Scholarlink

<p>Few-shot object detection (FSOD) aims to detect novel categories using only a small number of annotated samples. Meta-learning-based FSOD methods usually adopt a dual-branch framework, where the support branch extracts class-related prototypes and the query branch uses these prototypes to improve detection. However, the effectiveness of this paradigm depends heavily on the interaction and alignment between the support and query features. To strengthen cross-branch feature interaction, we propose a Decoupled-Projection Prototype Alignment (DPA) module and a Geometric Alignment (GA) module. DPA performs mid-level spatial alignment between Stage 1–3 query features and fine-grained support prototypes by using decoupled projections and a bounded residual gate. GA further calibrates high-level RoI features by softly shrinking prototype-inconsistent residual components. In addition, focal loss is introduced to improve classification learning under few-shot class imbalance. Experiments on PASCAL VOC and COCO demonstrate that the proposed method achieves competitive performance against representative FSOD methods.</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

Few-shot object detection via dual-feature alignment

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Dike Chen,
  • Zongyuan Ding,
  • Ji Zhang,
  • Hongyuan Wang,
  • Boyu Wang

摘要

Few-shot object detection (FSOD) aims to detect novel categories using only a small number of annotated samples. Meta-learning-based FSOD methods usually adopt a dual-branch framework, where the support branch extracts class-related prototypes and the query branch uses these prototypes to improve detection. However, the effectiveness of this paradigm depends heavily on the interaction and alignment between the support and query features. To strengthen cross-branch feature interaction, we propose a Decoupled-Projection Prototype Alignment (DPA) module and a Geometric Alignment (GA) module. DPA performs mid-level spatial alignment between Stage 1–3 query features and fine-grained support prototypes by using decoupled projections and a bounded residual gate. GA further calibrates high-level RoI features by softly shrinking prototype-inconsistent residual components. In addition, focal loss is introduced to improve classification learning under few-shot class imbalance. Experiments on PASCAL VOC and COCO demonstrate that the proposed method achieves competitive performance against representative FSOD methods.