https://zhaoylee.github.io/Blogs/ Recent content on zhaoyli's Blog Hugo zh-cn [©2024 zhaoyli&rsquo;s Blog] https://zhaoylee.github.io/) Mon, 16 Mar 2026 01:45:51 +0000 https://zhaoylee.github.io/Blogs/posts/plug_and_play/ocm3d--object-centric-monocular-3d-object-detection/ Mon, 16 Mar 2026 09:12:18 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/ocm3d--object-centric-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：arxiv<br> <strong>📅 发表年份</strong>：2021<br> <strong>💻 开源代码</strong>：<a href="https://github.com/mrsempress/OBMO_GUPNet/blob/main/tools/offline_OBMO.py">OBMO_GUPNet</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2104.06041">OCM3D: Object-Centric Monocular 3D Object Detection</a></p> </blockquote> <hr> <h3 id="1-文献背景研究目的与核心问题">1. 文献背景、研究目的与核心问题</h3> <ul> <li> <p><strong>研究背景</strong>：单目 3D 目标检测（Monocular 3D Object Detection）是一个高度病态（ill-posed）的问题。主流方法通常依赖纯图像或将其转化为伪激光雷达（Pseudo-LiDAR）点云。然而，前者难以捕捉像素间的 3D 空间几何关系，后者则受困于单目深度估计带来的巨大点云噪声。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/lr3d--improving-distant-3d-object-detection-using-2d-box-supervision/ Sun, 15 Mar 2026 22:23:00 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/lr3d--improving-distant-3d-object-detection-using-2d-box-supervision/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：CVPR<br> <strong>📅 发表年份</strong>：2024<br> <strong>💻 开源代码</strong>：<a href="%E5%A1%AB%E5%86%99%E4%BD%A0%E7%9A%84URL">无</a><br> <strong>📄 论文题目</strong>：<a href="https://openaccess.thecvf.com/content/CVPR2024/papers/Yang_Improving_Distant_3D_Object_Detection_Using_2D_Box_Supervision_CVPR_2024_paper.pdf">Improving Distant 3D Object Detection Using 2D Box Supervision</a></p> </blockquote> <hr> <p>这篇由 NVIDIA 等机构的研究人员发表在 CVPR 2024 的重磅论文 <strong>《Improving Distant 3D Object Detection Using 2D Box Supervision》(简称 LR3D)</strong>，切入了一个目前高阶自动驾驶极其头疼的落地难题：<strong>远距离感知（Long-Range Detection）</strong>。它展示了如何用最廉价的标注，榨取单目视觉在远距离上的极限潜力。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/streampetr-qaf2d--enhancing-3d-object-detection-with-2d-detection-guided-query-anchors/ Sun, 15 Mar 2026 21:59:16 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/streampetr-qaf2d--enhancing-3d-object-detection-with-2d-detection-guided-query-anchors/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：CVPR<br> <strong>📅 发表年份</strong>：2024<br> <strong>💻 开源代码</strong>：<a href="https://github.com/nullmax-vision/QAF2D">nullmax-vision/QAF2D-CVPR 2024</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2403.06093">Enhancing 3D Object Detection with 2D Detection-Guided Query Anchors</a></p> </blockquote> <hr> <p>这篇发表于 CVPR 2024 的论文 <strong>《Enhancing 3D Object Detection with 2D Detection-Guided Query Anchors》(简称 QAF2D)</strong> 极具工程实用价值。它没有死磕 3D 空间中的特征提取瓶颈，而是打出了一套极其聪明的“降维组合拳”，巧妙地利用成熟的 2D 视觉技术来为 3D 检测器“引路”。