https://worldhellooo.github.io/

1<!DOCTYPE html><html lang="zh-CN"><head><meta http-equiv="content-type" content="text/html; charset=utf-8"><meta content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=0" name="viewport"><meta content="yes" name="apple-mobile-web-app-capable"><meta content="black-translucent" name="apple-mobile-web-app-status-bar-style"><meta content="telephone=no" name="format-detection"><meta name="description"><title>WorldHellooo's Blog</title><link rel="stylesheet" type="text/css" href="//cdn.bootcss.com/normalize/5.0.0/normalize.min.css"><link rel="stylesheet" type="text/css" href="//cdn.bootcss.com/pure/0.6.2/pure-min.css"><link rel="stylesheet" type="text/css" href="//cdn.bootcss.com/pure/0.6.2/grids-responsive-min.css"><link rel="stylesheet" type="text/css" href="/css/style.css?v=0.0.0"><link rel="stylesheet" href="//cdn.bootcss.com/font-awesome/4.7.0/css/font-awesome.min.css">

1<script type="text/javascript" src="//cdn.bootcss.com/jquery/3.1.1/jquery.min.js"></script>

1<link rel="Shortcut Icon" type="image/x-icon" href="/favicon.ico"><link rel="apple-touch-icon" href="/apple-touch-icon.png"><link rel="apple-touch-icon-precomposed" href="/apple-touch-icon.png"><link rel="alternate" type="application/atom+xml" href="/atom.xml"></head><body><div class="body_container"><div id="header"><div class="site-name"><h1 class="hidden">WorldHellooo's Blog</h1><a id="logo" href="/.">WorldHellooo's Blog</a><p class="description"></p></div><div id="nav-menu"><a href="/." class="current"><i class="fa fa-home"> 首页</i></a><a href="/archives/"><i class="fa fa-archive"> 归档</i></a><a href="/about/"><i class="fa fa-user"> 关于</i></a></div></div><div id="layout" class="pure-g"><div class="pure-u-1 pure-u-md-3-4"><div class="content_container"><div class="post"><h2 class="post-title"><a href="/2018/12/12/关于概率和熵的一些概念/">关于概率和熵的一些概念</a></h2><div class="post-meta">2018-12-12</div><div class="post-content"><p><strong>本文基于维基百科相关词条整理</strong></p>
2<h2 id="熵"><a href="#熵" class="headerlink" title="熵"></a>熵</h2><h3 id="定义"><a href="#定义" class="headerlink" title="定义"></a>定义</h3><p>信息论中，熵（entropy）代表了消息中所包含的信息的平均量，是不确定性的度量。<br>熵$H(X) = E[I(X)] = E[-ln(P(X))]$<br>其中P为随机变量X的概率质量函数，E为期望，I(X)代表X的信息量。<br>当X为离散随机变量时，可以写成：$H(x) = \sum_{i}P(x_i)\log_2{P(x_i)}$</p></div><p class="readmore"><a href="/2018/12/12/关于概率和熵的一些概念/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/05/18/deepmind论文解读/">DeepMind论文：连接多巴胺与元强化学习的新方法</a></h2><div class="post-meta">2018-05-18</div><div class="post-content"><h2 id="ABSTRACT"><a href="#ABSTRACT" class="headerlink" title="ABSTRACT"></a>ABSTRACT</h2><p>过去20年间，神经科学中对于Reward-based Learning的研究已经收敛到了一类规范的模式上：神经递质多巴胺通过调整神经元之间的突触连接强度，以在情景、动作和奖励之间建立关联。