Scholarlink

<p>Temporal Action Localization is a critical task in behavior understanding, aiming to precisely localize start and end timestamps and identify action categories within untrimmed videos. While ActionFormer achieves end-to-end action localization via anchor-free boundary regression, it overlooks the complex periodic and aperiodic dynamic patterns inherent in actions, thereby limiting potential recognition performance. To address this limitation, we propose a periodicity-modeling-based method for temporal action detection. First, we design a Residual Cycle Refinement (RCR) module that utilizes backpropagation to optimize a learnable period matrix for extracting periodic patterns, while retaining aperiodic components in the residual branch, thus achieving effective separation of periodic and aperiodic features. Second, we introduce a Multi-Scale Cycle Fusion (MSCF) module, which employs depth-wise separable convolutions to capture periodic features across various temporal scales. Finally, we design an Asymmetric DIoU (A-DIoU) loss function that assigns higher penalty weights to ending boundaries, enabling more precise action boundary localization. Experimental results demonstrate that the proposed method significantly outperforms the baseline ActionFormer on the THUMOS14 and ActivityNet-1.3 datasets, achieving average mAPs of 67.95% and 36.71%, respectively, while maintaining comparable parameter counts and computational complexity.</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

A novel residual cycle refinement framework for temporal action detection(RCR-TAD)

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Fuhua Chen,
  • Dapeng Jiang,
  • Jinhao Chen,
  • Yizhuo Zhang

摘要

Temporal Action Localization is a critical task in behavior understanding, aiming to precisely localize start and end timestamps and identify action categories within untrimmed videos. While ActionFormer achieves end-to-end action localization via anchor-free boundary regression, it overlooks the complex periodic and aperiodic dynamic patterns inherent in actions, thereby limiting potential recognition performance. To address this limitation, we propose a periodicity-modeling-based method for temporal action detection. First, we design a Residual Cycle Refinement (RCR) module that utilizes backpropagation to optimize a learnable period matrix for extracting periodic patterns, while retaining aperiodic components in the residual branch, thus achieving effective separation of periodic and aperiodic features. Second, we introduce a Multi-Scale Cycle Fusion (MSCF) module, which employs depth-wise separable convolutions to capture periodic features across various temporal scales. Finally, we design an Asymmetric DIoU (A-DIoU) loss function that assigns higher penalty weights to ending boundaries, enabling more precise action boundary localization. Experimental results demonstrate that the proposed method significantly outperforms the baseline ActionFormer on the THUMOS14 and ActivityNet-1.3 datasets, achieving average mAPs of 67.95% and 36.71%, respectively, while maintaining comparable parameter counts and computational complexity.