https://simplestory.github.io/tags/deep-learning/Recent content in Deep Learning on SimplestoryHugozh-cnSimplestorySat, 17 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240817_d9d0d187039bec56d5fc1024c51a882c/Sat, 17 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240817_d9d0d187039bec56d5fc1024c51a882c/Denoising Diffusion Implicit Modelshttps://simplestory.github.io/posts/20240817_60a1868ea6b6f7b05126264f9d212a9c/Sat, 17 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240817_60a1868ea6b6f7b05126264f9d212a9c/Denoising Diffusion Probabilistic Modelshttps://simplestory.github.io/posts/20240817_86b1d339bba4e0c154df43f21e99fe50/Sat, 17 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240817_86b1d339bba4e0c154df43f21e99fe50/Latent Diffusion Modelshttps://simplestory.github.io/posts/20240811_787700a547309b9912534d7c3540b79b/Sun, 11 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240811_787700a547309b9912534d7c3540b79b/Generative Adversarial Networkhttps://simplestory.github.io/posts/20240811_a9423d60164f001a6715546c56c77154/Sun, 11 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240811_a9423d60164f001a6715546c56c77154/Varitional Autoencoderhttps://simplestory.github.io/posts/20240811_f09978b03d9ad93516b6003551892ddc/Sun, 11 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240811_f09978b03d9ad93516b6003551892ddc/Vector Quantized Generative Adversarial Networkhttps://simplestory.github.io/posts/20240811_520db3fb8aba368d31c1f3e5e9fbd8ba/Sun, 11 Aug 2024 00:00:00 +0000https://simplestory.github.io/posts/20240811_520db3fb8aba368d31c1f3e5e9fbd8ba/Vector Quantised Variational AutoEncoderhttps://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>https://simplestory.github.io/posts/20221006_1087bec7ef12911010976d203f36fd7f/Thu, 06 Oct 2022 00:00:00 +0000https://simplestory.github.io/posts/20221006_1087bec7ef12911010976d203f36fd7f/目标检测中的标签分配问题https://simplestory.github.io/posts/20210807_1c4ff5cd7d69cb58295eab75872ba2fd/Sat, 07 Aug 2021 00:00:00 +0000https://simplestory.github.io/posts/20210807_1c4ff5cd7d69cb58295eab75872ba2fd/<blockquote> <p>PointNet、PointNet++和Frustum PointNet。常用的3D点云目标检测。</p> </blockquote> <h2 id="pointnet">PointNet</h2> <p>PointNet将深度学习直接应用于点云数据。点云数据是欧式空间下的点的一个子集，有无序性、点与点的空间关系以及不变性这三个性质。</p>https://simplestory.github.io/posts/20210806_8e5f2ccbcaba22801139e42b929ed820/Fri, 06 Aug 2021 00:00:00 +0000https://simplestory.github.io/posts/20210806_8e5f2ccbcaba22801139e42b929ed820/MV3Dhttps://simplestory.github.io/posts/20210805_59094eb979620e0e11f92ffaa71cf228/Thu, 05 Aug 2021 00:00:00 +0000https://simplestory.github.io/posts/20210805_59094eb979620e0e11f92ffaa71cf228/<blockquote> <p>当前工业界常用的多目标跟踪框架。</p> </blockquote> <h2 id="sort">SORT</h2> <p>该算法基于目标检测网络（Faster Rcnn），并利用卡尔曼滤波和匈牙利算法（或KM算法），极大地提高了多目标跟踪的速度。