit the ability to differentiate similar actions due to a lack of fine-grained features, leading to hard samples. To address the above issues, we propose DPConv3D, a dual-pathway relearning framework based on 3D CNNs. The benchmark pathway learns general action representations and establishes a discriminator for hard samples using latent knowledge. In contrast, the relearning pathway incorporates multi-loss and joint spatiotemporal attention (JSTA) to enhance fine-grained feature learning. DPConv3D combines general and fine-grained features to enhance HAR performance and further improves it by fusing joint and limb streams. Notably, fine-tuning the relearning pathway alone enables the baseline to enhance fine-grained HAR. We evaluate DPConv3D on four public datasets: FineGYM, HMDB51, NTU RGB+D, and NTU RGB+D 120, demonstrating its effectiveness in enhancing fine-grained HAR while achieving state-of-the-art results on the first three benchmarks. The code is publicly available at https://github.com/Xie-Hao0818/arskl.