Scholarlink

<p>Goggles detection is a new research topic with important academic and application value. It further distinguishes whether people wear goggles correctly on the basis of face detection. There are two major challenges in the research of goggles detection, which are the lack of large-scale datasets and the high requirement of model generalization since the styles of goggles vary greatly. In this paper, a large-scale and high-quality goggles dataset is released to deal with the former problem and facilitate future research, named as GogglesDet2023. For the latter problem, a plug-and-play Clustering Transformer Module (CTM) is proposed to alleviate the inadequacy of model generalization caused by the intra-class diversity and the inter-class similarity between goggles and ordinary glasses. We have applied the Clustering Transformer Module to several typical detectors, such as YOLOX, YOLOv11, Faster R-CNN and Deformable-DETR. Extensive experiments have proved that the proposed method can effectively reduce confusion over categories, improve model generalization, and eventually achieve superior performance on the GogglesDet2023 dataset. The code and the dataset are public available at <a href="https://github.com/WCUSTC/GogglesDET">https://github.com/WCUSTC/GogglesDET</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

Gogglesdet: goggles detection with clustering transformer and a new dataset

  • 期刊论文

  • 发表日期：

    2025年1月3日

查看全文
  • Chong Wang,
  • Zhuozhi Cheng,
  • Adeel Akram,
  • Qixing Zhang,
  • Yongming Zhang

摘要

Goggles detection is a new research topic with important academic and application value. It further distinguishes whether people wear goggles correctly on the basis of face detection. There are two major challenges in the research of goggles detection, which are the lack of large-scale datasets and the high requirement of model generalization since the styles of goggles vary greatly. In this paper, a large-scale and high-quality goggles dataset is released to deal with the former problem and facilitate future research, named as GogglesDet2023. For the latter problem, a plug-and-play Clustering Transformer Module (CTM) is proposed to alleviate the inadequacy of model generalization caused by the intra-class diversity and the inter-class similarity between goggles and ordinary glasses. We have applied the Clustering Transformer Module to several typical detectors, such as YOLOX, YOLOv11, Faster R-CNN and Deformable-DETR. Extensive experiments have proved that the proposed method can effectively reduce confusion over categories, improve model generalization, and eventually achieve superior performance on the GogglesDet2023 dataset. The code and the dataset are public available at https://github.com/WCUSTC/GogglesDET.