Scholarlink

<p>Real-time object detection is of significance in many applications such as autonomous driving, security monitoring, and defect detection. The RT-DETR combines a CNN-based encoder with a Transformer-based decoder, achieving great performance. This paper proposes a lightweighted RT-DETR, LRT-DETR in short, with fewer parameters and higher FPS. By integrating the backbone of YOLOv7-tiny, the number of parameters is significantly reduced while improving detection speed. In the encoder, the introduction of the proposed SimpleUpsample operator mitigates noise caused by upsampling, and furthermore, a lightweight module RepFormer is designed to effectively enhancing feature extraction capability of it. Comparison experiments with the RT-DETR-R18 on the MS COCO2017 dataset demonstrate that there is merely 1% decrease in accuracy, at 44.6%, while the total parameter is just 12&#xa0;M, and the detection speed is up to 370FPS on a consumer-grade GPU (NVIDIA GeForce RTX 3090).</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

Lrt-detr: a lightweighted RT-DETR with RepFormer and SimpleUpsample for fast object detection

  • 期刊论文

  • 发表日期：

    2025年5月27日

查看全文
  • Aohui Li,
  • Ke Cheng,
  • Xuecheng Hua,
  • Yuanquan Wang

摘要

Real-time object detection is of significance in many applications such as autonomous driving, security monitoring, and defect detection. The RT-DETR combines a CNN-based encoder with a Transformer-based decoder, achieving great performance. This paper proposes a lightweighted RT-DETR, LRT-DETR in short, with fewer parameters and higher FPS. By integrating the backbone of YOLOv7-tiny, the number of parameters is significantly reduced while improving detection speed. In the encoder, the introduction of the proposed SimpleUpsample operator mitigates noise caused by upsampling, and furthermore, a lightweight module RepFormer is designed to effectively enhancing feature extraction capability of it. Comparison experiments with the RT-DETR-R18 on the MS COCO2017 dataset demonstrate that there is merely 1% decrease in accuracy, at 44.6%, while the total parameter is just 12 M, and the detection speed is up to 370FPS on a consumer-grade GPU (NVIDIA GeForce RTX 3090).