Scholarlink

<p>Transformer-based trackers achieve strong performance in visual object tracking by modeling long-range dependencies via multi-head attention (MHA). However, the global scope of MHA often blurs boundaries and introduces false associations with background regions. To overcome this, we propose LATrack—a locality-enhanced Transformer framework that structurally integrates two novel modules: Criss-Cross Attention (CCA) and Correlative Masked Attention (CMA). CCA aggregates horizontal and vertical features to sharpen spatial edges, while CMA enforces continuity constraints to suppress mismatches. Unlike prior works that independently apply attention variants, LATrack unifies both modules within a Siamese architecture, enabling coordinated local–global representation learning. Extensive experiments on nine benchmarks demonstrate that LATrack achieves state-of-the-art accuracy and real-time speed, especially under occlusion, deformation, and fast motion.</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

LATrack: enhancing transformer visual tracking via locality-aware attention

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Zhigang Shi,
  • Zhongyi Huang,
  • Zhiming Fang,
  • Feng Tang

摘要

Transformer-based trackers achieve strong performance in visual object tracking by modeling long-range dependencies via multi-head attention (MHA). However, the global scope of MHA often blurs boundaries and introduces false associations with background regions. To overcome this, we propose LATrack—a locality-enhanced Transformer framework that structurally integrates two novel modules: Criss-Cross Attention (CCA) and Correlative Masked Attention (CMA). CCA aggregates horizontal and vertical features to sharpen spatial edges, while CMA enforces continuity constraints to suppress mismatches. Unlike prior works that independently apply attention variants, LATrack unifies both modules within a Siamese architecture, enabling coordinated local–global representation learning. Extensive experiments on nine benchmarks demonstrate that LATrack achieves state-of-the-art accuracy and real-time speed, especially under occlusion, deformation, and fast motion.