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/obmo--one-bounding-box-multiple-objects-for-monocular-3d-object-detection/ Sun, 15 Mar 2026 21:59:12 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/obmo--one-bounding-box-multiple-objects-for-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：IEEE TIP<br> <strong>📅 发表年份</strong>：2023<br> <strong>💻 开源代码</strong>：<a href="https://github.com/mrsempress/OBMO_patchnet">mrsempress/OBMO_patchnet</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2212.10049">OBMO: One Bounding Box Multiple Objects for Monocular 3D Object Detection</a></p> </blockquote> <hr> <p>这篇发表于 IEEE TIP (2023) 的经典论文 <strong>《OBMO: One Bounding Box Multiple Objects for Monocular 3D Object Detection》</strong> 切入点非常犀利。它没有在复杂的网络主干上做文章，而是直击单目 3D 目标检测在“底层数学物理逻辑”上的痛点，提出了一种极其优雅的“即插即用（Plug-and-play）”训练策略。</p> https://zhaoylee.github.io/Blogs/posts/open-vocabulary-based/open-vocabulary-monocular-3d-object-detection/ Sun, 15 Mar 2026 21:14:37 +0800 https://zhaoylee.github.io/Blogs/posts/open-vocabulary-based/open-vocabulary-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：3DV<br> <strong>📅 发表年份</strong>：2026<br> <strong>💻 开源代码</strong>：<a href="https://github.com/UVA-Computer-Vision-Lab/ovmono3d">UVA-Computer-Vision-Lab/ovmono3d</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2411.16833">Open Vocabulary Monocular 3D Object Detection</a></p> </blockquote> <hr> <h3 id="一-背景研究目的与核心问题">一、 背景、研究目的与核心问题</h3> <ul> <li> <p><strong>研究背景：</strong> 传统的单目 3D 目标检测（M3OD）模型都属于“闭集（Closed-set）”学习。这意味着模型只能检测训练集中预先定义好的那几种类别（例如 KITTI 数据集里的车、人、自行车）。但在真实的自动驾驶或机器人场景中，会遇到无数的长尾目标（如遗落的轮胎、奇形怪状的施工路障、甚至是一只突然窜出的动物）。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/monoxiver--monocular-3d-object-detection-with-bounding-box-denoising-in-3d-by-perceivermonocular-3d-object-detection-with-bounding-box-denoising-in-3d-by-perceiver/ Sun, 15 Mar 2026 21:14:36 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/monoxiver--monocular-3d-object-detection-with-bounding-box-denoising-in-3d-by-perceivermonocular-3d-object-detection-with-bounding-box-denoising-in-3d-by-perceiver/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：ICCV<br> <strong>📅 发表年份</strong>：2023<br> <strong>💻 开源代码</strong>：<a href="https://github.com/Xianpeng919/monoxiver">Xianpeng919/monoxiver (ICCV'23)</a><br> <strong>📄 论文题目</strong>：<a href="https://openaccess.thecvf.com/content/ICCV2023/papers/Liu_Monocular_3D_Object_Detection_with_Bounding_Box_Denoising_in_3D_ICCV_2023_paper.pdf">Monocular 3D Object Detection with Bounding Box Denoising in 3D by Perceiver</a></p> </blockquote> <hr> <p>这篇发表于 ICCV 2023 的论文 <strong>《Monocular 3D Object Detection with Bounding Box Denoising in 3D by Perceiver》(简称 MonoXiver)</strong>，提供了一个非常经典且极具工程价值的“自上而下（Top-down）”纠错思路。它并没有试图发明一种全新的主干网络，而是设计了一个强大的“插件”，专门用来拯救那些定位不准的预测框。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/plot---pseudo-labeling-via-object-tracking-for-monocular-3d-object-detection/ Sun, 15 Mar 2026 20:52:51 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/plot---pseudo-labeling-via-object-tracking-for-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：ICLR<br> <strong>📅 发表年份</strong>：2026<br> <strong>💻 开源代码</strong>：<a href="%E5%A1%AB%E5%86%99%E4%BD%A0%E7%9A%84URL">无</a><br> <strong>📄 论文题目</strong>：<a href="https://openreview.net/pdf?id=3knS4J9isg">PLOT: Pseudo-Labeling via Object Tracking for Monocular 3D Object Detection</a></p> </blockquote> <hr> <h3 id="一-背景研究目的与核心问题">一、 背景、研究目的与核心问题</h3> <ul> <li> <p><strong>研究背景：</strong> 单目 3D 目标检测模型极度“吃数据”。