https://simplestory.github.io/categories/2023/Recent content in 2023 on SimplestoryHugozh-cnSimplestorySat, 24 Jun 2023 00:00:00 +0000https://simplestory.github.io/posts/20230624_593a3023b826e242be5651751b69d09e/Sat, 24 Jun 2023 00:00:00 +0000https://simplestory.github.io/posts/20230624_593a3023b826e242be5651751b69d09e/Lift, Splat, Shoothttps://simplestory.github.io/posts/20230623_f25271ca7acc8b883a6ebeba1a0b263d/Fri, 23 Jun 2023 00:00:00 +0000https://simplestory.github.io/posts/20230623_f25271ca7acc8b883a6ebeba1a0b263d/<p>单目3D检测在自动驾驶领域是一个重要且充满挑战的任务，目前的难点是如何准确对目标深度进行预测。MonoRCNN利用几何关系将深度估计问题解耦为目标实际高度和在图像上的投影高度问题。</p>https://simplestory.github.io/posts/20230623_e5106aece831780fe8decad391998e6a/Fri, 23 Jun 2023 00:00:00 +0000https://simplestory.github.io/posts/20230623_e5106aece831780fe8decad391998e6a/<p>在2D实时目标检测框架yolo上引入一个轻量级立体匹配模块来实现高效的3D物体检测。大致结构如下图：</p> <p><img loading="lazy" src="https://simplestory-blog-img.oss-cn-guangzhou.aliyuncs.com/in_posts/20230623/yolostereo3d_architecture.png" title="yolostereo3d_architecture"></p> <h2 id="anchors">Anchors</h2> <p>对于3D目标检测，原本yolo针对2D物体的anchors需要重新设计，作者采用了12个参数来表示一个anchor：</p>https://simplestory.github.io/posts/20230305_f73b955eec390396d75e03ababcc4909/Sun, 05 Mar 2023 00:00:00 +0000https://simplestory.github.io/posts/20230305_f73b955eec390396d75e03ababcc4909/<p>点云有着分布上稀疏、结构上不规则、几何上无序的特性，所以直接在点云上使用卷积操作可能会导致生成的特征极度地偏离它原本的表达。</p> <h1 id="voxelnet">VoxelNet</h1> <p>该模型是对点云进行体素化，然后利用3D卷积来获取特征。具体是：</p>https://simplestory.github.io/posts/20230127_2a9bce23520b48c68e00a87aa00aa611/Fri, 27 Jan 2023 00:00:00 +0000https://simplestory.github.io/posts/20230127_2a9bce23520b48c68e00a87aa00aa611/<blockquote> <p>该网络通过双目视觉来实现3D目标检测，属于3D detection中image based的result-lifting方法，这类方法就是将3D检测分解为2D object detection和depth estimation这两部分。</p> </blockquote>