<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[婷婷是信仰]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/hurricane91</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; hurricane91]]></copyright><item><title><![CDATA[Structural 3D Analysis]]></title><link>https://blog.youkuaiyun.com/hurricane91/article/details/112798805</link><guid>https://blog.youkuaiyun.com/hurricane91/article/details/112798805</guid><author>hurricane91</author><pubDate>Tue, 19 Jan 2021 08:48:29 +0800</pubDate><description><![CDATA[Unsupervised Learning of Intrinsic Structural Representation Points （CVPR2020）

原文arxiv链接

这篇工作能够无监督的学习出人造物体三维点云的结构。对于输入的三维点云，使用PointNet++对其进行编码和降采样，得到一系列特征点。接着继续用PointNet++预测这些点属于各个预定义的结构点的概率。依照概率计算各个结构点的坐标后，将结构点和原始点云进行双向Chamfer Loss计算，进行网络优化训练。其主要贡献是能够无监]]></description><category></category></item><item><title><![CDATA[View Synthesis]]></title><link>https://blog.youkuaiyun.com/hurricane91/article/details/112546242</link><guid>https://blog.youkuaiyun.com/hurricane91/article/details/112546242</guid><author>hurricane91</author><pubDate>Tue, 12 Jan 2021 20:39:13 +0800</pubDate><description><![CDATA[一、SynSin: End-to-End Synthesis from a Single Image

该论文给出了一个视角图片生成的方法，其pipeline如下图：



首先将图片输入特征和深度网络得到特征map和深度图，接着通过相机参数变换为带特征的点云，接着根据相对变换矩阵T，将带特征的点云渲染到二维像素位置上，接着通过一个GAN生成最终的新视角图片。

该论文的渲染方式值得关注，其并不是简单的使用zbuffer和固定投影像素位置渲染的，在z深度和平面广度上都是软映射的。



平面广度上是线性递减]]></description><category></category></item><item><title><![CDATA[Nerf: 可用于多视角生成的场景表示]]></title><link>https://blog.youkuaiyun.com/hurricane91/article/details/112433487</link><guid>https://blog.youkuaiyun.com/hurricane91/article/details/112433487</guid><author>hurricane91</author><pubDate>Sun, 10 Jan 2021 18:06:18 +0800</pubDate><description><![CDATA[Nerf: 可用于多视角生成的场景表示

一、初代Nerf

《NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis》一文提出了一种基于光照模型隐式场表示的场景表示方法Nerf，用于生成一个场景的不同视角。

Nerf作为一种场表示，其输入是三维空间某个顶点的三维坐标以及视角方向，是一个5D的向量。其输出为该位置发射的光线颜色以及该点的体密度。该隐式场的函数表示为。考虑到空间中点的体密度应该独立于视角方向，首先将输入的三]]></description><category></category></item><item><title><![CDATA[Samda添加用户]]></title><link>https://blog.youkuaiyun.com/hurricane91/article/details/101158537</link><guid>https://blog.youkuaiyun.com/hurricane91/article/details/101158537</guid><author>hurricane91</author><pubDate>Sun, 22 Sep 2019 14:55:14 +0800</pubDate><description><![CDATA[Samba为ubuntu系统的服务器和本地Windows电脑的文件共享提供了可能，以下是配置方案：

  1.首先找到samba的配置文件所在的位置，输入命令 &gt;&gt; find /etc -name smb.conf

  2.找到配置文件路径之后，用vim打开它 &gt;&gt; sudo vi /etc/samba/smb.conf。注意使用sudo以拥有读写权限。

 ...]]></description><category></category></item><item><title><![CDATA[AutoEncoder（自编码器）之风雨调参路]]></title><link>https://blog.youkuaiyun.com/hurricane91/article/details/90421313</link><guid>https://blog.youkuaiyun.com/hurricane91/article/details/90421313</guid><author>hurricane91</author><pubDate>Wed, 22 May 2019 01:17:14 +0800</pubDate><description><![CDATA[早在一年多前就接触到了自编码器这个简单巧妙的深度学习模型，还曾编写过简单的基于MNIST的自编码器。但是因为觉得其思想简单，就扔入武器库中任其生锈。

  一年后的今天，居然因为当初的不求甚解赔付了巨额罚单。。。

  任务如下：手语识别项目的大致思路是将视频编码成为二维向量，之后对其 Covariance Matrix 进行 SVD，保留前k大的奇异值，将该数据映射到其空间对应的...]]></description><category></category></item><item><title><![CDATA[Fraction Pooling]]></title><link>https://blog.youkuaiyun.com/hurricane91/article/details/86475731</link><guid>https://blog.youkuaiyun.com/hurricane91/article/details/86475731</guid><author>hurricane91</author><pubDate>Mon, 14 Jan 2019 13:18:05 +0800</pubDate><description><![CDATA[Compared to the max pooling, Fraction Pooling can have a subsampling ratio belonging to real number set.

Take the ratio from 1 to 2 as an example, the pool size could be 1 or 2.

Then you need to mak...]]></description><category></category></item><item><title><![CDATA[Notes of Python]]></title><link>https://blog.youkuaiyun.com/hurricane91/article/details/79534890</link><guid>https://blog.youkuaiyun.com/hurricane91/article/details/79534890</guid><author>hurricane91</author><pubDate>Mon, 12 Mar 2018 23:37:33 +0800</pubDate><description><![CDATA[I've learned how to check and change the path of current work.  I'd like to share it with my friends.1.How to check the working path? -Use import os, then type print(os.getcwd())2.To change the path, ...]]></description><category></category></item></channel></rss>