Scholarlink

<p>Temporal Sentence Grounding (TSG) is a challenging task in computer vision, which aims to localize the temporal span in untrimmed videos that corresponds to a natural language description. Recent approaches inspired by the Detection Transformer (DETR) paradigm decode a fixed set of learnable queries to predict temporal proposals, and establish supervision between proposals and ground truths via Hungarian matching. The absence of explicit task guidance often leads queries to produce overlapped and redundant proposals, blurring query responsibilities and impeding precise localization. To address this limitation, we focus on improving current training pipeline, proposing a selection-and-assignment strategy to determine which proposals actually contribute to the model training. This design encourage role specialization among learnable queries. Furthermore, a proposal suppression strategy is employed prior to matching to reduce redundancy among proposals. Our method achieves state-of-the-art results on three benchmarks. Extensive ablation studies demonstrate the effectiveness of our position-based proposal selection and anchor-based proposal assignment strategies.</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

Select and assign: anchor-guided proposals for temporal sentence grounding

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Ziyi Liu,
  • Jiawei Wang,
  • Ziyue Zou,
  • Qiang Zeng,
  • Xiaolong Sun,
  • Hongmin Liu

摘要

Temporal Sentence Grounding (TSG) is a challenging task in computer vision, which aims to localize the temporal span in untrimmed videos that corresponds to a natural language description. Recent approaches inspired by the Detection Transformer (DETR) paradigm decode a fixed set of learnable queries to predict temporal proposals, and establish supervision between proposals and ground truths via Hungarian matching. The absence of explicit task guidance often leads queries to produce overlapped and redundant proposals, blurring query responsibilities and impeding precise localization. To address this limitation, we focus on improving current training pipeline, proposing a selection-and-assignment strategy to determine which proposals actually contribute to the model training. This design encourage role specialization among learnable queries. Furthermore, a proposal suppression strategy is employed prior to matching to reduce redundancy among proposals. Our method achieves state-of-the-art results on three benchmarks. Extensive ablation studies demonstrate the effectiveness of our position-based proposal selection and anchor-based proposal assignment strategies.