大致流程如下：</p>https://simplestory.github.io/posts/20210804_697dd2d087a31ba4aab4ff575bde80be/Wed, 04 Aug 2021 00:00:00 +0000https://simplestory.github.io/posts/20210804_697dd2d087a31ba4aab4ff575bde80be/动态网络头https://simplestory.github.io/posts/20210803_2d8c6e5a2f995c3808422afc8967c85e/Tue, 03 Aug 2021 00:00:00 +0000https://simplestory.github.io/posts/20210803_2d8c6e5a2f995c3808422afc8967c85e/<blockquote> <p>让它动起来。</p> </blockquote> <h2 id="dynamic-convolution">Dynamic Convolution</h2> <p>一般的静态卷积用同一个卷积核对所有输入图像做相同的操作，而动态卷积会对不同的图像做调整，用更适合的卷积参数进行处理，这样可以在不增加网络深度或宽度的情况下增加模型的表达能力。</p>https://simplestory.github.io/posts/20210430_b30ab79aff271e301dde9ca18c2e8085/Fri, 30 Apr 2021 00:00:00 +0000https://simplestory.github.io/posts/20210430_b30ab79aff271e301dde9ca18c2e8085/<blockquote> <p>图像分割是对图像进行像素级的分类，主要有语义分割(Semantic Segmentation)和实例分割(Instance Segmentation)。语义分割只要求做到像素层面的分割，而实例分割要求在语义分割的基础上实现对不同个体的分割。这篇文章主要简单介绍几种基础的语义分割网络。</p> </blockquote>https://simplestory.github.io/posts/20201120_0595a0de31914bca4c526fb50d30b2dd/Fri, 20 Nov 2020 00:00:00 +0000https://simplestory.github.io/posts/20201120_0595a0de31914bca4c526fb50d30b2dd/Object Detection in 20 Yearshttps://simplestory.github.io/posts/20201025_93f382c06f3476ad1f59351e0d99f842/Sun, 25 Oct 2020 00:00:00 +0000https://simplestory.github.io/posts/20201025_93f382c06f3476ad1f59351e0d99f842/Objects as Pointshttps://simplestory.github.io/posts/20201018_25c3043509dadf35f8ecc40adf5a369c/Sun, 18 Oct 2020 00:00:00 +0000https://simplestory.github.io/posts/20201018_25c3043509dadf35f8ecc40adf5a369c/some loss functionhttps://simplestory.github.io/posts/20201003_fc3691951ecd38cc130a442bcc0566ed/Sat, 03 Oct 2020 00:00:00 +0000https://simplestory.github.io/posts/20201003_fc3691951ecd38cc130a442bcc0566ed/EfficientX basic familyhttps://simplestory.github.io/posts/20200926_ffb198e96e96a613e471e2f2eb1ae86c/Sat, 26 Sep 2020 00:00:00 +0000https://simplestory.github.io/posts/20200926_ffb198e96e96a613e471e2f2eb1ae86c/<blockquote> <p>IoU是目标检测里的一个重要指标，它是通过计算预测框与真实框的交集和并集的比值来衡量预测框的优劣。但通常的预测框调整函数一般采用的是L2范数，即以MSE的方式来计算损失，而目前来看，一些研究表明这并非最优的优化函数，所以就有了IOU损失函数。</p> </blockquote>https://simplestory.github.io/posts/20200426_02a6a4457d33491491495f7fe2b383d1/Sun, 26 Apr 2020 00:00:00 +0000https://simplestory.github.io/posts/20200426_02a6a4457d33491491495f7fe2b383d1/Inverted Residuals and Linear Bottleneckshttps://simplestory.github.io/posts/20200210_ba14ebd4efe67efd08cdc6410e758633/Mon, 10 Feb 2020 00:00:00 +0000https://simplestory.github.io/posts/20200210_ba14ebd4efe67efd08cdc6410e758633/without bells and whistleshttps://simplestory.github.io/posts/20200115_1789a504337212c61f0b46c458695b7c/Fri, 17 Jan 2020 00:00:00 +0000https://simplestory.github.io/posts/20200115_1789a504337212c61f0b46c458695b7c/An Incremental