<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[redfish95的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/redfish95</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; redfish95]]></copyright><item><title><![CDATA[SparseDrive原理及代码解析]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/157548578</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/157548578</guid><author>redfish95</author><pubDate>Mon, 02 Feb 2026 22:01:15 +0800</pubDate><description><![CDATA[SparseDrive是清华大学与地平线联合提出的端到端自动驾驶新范式，在感知、预测和规划任务上均达到SOTA性能。该方法采用对称稀疏感知模块，通过解耦的实例特征和锚框同时处理3D目标检测、跟踪和稀疏建图任务。运动规划模块则基于感知结果进行多模态轨迹预测和自车轨迹规划，并引入碰撞过滤策略确保安全性。实验表明，SparseDrive在NuScenes数据集上实现了全任务最优性能。该方法通过统一的稀疏表示实现了高效的多任务协同学习，为自动驾驶系统提供了一种简洁而强大的解决方案。]]></description><category></category></item><item><title><![CDATA[Sparse4DOE代码浅析]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/156130288</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/156130288</guid><author>redfish95</author><pubDate>Tue, 23 Dec 2025 19:53:18 +0800</pubDate><description><![CDATA[本文解析了地平线Sparse4D V3版本的BEV稀疏感知算法。该算法通过多路相机图像输入，输出BEV空间下的3D检测框，具有时序融合能力。模型采用类Transformer结构，包含6层重复模块，每层包含时序交互、实例交互、可变形注意力等子模块。相比原版，OE版本在模型结构上进行了简化以提高推理性能。算法创新点包括：1) 去噪训练策略，通过添加噪声anchor加速收敛；2) 引入centerness和yawness辅助训练；3) 采用深度监督学习稠密深度图。模型通过Memory Bank机制实现跨帧信息融合]]></description><category></category></item><item><title><![CDATA[BEVFormer算法及代码解析]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/153622187</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/153622187</guid><author>redfish95</author><pubDate>Mon, 20 Oct 2025 16:16:18 +0800</pubDate><description><![CDATA[BEVFormer是一种基于Transformer的BEV感知算法，通过多视角相机输入构建BEV特征图。其核心模块包括时空自注意力(TSA)和空间交叉注意力(SCA)，分别用于时序特征融合和多视角图像特征投影。算法采用可学习的BEV查询作为初始特征，通过6层编码器迭代优化，最终输出可用于3D检测或分割的BEV特征。关键创新点包括：1）时序对齐模块处理自车运动；2）可变形注意力机制实现高效特征采样；3）多尺度特征融合策略。代码实现中包含了网格掩码数据增强、可学习位置编码等细节设计，为后续BEV研究提供了重要参]]></description><category></category></item><item><title><![CDATA[GUPNet原理解析]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/153526966</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/153526966</guid><author>redfish95</author><pubDate>Sat, 18 Oct 2025 21:59:15 +0800</pubDate><description><![CDATA[GUPNet是一种基于单目图像的3D目标检测网络，采用两阶段框架：首先通过2D检测头获取2D框，再通过ROI特征和几何投影预测3D信息。核心创新包括：1）基于小孔成像原理的深度预测模块，引入几何不确定度概念；2）采用Hierarchical Task Learning（HTL）策略，通过评估前置任务状态动态调整当前任务权重。网络输出包含目标类别、2D/3D框等属性，其中3D预测通过热图、偏移量和尺寸回归实现，深度计算结合投影几何和修正值。损失函数设计涵盖热图（Focal Loss）、偏移量（L1 Loss）]]></description><category></category></item><item><title><![CDATA[LSS原理及代码解析]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/143276098</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/143276098</guid><author>redfish95</author><pubDate>Sun, 22 Dec 2024 09:48:40 +0800</pubDate><description><![CDATA[BEV感知的核心思想是将多路传感器的感知数据转换到统一的BEV空间去提取特征，实现目标检测、语义分割、地图构建等任务，对于相机感知模块，转向BEV空间可带来很大的收益，主要体现在三个方面：目前BEV感知主要出现了3类视图转换模块的方案：IPM（Inverse Perspective Mapping）、Lift-splat和Transformer。本文主要对基于Lift-splat（后文简称LS或LSS）的方案进行解析。LSS是英伟达（NVIDIA）在ECCV2020上发表的文章（Lift，Splat，Sho]]></description><category></category></item><item><title><![CDATA[Python 并发编程]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/129250367</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/129250367</guid><author>redfish95</author><pubDate>Wed, 08 Mar 2023 18:39:07 +0800</pubDate><description><![CDATA[python并发编程介绍]]></description><category></category></item><item><title><![CDATA[释放cuda显存及进程]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/129162867</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/129162867</guid><author>redfish95</author><pubDate>Wed, 22 Feb 2023 15:02:20 +0800</pubDate><description><![CDATA[如何释放cuda显存及进程]]></description><category></category></item><item><title><![CDATA[Docker教程及完整讲义（入门级）]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/129007504</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/129007504</guid><author>redfish95</author><pubDate>Mon, 13 Feb 2023 21:56:31 +0800</pubDate><description><![CDATA[Docker教程及完整讲义]]></description><category></category></item><item><title><![CDATA[nvidia-smi 命令详解]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/129005910</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/129005910</guid><author>redfish95</author><pubDate>Mon, 13 Feb 2023 11:57:22 +0800</pubDate><description><![CDATA[nvidia-smi命令详解]]></description><category></category></item><item><title><![CDATA[深度学习中的随机数种子]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/129000000</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/129000000</guid><author>redfish95</author><pubDate>Sun, 12 Feb 2023 22:27:27 +0800</pubDate><description><![CDATA[本文参考网络上对于深度学习中seed函数的讲解，记录一下我对于这个函数的理解。]]></description><category></category></item><item><title><![CDATA[Tensorboard的使用 ---- SummaryWriter类（pytorch版）]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/128990282</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/128990282</guid><author>redfish95</author><pubDate>Sun, 12 Feb 2023 11:06:20 +0800</pubDate><description><![CDATA[在进行模型训练时，对训练进行可视化可以帮助我们更直观查看模型训练情况，从而更容易发现问题。本文用到的方法为tensorboard可视化方法。]]></description><category></category></item><item><title><![CDATA[Git简易指令]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/128989493</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/128989493</guid><author>redfish95</author><pubDate>Sat, 11 Feb 2023 22:06:34 +0800</pubDate><description><![CDATA[Git简易指令]]></description><category></category></item><item><title><![CDATA[ubuntu22.04安装显卡驱动、CUDA及cudnn]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/128909073</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/128909073</guid><author>redfish95</author><pubDate>Tue, 07 Feb 2023 22:40:01 +0800</pubDate><description><![CDATA[教你如何在ubuntu22.04上安装显卡驱动、CUDA以及cudnn]]></description><category></category></item><item><title><![CDATA[目标检测评测简介]]></title><link>https://blog.youkuaiyun.com/redfish95/article/details/128899889</link><guid>https://blog.youkuaiyun.com/redfish95/article/details/128899889</guid><author>redfish95</author><pubDate>Mon, 06 Feb 2023 13:29:32 +0800</pubDate><description><![CDATA[目标检测评测目标检测评测简介]]></description><category></category></item></channel></rss>