Scholarlink

<p>We propose improved DiffusionDet, an enhanced framework that integrates diffusion models into object detection. By introducing a deformable sigmoid variance schedule and an adjustable sampling strategy, improved DiffusionDet achieves improved sampling efficiency and accuracy, demonstrated by a 6.8AP increase in the VOC dataset using 500 initial boxes and 50 timesteps. Our framework effectively captures sparse features, addressing limitations in low-ground-truth density scenarios. The improved sampling strategy makes our detect network superior ability to sampling small object boxes on image boundaries. This work addresses the disadvantage of over-reliance on random coverage when sampling detect boxes in DiffusionDet as well as contributing to the development of more robust and efficient diffusion models for object detection. The code is available at: <a href="https://github.com/tengfeifun/diffusiondet-v2">https://github.com/tengfeifun/diffusiondet-v2</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

Improved diffusion object detection through deformable sigmoid variance and adjustable sampling strategy

  • 期刊论文

  • 发表日期：

    2025年3月31日

查看全文
  • Teng Fei,
  • Tianle Huang,
  • Wuzhi Li,
  • Huankang Cui,
  • Guowei Zhang

摘要

We propose improved DiffusionDet, an enhanced framework that integrates diffusion models into object detection. By introducing a deformable sigmoid variance schedule and an adjustable sampling strategy, improved DiffusionDet achieves improved sampling efficiency and accuracy, demonstrated by a 6.8AP increase in the VOC dataset using 500 initial boxes and 50 timesteps. Our framework effectively captures sparse features, addressing limitations in low-ground-truth density scenarios. The improved sampling strategy makes our detect network superior ability to sampling small object boxes on image boundaries. This work addresses the disadvantage of over-reliance on random coverage when sampling detect boxes in DiffusionDet as well as contributing to the development of more robust and efficient diffusion models for object detection. The code is available at: https://github.com/tengfeifun/diffusiondet-v2.