Scholarlink

<p>Deep convolutional network developments have contributed to the widespread application of object detection algorithms in daily life. Furthermore, the most recent models have achieved high accuracy levels. However, accurate detection remains a challenge. This challenge arises from the abundance of small objects and the significant scale disparities between different classes of objects in traffic settings. In this paper, we present SOD-GC, an innovative architecture for improving the detection accuracy of small objects. Firstly, an entirely new backbone network called Global Context Network (GCNet) is developed. The shortcomings of traditional backbone networks in managing global contextual information are addressed by GCNet. Secondly, we employ Convolutional Feature Space Pyramid (CFSP). By utilizing convolutions with smaller dilation rates, CFSP focuses on capturing local textures and edge details. Finally, the multi-scale feature fusion recalibration component (MFRC) is introduced. It focuses on recalibrating the feature weights according to the resolution and content of different scale feature maps. The capacity of shallow and deep features to interact is enhanced. GCNet provides a global perspective, while CFSP enhances local details. MFRC performs multi-scale fusion and recalibrating the feature weights extracted by GCNet and the local information focused on by CFSP. These three components collaboratively provide a significant enhancement over existing detectors, as demonstrated by experimental results. SOD-GC achieves a mAP@0.5 of 94.02% on the KITTI dataset, exceeding the performance of baseline by 4.80%. Meanwhile, SOD-GC is 3.36% higher than baseline in mAP@0.5 on the COCO-traffic dataset. The code is available at: <a href="https://github.com/Lus-hub/SOD-GC.">https://github.com/Lus-hub/SOD-GC.</a></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

SOD-GC: Small object detection with global context network and multi-scale feature fusion recalibration

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Zhong Qu,
  • Shuang Lu,
  • Shufang Xia

摘要

Deep convolutional network developments have contributed to the widespread application of object detection algorithms in daily life. Furthermore, the most recent models have achieved high accuracy levels. However, accurate detection remains a challenge. This challenge arises from the abundance of small objects and the significant scale disparities between different classes of objects in traffic settings. In this paper, we present SOD-GC, an innovative architecture for improving the detection accuracy of small objects. Firstly, an entirely new backbone network called Global Context Network (GCNet) is developed. The shortcomings of traditional backbone networks in managing global contextual information are addressed by GCNet. Secondly, we employ Convolutional Feature Space Pyramid (CFSP). By utilizing convolutions with smaller dilation rates, CFSP focuses on capturing local textures and edge details. Finally, the multi-scale feature fusion recalibration component (MFRC) is introduced. It focuses on recalibrating the feature weights according to the resolution and content of different scale feature maps. The capacity of shallow and deep features to interact is enhanced. GCNet provides a global perspective, while CFSP enhances local details. MFRC performs multi-scale fusion and recalibrating the feature weights extracted by GCNet and the local information focused on by CFSP. These three components collaboratively provide a significant enhancement over existing detectors, as demonstrated by experimental results. SOD-GC achieves a mAP@0.5 of 94.02% on the KITTI dataset, exceeding the performance of baseline by 4.80%. Meanwhile, SOD-GC is 3.36% higher than baseline in mAP@0.5 on the COCO-traffic dataset. The code is available at: https://github.com/Lus-hub/SOD-GC.