https://vilab.team/tag/%E8%BE%B9%E7%95%8C%E6%A1%86%E5%9B%9E%E5%BD%92/ 边界框回归 Hugo Blox Builder (https://hugoblox.com)en-usThu, 01 May 2025 00:00:00 +0000 https://vilab.team/media/icon_hu2896232876136423579.png https://vilab.team/tag/%E8%BE%B9%E7%95%8C%E6%A1%86%E5%9B%9E%E5%BD%92/ https://vilab.team/publication/d-fine-redefine-regression-task-of-detrs-as-fine-grained-dis/ Thu, 01 May 2025 00:00:00 +0000 https://vilab.team/publication/d-fine-redefine-regression-task-of-detrs-as-fine-grained-dis/ <p>D-FINE是一种实时目标检测器，通过重新定义DETR中的边界框回归任务实现高精度定位。其核心包含细粒度分布细化（FDR）和全局最优定位自蒸馏（GO-LSD）。FDR将回归从预测固定坐标改为迭代细化概率分布，提供细粒度中间表示；GO-LSD通过自蒸馏将定位知识从最终层传递到浅层，并简化深层残差预测。在COCO上达到54.0%/55.8% AP，124/78 FPS，预训练后达57.1%/59.3% AP，超越现有实时检测器，并显著提升多种DETR模型性能。</p> https://vilab.team/event/publication-news-06e581b9272339ee/ Thu, 01 May 2025 00:00:00 +0000 https://vilab.team/event/publication-news-06e581b9272339ee/ <p>热烈祝贺彭岩松同学！论文《D-FINE: Redefine regression task of DETRs as fine-grained distribution refinement》已发表在 <em>ICLR 2025</em>。</p> <h2 id="d-fine-redefine-regression-task-of-detrs-as-fine-grained-distribution-refinement">D-FINE: Redefine regression task of DETRs as fine-grained distribution refinement</h2> <p><strong>祝贺彭岩松同学！该论文已发表在 <em>ICLR</em>。</strong></p> <p> <figure > <div class="d-flex justify-content-center"> <div class="w-100" ><img alt="D-FINE: Redefine regression task of DETRs as fine-grained distribution refinement" srcset=" /event/publication-news-06e581b9272339ee/images/paper-01_hu14730101056268865597.webp 400w, /event/publication-news-06e581b9272339ee/images/paper-01_hu1666356483600992880.webp 760w, /event/publication-news-06e581b9272339ee/images/paper-01_hu15779176642512148333.webp 1200w" src="https://vilab.team/event/publication-news-06e581b9272339ee/images/paper-01_hu14730101056268865597.webp" width="760" height="314" loading="lazy" data-zoomable /></div> </div></figure> </p> <ul> <li><strong>作者：</strong> Yansong Peng、Hebei Li、Peixi Wu、Yueyi Zhang、Xiaoyan Sun、Feng Wu</li> <li><strong>发表载体：</strong> <em>ICLR</em></li> <li><strong>发表时间：</strong> 2025年5月1日</li> <li><strong>相关链接：</strong> <a href="https://proceedings.iclr.cc/paper_files/paper/2025/hash/6cf58a87e3097e7d1f9be3e8693a93de-Abstract-Conference.html" target="_blank" rel="noopener">论文链接</a> · <a href="https://proceedings.iclr.cc/paper_files/paper/2025/file/6cf58a87e3097e7d1f9be3e8693a93de-Paper-Conference.pdf" target="_blank" rel="noopener">PDF</a> · <a href="https://github.com/Peterande/D-FINE" target="_blank" rel="noopener">代码</a></li> </ul> <h3 id="论文介绍">论文介绍</h3> <p>D-FINE是一种实时目标检测器，通过重新定义DETR中的边界框回归任务实现高精度定位。其核心包含细粒度分布细化（FDR）和全局最优定位自蒸馏（GO-LSD）。FDR将回归从预测固定坐标改为迭代细化概率分布，提供细粒度中间表示；GO-LSD通过自蒸馏将定位知识从最终层传递到浅层，并简化深层残差预测。在COCO上达到54.0%/55.8% AP，124/78 FPS，预训练后达57.1%/59.3% AP，超越现有实时检测器，并显著提升多种DETR模型性能。</p>