Scholarlink

<p>Balancing computational efficiency with fine-grained discrimination remains a persistent bottleneck in remote sensing object detection. While emerging efficient DETR architectures mitigate overhead via interleaved updates, this asynchronous propagation risks context leakage, where global noise obscures high-frequency details. Concurrently, standard absolute positional encodings lack the relative geometric precision to resolve dense clusters, leading to query collapse. We propose AGLG-DINO, a unified framework addressing these limitations. First, the Adaptive Global–Local Gating (AGLG) module functions as a semantic filter, suppressing background clutter by dynamically conditioning global context injection on pixel-wise relevance. Second, 2D Mixed-Rotary Positional Embeddings (RoPE) are integrated into the decoder to enforce relative spatial constraints, effectively disambiguating closely packed instances. Extensive experiments on DIOR and RSOD benchmarks demonstrate that AGLG-DINO establishes a new state-of-the-art efficiency-accuracy trade-off, significantly outperforming baselines with negligible computational overhead.</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

AGLG-DINO: adaptive global–local gating and geometry-aware efficient DETR for remote sensing object detection

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Shoubin Wang,
  • Zefeng Ding,
  • Guili Peng,
  • Pengcheng Gao,
  • Wenjie Zhu,
  • Youbing Li,
  • Lewei Jing,
  • Xinchang Fang

摘要

Balancing computational efficiency with fine-grained discrimination remains a persistent bottleneck in remote sensing object detection. While emerging efficient DETR architectures mitigate overhead via interleaved updates, this asynchronous propagation risks context leakage, where global noise obscures high-frequency details. Concurrently, standard absolute positional encodings lack the relative geometric precision to resolve dense clusters, leading to query collapse. We propose AGLG-DINO, a unified framework addressing these limitations. First, the Adaptive Global–Local Gating (AGLG) module functions as a semantic filter, suppressing background clutter by dynamically conditioning global context injection on pixel-wise relevance. Second, 2D Mixed-Rotary Positional Embeddings (RoPE) are integrated into the decoder to enforce relative spatial constraints, effectively disambiguating closely packed instances. Extensive experiments on DIOR and RSOD benchmarks demonstrate that AGLG-DINO establishes a new state-of-the-art efficiency-accuracy trade-off, significantly outperforming baselines with negligible computational overhead.