<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[soarflight_ing的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/soarflight_ing</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; soarflight_ing]]></copyright><item><title><![CDATA[Pytorch Ignite 使用方法]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/115591511</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/115591511</guid><author>soarflight_ing</author><pubDate>Sun, 11 Apr 2021 11:00:13 +0800</pubDate><description><![CDATA[Pytorch Ignite 使用方法

下载 pip install ignite
官方网址：https://pytorch.org/ignite/concepts.html

Engine
该框架的本质是class Engine，它是一种抽象形式，它在提供的数据上循环给定的次数，执行处理函数并返回结果：
  while epoch &lt; max_epochs:
    # run an epoch on data
    data_iter = iter(data)
    while True:
]]></description><category></category></item><item><title><![CDATA[Maven 的下载和idea中的配置]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/114789734</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/114789734</guid><author>soarflight_ing</author><pubDate>Sun, 14 Mar 2021 15:23:34 +0800</pubDate><description><![CDATA[Maven 的下载和配置
一、Maven介绍

Maven 是一个项目管理工具，它包含了一个 项目对象模型 (POM：Project Object Model)，一组标准集合，一个项目生命周期(Project Lifecycle)，一个依赖管理系统(Dependency Management System)，和用来运行定义在生命周期阶段(phase)中插件(plugin)目标(goal)的逻辑。

二、Maven 解决的问题

项目开发不仅仅是写写代码而已，期间会伴随着各种必不可少的事情要做，下面列举几个感]]></description><category></category></item><item><title><![CDATA[SSM整合笔记]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/114788529</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/114788529</guid><author>soarflight_ing</author><pubDate>Sun, 14 Mar 2021 14:58:26 +0800</pubDate><description><![CDATA[SSM整合笔记
一、配置准备
在进行配置前，先想清楚要配置哪些文件，如图，除web.xml外，其余三个配置文件名称均可自定义。

如图所示，一共有四个需要手动配置的文件：

web.xml:配置servlet、filter、listener
applicationContext.xml:配置相关的bean，与mybatis整合
SpringMvc.xml:用于配置视图解析器、控制器等
Mybatis.xml:用于生成Mybatis用到的实体类、Mapper接口、Map配置文件 (由于要与spring进行整合]]></description><category></category></item><item><title><![CDATA[Pyramidal Person Re-Identification via Multi-Loss Dynamic Training]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/114585421</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/114585421</guid><author>soarflight_ing</author><pubDate>Tue, 09 Mar 2021 15:46:20 +0800</pubDate><description><![CDATA[Pyramidal Person Re-Identification via Multi-Loss Dynamic Training

论文：Pyramidal Person Re-IDentification via Multi-Loss Dynamic Training,cvpr,2019.
链接：paper
代码：github

摘要
论文指出基于特征图水平切块（part-based）的行人重识别很大程度上依赖于目标检测算法的优劣。差的目标检测算法由于无法精确的框出行人，导致在进行局部特征相似度计算和]]></description><category></category></item><item><title><![CDATA[Pyramidal Convolution Rethinking Convolutional Neural Networks for Visual Recognition]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/114538449</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/114538449</guid><author>soarflight_ing</author><pubDate>Mon, 08 Mar 2021 16:59:08 +0800</pubDate><description><![CDATA[Pyramidal Convolution: Rethinking Convolutional Neural Networks for Visual Recognition

Pyramidal Convolution:Rethinking Convolutional Neural Networks for Visual Recognition,cvpr,2020.
链接：paper
代码：github


这篇论文不同于其他论文的关键点在于：卷积核的多尺度。同时为了尽可能的降低整体的计算复杂度，在每个gr]]></description><category></category></item><item><title><![CDATA[Salience-Guided Cascaded Suppression Network for Person Re-identification]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/114536967</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/114536967</guid><author>soarflight_ing</author><pubDate>Mon, 08 Mar 2021 16:33:19 +0800</pubDate><description><![CDATA[Salience-Guided Cascaded Suppression Network for Person Re-identification