然而，人工标注 3D 边界框极其昂贵且耗时，导致目前带 3D 标签的数据集规模很小，严重限制了模型的泛化能力。</p> https://zhaoylee.github.io/Blogs/posts/transformer-based/monodlgd--difficulty-aware-label-guided-denoising-for-monocular-3d-object-detection/ Sun, 15 Mar 2026 20:52:49 +0800 https://zhaoylee.github.io/Blogs/posts/transformer-based/monodlgd--difficulty-aware-label-guided-denoising-for-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：AAAI <br> <strong>📅 发表年份</strong>：2026<br> <strong>💻 开源代码</strong>：<a href="https://github.com/lsy010857/MonoDLGD">MonoDLGD</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2511.13195">Difficulty-Aware Label-Guided Denoising for Monocular 3D Object Detection</a></p> </blockquote> <hr> <h3 id="一-背景研究目的与核心问题">一、 背景、研究目的与核心问题</h3> <ul> <li> <p><strong>研究背景：</strong> 在基于 Transformer 的单目 3D 目标检测中，通过向真实标签注入噪声并让模型去重构（即查询去噪 Query Denoising），能有效加速模型收敛并提升几何感知能力。</p> https://zhaoylee.github.io/Blogs/posts/transformer-based/mono3dv--monocular-3d-object-detection-with-3d-aware-bipartite-matching-and-variational-query-denoising/ Sun, 15 Mar 2026 20:36:42 +0800 https://zhaoylee.github.io/Blogs/posts/transformer-based/mono3dv--monocular-3d-object-detection-with-3d-aware-bipartite-matching-and-variational-query-denoising/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：CVPR / ICCV / ECCV<br> <strong>📅 发表年份</strong>：2026<br> <strong>💻 开源代码</strong>： 无<br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2601.01036">Mono3DV: Monocular 3D Object Detection with 3D-Aware Bipartite Matching and Variational Query DeNoising</a></p> </blockquote> <hr> <h3 id="一-背景研究目的与核心问题">一、 背景、研究目的与核心问题</h3> <ul> <li> <p><strong>研究背景：</strong> 近年来，基于 Transformer（特别是 DETR 架构）的模型在 2D 目标检测中取得了巨大成功，并顺理成章地被引入到单目 3D 目标检测（M3OD）领域。这类模型依赖“查询（Query）”机制和“二分图匹配（Bipartite Matching）”来端到端地输出检测结果，无需繁琐的非极大值抑制（NMS）。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/dr-traversal-m3d--object-scene-camera-decomposition-and-recomposition-for-data-efficient-monocular-3d-object-detection/ Sun, 15 Mar 2026 19:59:15 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/dr-traversal-m3d--object-scene-camera-decomposition-and-recomposition-for-data-efficient-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：IJCV<br> <strong>📅 发表年份</strong>：2026<br> <strong>💻 开源代码</strong>：<a href="https://github.com/kwong292521/DR-Traversal-M3D">GitHub 链接</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2602.20627">Object-Scene-Camera Decomposition and Recomposition for Data-Efficient Monocular 3D Object Detection</a></p> </blockquote> <hr> <h3 id="一-背景研究目的与核心问题">一、 背景、研究目的与核心问题</h3> <ul> <li> <p><strong>研究背景：</strong> 单目 3D 目标检测本身是一个病态（ill-posed）问题，由于缺乏深度信息，模型极度依赖海量、多样化的训练数据来学习鲁棒的特征。