Scholarlink

<p>Temporal action detection aims to classify and locate the actions in the video. However, the existing methods adopt a unified feature processing paradigm for the characteristics of the two sub-tasks, ignoring the mutual influence between two sub-tasks, which leads to a degradation in the performance of temporal action detection. To address this problem, we propose a Cross-Task Attention Alignment (CTAA) technique, which enhances the detection performance by specifically differentiating and refines the classification features and regression features. Specifically, in order to achieve effective separation and adaptive interaction between classification features and regression features, we design a novel Cross-task Refinement Module (CRM). CRM utilizes the attention mechanism to capture the intra-task feature correlations within each specific task and the cross-task connections between different tasks, thereby achieving fine feature separation. Subsequently, we further propose a Cross-channel Interaction Module (CIM), which integrates Mamba and Cross-channel Attention Fusion (CAF). It uses Mamba to capture the long-range temporal dependencies within each sub-channel, while our proposed CAF is designed to mine the fine-grained information within individual sub-channels and promote effective information exchange between sub-channels and the global channels. We conduct extensive experiments to validate our CTAA on four challenging datasets, i.e., THUMOS14, ActivityNet-1.3, MultiTHUMOS, and EPIC-Kitchens 100.</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

CTAA: Cross-Task Attention Alignment for temporal action detection

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Tianwei Qu,
  • Qixian Zhang,
  • Yuze Wang

摘要

Temporal action detection aims to classify and locate the actions in the video. However, the existing methods adopt a unified feature processing paradigm for the characteristics of the two sub-tasks, ignoring the mutual influence between two sub-tasks, which leads to a degradation in the performance of temporal action detection. To address this problem, we propose a Cross-Task Attention Alignment (CTAA) technique, which enhances the detection performance by specifically differentiating and refines the classification features and regression features. Specifically, in order to achieve effective separation and adaptive interaction between classification features and regression features, we design a novel Cross-task Refinement Module (CRM). CRM utilizes the attention mechanism to capture the intra-task feature correlations within each specific task and the cross-task connections between different tasks, thereby achieving fine feature separation. Subsequently, we further propose a Cross-channel Interaction Module (CIM), which integrates Mamba and Cross-channel Attention Fusion (CAF). It uses Mamba to capture the long-range temporal dependencies within each sub-channel, while our proposed CAF is designed to mine the fine-grained information within individual sub-channels and promote effective information exchange between sub-channels and the global channels. We conduct extensive experiments to validate our CTAA on four challenging datasets, i.e., THUMOS14, ActivityNet-1.3, MultiTHUMOS, and EPIC-Kitchens 100.