Scholarlink

<p>The recently proposed DEtection TRansformer (DETR) and its variants have achieved good performance in end-to-end object detection. However, these methods do not take into account the positional relationships between instances in the image and the importance of local feature information. To this end, this paper proposes an object detection method based on instance position and local feature enhancement, named PF-DETR. Firstly, a Dual Positional Encoding Attention (DPEA) is designed, which can simultaneously embed the absolute and relative position information of the object queries, thereby enhancing the localization ability of the queries. Secondly, we introduce a Local Feature Enhancement Module (LFEM), which makes full use of local information to enrich and refine object queries, further enhancing the queries’ perception of local features. Finally, a positional relation is established between the queries and the encoded image feature, and a Positional Interaction Cross-Attention (PICA) is proposed to allow the queries to find relevant regions in the encoded image feature more quickly. The experimental results on the MS COCO 2017 dataset show that the proposed method can effectively improve the convergence and performance of the model.</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

PF-DETR: Instance position and local feature enhancement for DETR

  • 期刊论文

  • 发表日期：

    2025年1月18日

查看全文
  • Xinfang Zhong,
  • Wenlan Kuang,
  • Zhixin Li

摘要

The recently proposed DEtection TRansformer (DETR) and its variants have achieved good performance in end-to-end object detection. However, these methods do not take into account the positional relationships between instances in the image and the importance of local feature information. To this end, this paper proposes an object detection method based on instance position and local feature enhancement, named PF-DETR. Firstly, a Dual Positional Encoding Attention (DPEA) is designed, which can simultaneously embed the absolute and relative position information of the object queries, thereby enhancing the localization ability of the queries. Secondly, we introduce a Local Feature Enhancement Module (LFEM), which makes full use of local information to enrich and refine object queries, further enhancing the queries’ perception of local features. Finally, a positional relation is established between the queries and the encoded image feature, and a Positional Interaction Cross-Attention (PICA) is proposed to allow the queries to find relevant regions in the encoded image feature more quickly. The experimental results on the MS COCO 2017 dataset show that the proposed method can effectively improve the convergence and performance of the model.