Scholarlink

<p>Real-time recognition of instantaneous student behaviors in cluttered environments presents a significant challenge for existing object detectors, which often fail to balance accuracy with computational efficiency. This paper introduces HawkEye, a lightweight detector designed to resolve this accuracy-efficiency trade-off. HawkEye’s architecture integrates two key innovations: a Global Edge Fusion Enhancer (GEFE) to counteract occlusion by recovering structural features, and a Task-Aligned Offset-Aware Head (TAOAH) to precisely localize small, non-rigid objects. Following model training with EMASlideLoss, a final optimized model, HawkEye-P, was produced via structured LAMP pruning. HawkEye-P establishes a new state-of-the-art on the SCB5-A benchmark, surpassing its baseline with a 4.6% relative improvement in mAP@50. Crucially, this accuracy gain is achieved with a massive leap in efficiency: the pruning process reduces model parameters by 61.1% and FLOPs by 50.6%, resulting in an 18.3% faster inference speed (168 FPS). These concurrent gains in both accuracy and efficiency validate HawkEye-P as a compelling framework with significant potential for deployment 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

HawkEye: a lightweight edge-aware detector for classroom behavior recognition

  • 期刊论文

  • 发表日期：

    2026年1月11日

查看全文
  • Minggen Xie,
  • Zhiwen Ran,
  • Di Chen,
  • Renzhang Chen,
  • Zonglin Liu,
  • Zhongchun He,
  • Quanlong Guan,
  • Zhanchuan Cai

摘要

Real-time recognition of instantaneous student behaviors in cluttered environments presents a significant challenge for existing object detectors, which often fail to balance accuracy with computational efficiency. This paper introduces HawkEye, a lightweight detector designed to resolve this accuracy-efficiency trade-off. HawkEye’s architecture integrates two key innovations: a Global Edge Fusion Enhancer (GEFE) to counteract occlusion by recovering structural features, and a Task-Aligned Offset-Aware Head (TAOAH) to precisely localize small, non-rigid objects. Following model training with EMASlideLoss, a final optimized model, HawkEye-P, was produced via structured LAMP pruning. HawkEye-P establishes a new state-of-the-art on the SCB5-A benchmark, surpassing its baseline with a 4.6% relative improvement in mAP@50. Crucially, this accuracy gain is achieved with a massive leap in efficiency: the pruning process reduces model parameters by 61.1% and FLOPs by 50.6%, resulting in an 18.3% faster inference speed (168 FPS). These concurrent gains in both accuracy and efficiency validate HawkEye-P as a compelling framework with significant potential for deployment in resource-constrained environments.