Improvementhttps://simplestory.github.io/posts/20200110_aa4020065e766e48007d2b3ae229757d/Fri, 10 Jan 2020 00:00:00 +0000https://simplestory.github.io/posts/20200110_aa4020065e766e48007d2b3ae229757d/Better, Faster, Strongerhttps://simplestory.github.io/posts/20200101_83ba412c3f5a7a54968bf73cc917cd13/Wed, 01 Jan 2020 00:00:00 +0000https://simplestory.github.io/posts/20200101_83ba412c3f5a7a54968bf73cc917cd13/You Only Look Oncehttps://simplestory.github.io/posts/20191003_34df20bab5e85dc75bfc94ef569cced9/Thu, 03 Oct 2019 00:00:00 +0000https://simplestory.github.io/posts/20191003_34df20bab5e85dc75bfc94ef569cced9/Single Shot MultiBox Detectorhttps://simplestory.github.io/posts/20190922_5e8040095d43e97ccedd09771f9eab38/Sun, 22 Sep 2019 00:00:00 +0000https://simplestory.github.io/posts/20190922_5e8040095d43e97ccedd09771f9eab38/目标检测识别算法https://simplestory.github.io/posts/20190921_817fe11366b1149fdc44a5a4c92685c4/Sat, 21 Sep 2019 00:00:00 +0000https://simplestory.github.io/posts/20190921_817fe11366b1149fdc44a5a4c92685c4/经典目标检测识别算法https://simplestory.github.io/posts/20190919_13243638926ee79bc7810c80a460408c/Thu, 19 Sep 2019 00:00:00 +0000https://simplestory.github.io/posts/20190919_13243638926ee79bc7810c80a460408c/经典目标检测识别算法https://simplestory.github.io/posts/20190910_cefb8aa8290a0b4cfddf3f88bf5a2a3c/Tue, 10 Sep 2019 00:00:00 +0000https://simplestory.github.io/posts/20190910_cefb8aa8290a0b4cfddf3f88bf5a2a3c/<blockquote> <p>首先祝所有教师节日快乐，感谢你们赠人玫瑰手留余香的奉献精神。然后时隔将近一个月，我又滚回来更新博客了，之前一直在忙着入职这件事，希望能尽快融入公司开发业务中而”忙碌“了一个月。接下来继续经典图像识别网络模型的介绍。上文我们已经介绍了AlexNet和VGG网络，这两个网络都将当时的神经网络的深度发挥到了极致，后续的神经网络模型性能基本上好不了多少。这是由于随着网络的加深，在误差后向传播时，靠前的网络层权重基本得不到更新，一直维持在某一个值附近。在2015年，何凯明大神提出了一种快捷连接方式将误差传播到了前面的网络层，很大程度上提高了网络的性能。下面简单解析一下ResNet网络。</p> </blockquote>https://simplestory.github.io/posts/20190811_58bd041ca5bf7b009cdc8d07392dc061/Sun, 11 Aug 2019 00:00:00 +0000https://simplestory.github.io/posts/20190811_58bd041ca5bf7b009cdc8d07392dc061/<blockquote> <p>神经网络已经有几个归一化算法了，它们长得都很相似，基本上都有统一的形式。本文参考了知乎专栏SIGAI关于归一化算法的文章，具体链接在文章底部。</p> </blockquote>https://simplestory.github.io/posts/20190730_1036fd1af9bc59d1a811b4e2ab23d60c/Tue, 30 Jul 2019 00:00:00 +0000https://simplestory.github.io/posts/20190730_1036fd1af9bc59d1a811b4e2ab23d60c/经典的卷积神经网络https://simplestory.github.io/posts/20190722_cdebdc3cc4ecd47a04cff48d03a688d3/Mon, 22 Jul 2019 00:00:00 +0000https://simplestory.github.io/posts/20190722_cdebdc3cc4ecd47a04cff48d03a688d3/经典的卷积神经网络https://simplestory.github.io/posts/20181103_7b202c9a380844d1b8ae2c4f5ee3461f/Sat, 03 Nov 2018 00:00:00 +0000https://simplestory.github.io/posts/20181103_7b202c9a380844d1b8ae2c4f5ee3461f/<blockquote> <p>最优化是一个十分困难的问题，通常要精心设计一个目标函数及相关约束，同时还要确保所要优化的问题为凸函数，但现实中的问题大多为非凸函数，针对非凸优化问题，这里有一些基本的优化方法：SGD、Momentum、AdaGrad、RMSProp及Adam等</p> </blockquote>