但是，近期越来越多的发现开始挑战这一标准模型。因此，本文提出一种新的基于奖励的学习机制：多巴胺系统训练了另一个大脑区域————前额叶，来将其作为独立的学习系统。这个全新的视角既能够继承标准模型的那些依据，也能够很好地处理宽泛的经验观察，为未来的研究提供全新的基础。</p></div><p class="readmore"><a href="/2018/05/18/deepmind论文解读/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/26/pytorch doc阅读笔记/">pytorch doc阅读笔记</a></h2><div class="post-meta">2018-04-26</div><div class="post-content"><p>[TOC]</p>
3<h2 id="torch特性"><a href="#torch特性" class="headerlink" title="torch特性"></a>torch特性</h2><p>待补充</p>
4<h2 id="torch-Tensor"><a href="#torch-Tensor" class="headerlink" title="torch.Tensor"></a>torch.Tensor</h2><h3 id="expand方法"><a href="#expand方法" class="headerlink" title="expand方法"></a>expand方法</h3><p>expand方法并不会开辟新的内存，只是在已有tensor上创建了新的view，修改原tensor，会影响到expand返回的变量。<br>举例：<br><figure class="highlight python"><table><tr><td class="gutter"><pre><div class="line">1</div><div class="line">2</div><div class="line">3</div><div class="line">4</div><div class="line">5</div><div class="line">6</div><div class="line">7</div><div class="line">8</div><div class="line">9</div></pre></td><td class="code"><pre><div class="line">a.unsqueeze_(<span class="number">1</span>)</div><div class="line">b = a.expand(<span class="number">-1</span>,<span class="number">4</span>)</div><div class="line">b = <span class="number">1</span> <span class="number">1</span> <span class="number">1</span> <span class="number">1</span></div><div class="line">    <span class="number">2</span> <span class="number">2</span> <span class="number">2</span> <span class="number">2</span></div><div class="line">    <span class="number">3</span> <span class="number">3</span> <span class="number">3</span> <span class="number">3</span></div><div class="line">a[<span class="number">1</span>] = <span class="number">4</span></div><div class="line">b = <span class="number">1</span> <span class="number">1</span> <span class="number">1</span> <span class="number">1</span></div><div class="line">    <span class="number">4</span> <span class="number">4</span> <span class="number">4</span> <span class="number">4</span></div><div class="line">    <span class="number">3</span> <span class="number">3</span> <span class="number">3</span> <span class="number">3</span></div></pre></td></tr></table></figure></p></div><p class="readmore"><a href="/2018/04/26/pytorch doc阅读笔记/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/18/大数据竞赛经验谈/">大数据竞赛经验谈</a></h2><div class="post-meta">2018-04-18</div><div class="post-content"><h2 id="随笔"><a href="#随笔" class="headerlink" title="随笔"></a>随笔</h2><p>最近在和同学一起做阿里天池的服装属性识别的竞赛，训了几个不å

