Scholarlink

<p>Small object detection (SOD) remains a critical yet unresolved challenge in domains such as aerial monitoring, traffic management, and autonomous navigation, where targets occupy only a few pixels and real-time inference is essential. Conventional strategies often rely on uniform scaling or ad hoc module additions, which increase computational cost without guaranteeing optimal accuracy–latency trade-offs. In this work, we present a systematic study of <i>computational resource reallocation</i> across the backbone, neck, and detection head of lightweight YOLOv11 models. By selectively widening early backbone stages, strengthening high-resolution fusion pathways, and integrating a P2/4-scale detection head, we establish principled guidelines for allocating limited FLOPs to maximize both detection performance and execution efficiency. Extensive experiments on VisDrone and multiple cross-domain benchmarks demonstrate consistent improvements in the mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation>–GFLOPs Pareto frontier while reducing per-frame latency by up to 23% under identical hardware conditions. These findings confirm that <i>where</i> computation is placed—rather than how much—is crucial for real-time SOD, and that strategic redistribution of computation outperforms uniform scaling and module-based approaches. This study provides both empirical insights and hardware-aware design principles for efficient, low-latency small-object detection in resource-constrained environments.</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

A systematic study of resource allocation for real-time small-object detection

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Dahee Lim,
  • Minjae Kim,
  • Jihun Park

摘要

Small object detection (SOD) remains a critical yet unresolved challenge in domains such as aerial monitoring, traffic management, and autonomous navigation, where targets occupy only a few pixels and real-time inference is essential. Conventional strategies often rely on uniform scaling or ad hoc module additions, which increase computational cost without guaranteeing optimal accuracy–latency trade-offs. In this work, we present a systematic study of computational resource reallocation across the backbone, neck, and detection head of lightweight YOLOv11 models. By selectively widening early backbone stages, strengthening high-resolution fusion pathways, and integrating a P2/4-scale detection head, we establish principled guidelines for allocating limited FLOPs to maximize both detection performance and execution efficiency. Extensive experiments on VisDrone and multiple cross-domain benchmarks demonstrate consistent improvements in the mAP \(_{50}\) 50 –GFLOPs Pareto frontier while reducing per-frame latency by up to 23% under identical hardware conditions. These findings confirm that where computation is placed—rather than how much—is crucial for real-time SOD, and that strategic redistribution of computation outperforms uniform scaling and module-based approaches. This study provides both empirical insights and hardware-aware design principles for efficient, low-latency small-object detection in resource-constrained environments.