论文：Salience-Guided Cascaded Suppression Network for Person Re-identification，CVPR,2020.
链接:paper

摘要
动机：
使用注意力机制来建模全局和局部特征作为最终的行人表示已经成为人再识别(Re-ID)算法的趋势。这些方法的一个潜在限制是它们集中于最显著的特征，但是对一个]]></description><category></category></item><item><title><![CDATA[ReID done right towards good practices for person re-identification]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/113337050</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/113337050</guid><author>soarflight_ing</author><pubDate>Thu, 28 Jan 2021 14:23:36 +0800</pubDate><description><![CDATA[ReID done right: towards good practices for person re-identification
本文中，我们采用了一种不同的方法，并仔细设计了一个简单的深层架构的每个组件，以及关键的策略，以有效地对其进行培训，从而实现人员的重新识别。我们对每个设计选择都进行了广泛的评估，得出了一份人员再识别的良好实践清单。通过遵循这些实践，我们的方法在四个基准数据集上以很大的优势超越了最先进的方法，包括带有辅助组件的更复杂的方法。
提升训练结果的技巧：


模型预训练
足够大的图像]]></description><category></category></item><item><title><![CDATA[Relation-Aware Global Attention]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/113183591</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/113183591</guid><author>soarflight_ing</author><pubDate>Tue, 26 Jan 2021 16:37:06 +0800</pubDate><description><![CDATA[Relation-Aware Global Attention

论文：Relation-Aware Global Attention for Person Re-identification,CVPR,2020.
链接：paper
代码：github

摘要
注意力机制通过关注重要特征和抑制不重要的特征来提到表示能力。在卷积神经网络中，注意力通常是通过局部卷积来学习的，它忽略了全局信息和隐藏关系。如何有效的利用远程上下文信息来全局的学习注意力还没有得到充分的研究。文中提出了一个有效的关系感知全局注意力模块]]></description><category></category></item><item><title><![CDATA[Feature Mask Network for Person Re-identification]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/111316907</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/111316907</guid><author>soarflight_ing</author><pubDate>Thu, 17 Dec 2020 15:46:51 +0800</pubDate><description><![CDATA[Feature Mask Network for Person Re-identification

论文：Let Features Decide for Themselves: Feature Mask Network for Person
Re-identification，cvpr,2018.
链接：paper
代码：github

摘要
虽然大多数方法侧重于学习特征和度量以获得更好的表示，但我们假设局部和全局上下文线索对于准确的身份匹配至关重要。为此文中提出了一种特征掩码网络(FMN)，利用ResN]]></description><category></category></item><item><title><![CDATA[Interaction-and-Aggregation Network for Person Re-identification]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/111053173</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/111053173</guid><author>soarflight_ing</author><pubDate>Fri, 11 Dec 2020 20:52:18 +0800</pubDate><description><![CDATA[Interaction-and-Aggregation Network for Person Re-identification

论文：Interaction-and-Aggregation Network for Person Re-identification，cvpr,2019
链接：paper
代码：github

摘要
背景：人的再识别(reID)极大地受益于学习鲁棒特征嵌入的深度卷积神经网络(CNNs)。然而，CNNs由于其固定的几何结构，在建模人的姿态和尺度的巨大变化时具有固有的局限性。
贡]]></description><category></category></item><item><title><![CDATA[Masked Graph Attention Network for Person Re-identification]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110882289</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110882289</guid><author>soarflight_ing</author><pubDate>Tue, 08 Dec 2020 21:03:21 +0800</pubDate><description><![CDATA[Masked Graph Attention Network for Person Re-identification

论文：Masked Graph Attention Network for Person Re-identification，cvpr,2019
链接：paper
代码：github

摘要
​		主流的行人重识别方法（ReID)主要关注个体样本图像与标签之间的对应关系，而忽略了整个样本集中丰富的全局互信息。为此，文中提出了一种掩码图注意网络（MGAT）方法。MGAT 利用所提取的特征信]]></description><category></category></item><item><title><![CDATA[GCNet]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110540129</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110540129</guid><author>soarflight_ing</author><pubDate>Thu, 03 Dec 2020 16:11:33 +0800</pubDate><description><![CDATA[GCNet