4Œæž¶æž„的分类网络，然后手动融合了一下结果。但是感觉黔驴技穷，不知道该怎么提升了。所以今天跟实验室的Kaggle大神师兄讨教了一下，学到了很多做比赛的经验，在此记录。</p></div><p class="readmore"><a href="/2018/04/18/大数据竞赛经验谈/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/12/基于PyQt的多源图像文字描述生成软件界面/">基于PyQt的多源图像文字描述生成软件界面</a></h2><div class="post-meta">2018-04-12</div><div class="post-content"><h2 id="任务简介"><a href="#任务简介" class="headerlink" title="任务简介"></a>任务简介</h2><p>因为项目要求，先需要设计一个软件界面。业务逻辑主要是基于深度学习实现图像的自动文字描述。界面主要要包括基本的菜单栏、待处理图像的展示、文字生成结果展示、图像批次切换按键等。</p></div><p class="readmore"><a href="/2018/04/12/基于PyQt的多源图像文字描述生成软件界面/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/08/翻译 Region of interest pooling explained/">翻译 Region of interest pooling explained</a></h2><div class="post-meta">2018-04-08</div><div class="post-content"><p>原文链接：<br><a href="https://deepsense.io/region-of-interest-pooling-explained/" target="_blank" rel="external">https://deepsense.io/region-of-interest-pooling-explained/</a></p></div><p class="readmore"><a href="/2018/04/08/翻译 Region of interest pooling explained/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/08/深度网络训练中的一些参数/">深度网络训练中的一些技巧</a></h2><div class="post-meta">2018-04-08</div><div class="post-content"><h2 id="epoch"><a href="#epoch" class="headerlink" title="epoch"></a>epoch</h2><p>机器学习，或者说神经网络训练中，一个很重要的参数就是epoch。一个epoch就表示网络对整个数据集完成了一次训练。</p>
5<h2 id="iteration"><a href="#iteration" class="headerlink" title="iteration"></a>iteration</h2><p>迭代次数。即在一个epoch中，分成iteration个小的数据批次即minibatch，进行训练。例如样本总数为N，每个minibatch的batchsize为n，则$iteration = N / n$。</p></div><p class="readmore"><a href="/2018/04/08/深度网络训练中的一些参数/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/08/目标追踪经典算法系列/">目标追踪算法之研究背景</a></h2><div class="post-meta">2018-04-08</div><div class="post-content"><h3 id="研究背景"><a href="#研究背景" class="headerlink" title="研究背景"></a>研究背景</h3><p><strong>目标追踪之定义</strong></p>
6<p>目标跟踪就是在一段序列图像中的每幅图像中实时地找到感兴趣的运动目标。</p>
7<p>在计算机视觉应用中，追踪是一个非常重要的部分，涉及活动分析分类（activity analysis classification）、目标识别（object recognition）以及人机交互（human computer interference）。在军事制导、视觉导航、安全监控、智能交通等等领域有着广泛的应用。</p></div><p class="readmore"><a href="/2018/04/08/目标追踪经典算法系列/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/08/SINT/">Introduction of Siamese Instance Search for Tracking</a></h2><div class="post-meta">2018-04-08</div><div class="post-content"><h2 id="特点"><a href="#特点" class="headerlink" title="特点"></a>特点</h2><ul>
8<li><strong>孪生结构</strong></li>
9<li>没有应用模型更新</li>
10<li>没有遮挡检测</li>
11<li>没有采用追踪器集合</li></div><p class="readmore"><a href="/2018/04/08/SINT/">阅读更多</a></p></div><div class="post"><h2 class="post-title"><a href="/2018/04/08/SiamFC改进思路/">SiamFC算法改进思路</a></h2><div class="post-meta">2018-04-08</div><div class="post-content"><p>视频追踪问题中，目标通常是连续可微的。</p>
12<p>SiamFC利用全卷积孪生网络结构对搜索域和样本图像进行相似度匹配，实现追踪目标。</p>
13<p>本文分析了SiamFC在vot2015数据集上的追踪结果，总结出以下问题，并提出针对性的改进方案。</p>
14<h2 id="表现鲁棒"><a href="#表现鲁棒" class="headerlink" title="表现鲁棒"></a>表现鲁棒</h2><ul></div><p class="readmore"><a href="/2018/04/08/SiamFC改进思路/">阅读更多</a></p></div><nav class="page-navigator"><span class="page-number current">1</span><a class="page-number" href="/page/2/">2</a><a class="extend next" rel="next" href="/page/2/">下一页</a></nav></div></div><div class="pure-u-1-4 hidden_mid_and_down"><div id="sidebar"><div class="widget"><form action="//www.google.com/search" method="get" accept-charset="utf-8" target="_blank" class="search-form"><input type="text" name="q" maxlength="20" placeholder="Search"/><input type="hidden" name="sitesearch" value="http://worldhellooo.github.io"/></form></div><div class="widget"><div class="widget-title"><i class="fa fa-folder-o"> 分类</i></div><ul class="category-list"><li class="category-list-item"><a class="category-list-link" href="/categories/Computer-Vision/">Computer Vision</a></li><li class="category-list-item"><a class="category-list-link" href="/categories/Deep-Learning/">Deep Learning</a></li><li class="category-list-item"><a class="category-list-link" href="/categories/GUI/">GUI</a></li><li class="category-list-item"><a class="category-list-link" href="/categories/Reinforcement-Learing/">Reinforcement Learing</a></li><li class="category-list-item"><a class="category-list-link" href="/categories/Visual-Tracking/">Visual Tracking</a></li><li class="category-list-item">

