Scholarlink

<p>Fashion item detection is challenging due to ambiguities caused by diverse item appearances and similarities among subcategories. To address this, we propose <i>Holi-DETR</i>, a <b>Holi</b>stic <b>De</b>tection <b>Tr</b>ansformer that detects fashion items by leveraging contextual information. In contrast to conventional detectors that treat each item independently, Holi-DETR alleviates ambiguity by incorporating three types of contextual cues that capture inter-item relationships: (1) inter-item co-occurrence relationship, (2) relative positions and sizes based on inter-item spatial arrangements, and (3) spatial relations between items and human body keypoints. Holi-DETR integrates these heterogeneous cues into the Detection Transformer (DETR) architecture in a learnable manner. Experimental results demonstrate that the proposed methods improve the average precision (AP) by 3.6 pp over DETR and 1.1 pp over Co-DETR.</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

Holi-DETR: holistic fashion item detection leveraging contextual information

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Youngchae Kwon,
  • Jinyoung Choi,
  • Injung Kim

摘要

Fashion item detection is challenging due to ambiguities caused by diverse item appearances and similarities among subcategories. To address this, we propose Holi-DETR, a Holistic Detection Transformer that detects fashion items by leveraging contextual information. In contrast to conventional detectors that treat each item independently, Holi-DETR alleviates ambiguity by incorporating three types of contextual cues that capture inter-item relationships: (1) inter-item co-occurrence relationship, (2) relative positions and sizes based on inter-item spatial arrangements, and (3) spatial relations between items and human body keypoints. Holi-DETR integrates these heterogeneous cues into the Detection Transformer (DETR) architecture in a learnable manner. Experimental results demonstrate that the proposed methods improve the average precision (AP) by 3.6 pp over DETR and 1.1 pp over Co-DETR.