</p> https://zhaoylee.github.io/Blogs/posts/centernet-based/spikesmoke--spiking-neural-networks-for-monocular-3d-object-detection-with-cross-scale-gated-coding/ Sun, 15 Mar 2026 19:59:10 +0800 https://zhaoylee.github.io/Blogs/posts/centernet-based/spikesmoke--spiking-neural-networks-for-monocular-3d-object-detection-with-cross-scale-gated-coding/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：CVPR / ICCV / ECCV<br> <strong>📅 发表年份</strong>：2026<br> <strong>💻 开源代码</strong>：<a href="%E5%A1%AB%E5%86%99%E4%BD%A0%E7%9A%84URL">无</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2506.07737">SpikeSMOKE: Spiking Neural Networks for Monocular 3D Object Detection with Cross-Scale Gated Coding</a></p> </blockquote> <hr> <h3 id="一-背景研究目的与核心问题">一、 背景、研究目的与核心问题</h3> <ul> <li> <p><strong>研究背景：</strong> 在自动驾驶等领域，3D 目标检测是核心技术。其中，“单目 3D 目标检测”由于仅依赖单张图像，硬件成本极低，备受青睐。然而，传统基于人工神经网络（ANNs）的模型计算量大、能耗极高，给边缘计算设备的电池续航和散热带来了巨大压力。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/span--spatial-projection-alignment-for-monocular-3d-object-detection/ Sun, 15 Mar 2026 19:38:47 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/span--spatial-projection-alignment-for-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：CVPR / ICCV / ECCV<br> <strong>📅 发表年份</strong>：2026<br> <strong>💻 开源代码</strong>：<a href="https://github.com/WYFDUT/SPAN">GitHub 链接</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/pdf/2511.06702">SPAN: Spatial-Projection Alignment for Monocular 3D Object Detection</a></p> </blockquote> <hr> <h2 id="1-文献背景与研究动机">1. 文献背景与研究动机</h2> <h2 id="背景与现状">背景与现状</h2> <p>单目3D目标检测（Monocular 3D Object Detection）是自动驾驶和机器人视觉中的核心任务，旨在仅通过单张RGB图像预测物体的3D边界框。</p> https://zhaoylee.github.io/Blogs/posts/plug_and_play/digging-into-output-representation-for-monocular-3d-object-detection/ Fri, 13 Mar 2026 15:27:43 +0800 https://zhaoylee.github.io/Blogs/posts/plug_and_play/digging-into-output-representation-for-monocular-3d-object-detection/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：ICLR<br> <strong>📅 发表年份</strong>：2022<br> <strong>💻 开源代码</strong>：<a href="https://github.com/Owen-Liuyuxuan/visualDet3D/releases/tag/1.1.1">GitHub 链接</a><br> <strong>📄 论文题目</strong>：<a href="https://openreview.net/pdf?id=mPlm356yMIP">DIGGING INTO OUTPUT REPRESENTATION FOR MONOCULAR 3D OBJECT DETECTION</a></p> </blockquote> <hr> <h2 id="0-一句话总结-tldr">0. 一句话总结 (TL;DR)</h2> <p><em>(这篇论文用什么方法，解决了什么问题，达到了什么效果)</em></p> https://zhaoylee.github.io/Blogs/posts/centernet-based/monoct-overcoming-monocular-3d-detection-domain-shift-with-consistent-teacher-models/ Thu, 12 Mar 2026 10:25:13 +0800 https://zhaoylee.github.io/Blogs/posts/centernet-based/monoct-overcoming-monocular-3d-detection-domain-shift-with-consistent-teacher-models/ <hr> <blockquote> <p><strong>🏛️ 会议/期刊</strong>：ICRA<br> <strong>📅 发表年份</strong>：2025<br> <strong>💻 开源代码</strong>：<a href="%E5%A1%AB%E5%86%99%E4%BD%A0%E7%9A%84URL">GitHub 链接</a><br> <strong>📄 论文题目</strong>：<a href="https://arxiv.org/abs/2503.13743">MonoCT: Overcoming Monocular 3D Detection Domain Shift with Consistent Teacher Models</a></p> </blockquote> <hr> <h2 id="0-一句话总结-tldr">0. 一句话总结 (TL;DR)</h2> <p><em>(这篇论文用什么方法，解决了什么问题，达到了什么效果)</em><br> MonoCT 提出了一种基于<strong>一致性教师模型（Consistent Teacher）的</strong>半监督自适应框架，通过在目标域（Target Domain）引入伪标签一致性约束，有效解决了单目 3D 检测在不同数据集间迁移时的深度估计偏差问题。</p> https://zhaoylee.github.io/Blogs/about/ Mon, 01 Jan 0001 00:00:00 +0000 https://zhaoylee.github.io/Blogs/about/ about