Scholarlink

<p>Temporal Action Detection (TAD) aims to identify and locate all action instances and their boundaries in untrimmed videos. Despite the various network architectures proposed by researchers in recent years to improve the modeling effect of TAD, existing methods still have significant limitations: the features obtained by the Transformer’s attention mechanism are treated equally; most current models fail to fully utilize the historical information in the changes of action boundaries; existing methods for capturing channel correlations increase complexity significantly while enhancing robustness. To address these issues, we introduce ECMFC-Net, which integrates a pyramid structure to combine multi-scale features. The key component is the GlobalLocal Fusion Block (GLF), which adopts a parallel dual-branch architecture. There are three key improvements: ❶The Cross-axis Attention Gate (CAG) module dynamically fuses the channel and spatial attention branches, adaptively selecting and fusing the features obtained by causal attention. ❷The causal attention module combines temporal flipping and causal masking multi-head self-attention to explicitly model unidirectional temporal dependencies, allowing the model to simultaneously learn forward and backward temporal dependencies. ❸The Feature Consensus Net (FCN) uses a lightweight aggregation and redistribution mechanism combined with self-gating to efficiently model channel dependencies, enhancing module robustness with only linear complexity. ECMFC-Net outperforms many state-of-the-art methods on five challenging benchmark datasets (THUMOS14, MultiTHUMOS, EPIC-KITCHENS-100, ActivityNet-1.3, and FineAction).</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

ECMFC-Net:An enhanced causal modeling and feature consensus framework for temporal action detection

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Liming Guan,
  • Hongjie Zhang,
  • Haiping Zhang,
  • Xujian Fang

摘要

Temporal Action Detection (TAD) aims to identify and locate all action instances and their boundaries in untrimmed videos. Despite the various network architectures proposed by researchers in recent years to improve the modeling effect of TAD, existing methods still have significant limitations: the features obtained by the Transformer’s attention mechanism are treated equally; most current models fail to fully utilize the historical information in the changes of action boundaries; existing methods for capturing channel correlations increase complexity significantly while enhancing robustness. To address these issues, we introduce ECMFC-Net, which integrates a pyramid structure to combine multi-scale features. The key component is the GlobalLocal Fusion Block (GLF), which adopts a parallel dual-branch architecture. There are three key improvements: ❶The Cross-axis Attention Gate (CAG) module dynamically fuses the channel and spatial attention branches, adaptively selecting and fusing the features obtained by causal attention. ❷The causal attention module combines temporal flipping and causal masking multi-head self-attention to explicitly model unidirectional temporal dependencies, allowing the model to simultaneously learn forward and backward temporal dependencies. ❸The Feature Consensus Net (FCN) uses a lightweight aggregation and redistribution mechanism combined with self-gating to efficiently model channel dependencies, enhancing module robustness with only linear complexity. ECMFC-Net outperforms many state-of-the-art methods on five challenging benchmark datasets (THUMOS14, MultiTHUMOS, EPIC-KITCHENS-100, ActivityNet-1.3, and FineAction).