14<a class="category-list-link" href="/categories/competition/">competition</a></li><li class="category-list-item"><a class="category-list-link" href="/categories/笔记/">笔记</a></li></ul></div><div class="widget"><div class="widget-title"><i class="fa fa-star-o"> 标签</i></div><div class="tagcloud"><a href="/tags/增强学习-目标追踪/" style="font-size: 15px;">增强学习 目标追踪</a> <a href="/tags/目标追踪/" style="font-size: 15px;">目标追踪</a> <a href="/tags/实例搜索-孪生结果-目标追踪/" style="font-size: 15px;">实例搜索 孪生结果 目标追踪</a> <a href="/tags/增强学习/" style="font-size: 15px;">增强学习</a> <a href="/tags/目标追踪-研究背景-主流思路/" style="font-size: 15px;">目标追踪 研究背景 主流思路</a> <a href="/tags/RoIPooling-翻译/" style="font-size: 15px;">RoIPooling 翻译</a></div></div><div class="widget"><div class="widget-title"><i class="fa fa-file-o"> 最新文章</i></div><ul class="post-list"><li class="post-list-item"><a class="post-list-link" href="/2018/12/12/关于概率和熵的一些概念/">关于概率和熵的一些概念</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/05/18/deepmind论文解读/">DeepMind论文：连接多巴胺与元强化学习的新方法</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/26/pytorch doc阅读笔记/">pytorch doc阅读笔记</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/18/大数据竞赛经验谈/">大数据竞赛经验谈</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/12/基于PyQt的多源图像文字描述生成软件界面/">基于PyQt的多源图像文字描述生成软件界面</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/08/翻译 Region of interest pooling explained/">翻译 Region of interest pooling explained</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/08/深度网络训练中的一些参数/">深度网络训练中的一些技巧</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/08/目标追踪经典算法系列/">目标追踪算法之研究背景</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/08/SINT/">Introduction of Siamese Instance Search for Tracking</a></li><li class="post-list-item"><a class="post-list-link" href="/2018/04/08/SiamFC改进思路/">SiamFC算法改进思路</a></li></ul></div><div class="widget"><div class="widget-title"><i class="fa fa-external-link"> 友情链接</i></div><ul></ul><a href="http://blog.jcix.top" title="Zhangjc's blog" target="_blank">Zhangjc's blog</a><ul></ul><a href="http://chaway.github.io" title="Chaway's blog" target="_blank">Chaway's blog</a><ul></ul><a href="http://www.votchallenge.net" title="VOT Challenge" target="_blank">VOT Challenge</a><ul></ul><a href="http://www.visual-tracking.net" title="OTB Benchmark" target="_blank">OTB Benchmark</a></div></div></div><div class="pure-u-1 pure-u-md-3-4"><div id="footer">© <a href="/." rel="nofollow">WorldHellooo's Blog.</a> Powered by<a rel="nofollow" target="_blank" href="https://hexo.io"> Hexo.</a><a rel="nofollow" target="_blank" href="https://github.com/tufu9441/maupassant-hexo"> Theme</a> by<a rel="nofollow" target="_blank" href="https://github.com/pagecho"> Cho.</a></div></div></div><a id="rocket" href="#top" class="show"></a>

14<script type="text/javascript" src="/js/totop.js?v=0.0.0" async></script>

14<script type="text/javascript" src="//cdn.bootcss.com/fancybox/2.1.5/jquery.fancybox.pack.js" async></script>

14<script type="text/javascript" src="/js/fancybox.js?v=0.0.0" async></script>

14<link rel="stylesheet" type="text/css" href="/css/jquery.fancybox.css?v=0.0.0">

14<script type="text/javascript" src="/js/codeblock-resizer.js?v=0.0.0"></script>

14<script type="text/javascript" src="/js/smartresize.js?v=0.0.0"></script>

14</div></body></html>