Scholarlink

<p>Temporal action detection (TAD) in untrimmed videos requires both effective long-range temporal dependency modeling and accurate action boundary localization. Although Vision Transformer (ViT)-based methods have shown strong representation capabilities, they often lack explicit mechanisms for boundary-sensitive temporal interaction and fine-grained boundary modeling. To address these issues, we propose Effi-TAD, an efficient temporal action detection framework built upon a plain ViT backbone. Specifically, we introduce a Boundary-Aware Temporal Feature Interaction Module (BA-TFIM) to refine post-backbone temporal features by integrating local boundary-aware residual cues with global temporal interaction. Based on the refined representations, we further design a Tri-Branch Boundary-Aware Head (TriBE Head), which decouples start-boundary prediction, end-boundary prediction, and center-offset estimation into three specialized branches for more accurate temporal localization. In addition, a localization-aware classification objective is incorporated to improve the consistency between classification confidence and localization quality. Extensive experiments on four benchmark datasets, including THUMOS14, ActivityNet−1.3, Multi-THUMOS, and Charades, demonstrate the effectiveness and generalization capability of the proposed framework. With the VideoMAE-B backbone, Effi-TAD achieves 72.4% average mAP on THUMOS14, 38.99% on ActivityNet−1.3, 44.3% on Multi-THUMOS, and 28.9% on Charades. When equipped with the larger VideoMAE-L backbone, Effi-TAD further improves the average mAP on THUMOS14 to 74.0%. The code is available at <a href="https://github.com/guojiayi1209/Effi-TAD">https://github.com/guojiayi1209/Effi-TAD</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

Effi-TAD: efficient temporal action detection via temporal interaction and boundary-aware modeling

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Jiayi Guo,
  • Yumeng Wang,
  • Kexin Ma,
  • Zongfang Ma

摘要

Temporal action detection (TAD) in untrimmed videos requires both effective long-range temporal dependency modeling and accurate action boundary localization. Although Vision Transformer (ViT)-based methods have shown strong representation capabilities, they often lack explicit mechanisms for boundary-sensitive temporal interaction and fine-grained boundary modeling. To address these issues, we propose Effi-TAD, an efficient temporal action detection framework built upon a plain ViT backbone. Specifically, we introduce a Boundary-Aware Temporal Feature Interaction Module (BA-TFIM) to refine post-backbone temporal features by integrating local boundary-aware residual cues with global temporal interaction. Based on the refined representations, we further design a Tri-Branch Boundary-Aware Head (TriBE Head), which decouples start-boundary prediction, end-boundary prediction, and center-offset estimation into three specialized branches for more accurate temporal localization. In addition, a localization-aware classification objective is incorporated to improve the consistency between classification confidence and localization quality. Extensive experiments on four benchmark datasets, including THUMOS14, ActivityNet−1.3, Multi-THUMOS, and Charades, demonstrate the effectiveness and generalization capability of the proposed framework. With the VideoMAE-B backbone, Effi-TAD achieves 72.4% average mAP on THUMOS14, 38.99% on ActivityNet−1.3, 44.3% on Multi-THUMOS, and 28.9% on Charades. When equipped with the larger VideoMAE-L backbone, Effi-TAD further improves the average mAP on THUMOS14 to 74.0%. The code is available at https://github.com/guojiayi1209/Effi-TAD.