GCNet：Non-local Networks Meet Squeeze-Excitation Networks and Beyond,cvpr,2019.
链接：paper
代码：github

摘要
背景：Non-local Network 提供了一种捕获远程依赖关系的开创性方法，通过将特定于查询的全局上下文聚合得到每个查询位置。但是，本文发现，对于图中不同的查询位置，non-local network 网络建模的上下文信息几乎是相同的。
贡献：基于上述问题，文中建立了一个简化的基于查询]]></description><category></category></item><item><title><![CDATA[Human Semantic Parsing for Person Re-identification]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110489596</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110489596</guid><author>soarflight_ing</author><pubDate>Wed, 02 Dec 2020 16:28:50 +0800</pubDate><description><![CDATA[Human Semantic Parsing for Person Re-identification

论文：Human Semantic Parsing for Person Re-identification,cvpr,2018
链接：paper
代码：github

摘要
背景：现有方法都是利用检测框来对局部特征进行提取，这样的框精度较低，并且现有的方法涉及很多模块，相对比较复杂，通过大量实验证明使用简单且有效的训练方法效果是非常好的；
贡献：文中提出了SPReID，利用human semantic]]></description><category></category></item><item><title><![CDATA[Asymmetric Non-local Neural Networks]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110440411</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110440411</guid><author>soarflight_ing</author><pubDate>Tue, 01 Dec 2020 17:10:02 +0800</pubDate><description><![CDATA[Asymmetric Non-local Neural Networks

论文：Asymmetric Non-local Neural Networks for Semantic Segmentation,ICCV,2019.
链接：paper
代码：github

本文是在Non-local Neural Networks（cvpr,2018) 这篇论文的基础上改进得来的，non-local block Non-local模块作为语义分割任务中很有用的技术，但因为计算量较大，需要较大的显存开销，这阻碍了]]></description><category></category></item><item><title><![CDATA[Non-local Neural Networks]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110404071</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110404071</guid><author>soarflight_ing</author><pubDate>Mon, 30 Nov 2020 20:10:06 +0800</pubDate><description><![CDATA[Non-local Neural Networks

论文：Non - local Neural Networks，cvpr,2017.
链接：paper
代码：github

摘要
卷积神经网络和循环神经网络都是处理局部区域的操作，是典型的local operations。受经典的非局部均值（non-local means ）的启发，文中提出了一种non-local operations 用于捕获长距离依赖（long-range dependencies),(即如何建立图像上两个有一定距离的像素之间的联系]]></description><category></category></item><item><title><![CDATA[Multi-Scale Body-Part Mask Guide Attention for Person Re-identification]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110140613</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110140613</guid><author>soarflight_ing</author><pubDate>Wed, 25 Nov 2020 19:17:46 +0800</pubDate><description><![CDATA[Multi-Scale Body-Part Mask Guide Attention for Person Re-identification

论文:Multi-Scale Boyd-Part Mask Guide Attention for Person Re-identification
论文链接：MMGA
代码：

摘要
背景：由于人的姿态变化，不同的光照、遮挡、错位、背景杂乱等原因，对人的重识别仍然具有挑战性。
贡献：文中提出了一种多尺度身体局部语义掩模引导注意力网络（MMGA)，可以联合全身和局]]></description><category></category></item><item><title><![CDATA[MGCAM]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110140475</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110140475</guid><author>soarflight_ing</author><pubDate>Wed, 25 Nov 2020 19:01:56 +0800</pubDate><description><![CDATA[MGCAM

论文：Mask-guided Contrastive Attention Model for Person Re-Identifification，cvpr,2018
论文链接：paper
代码：github

摘要
背景：由于不同的背景杂乱，不同的视点和身体姿势，远远没有解决。如何提取对背景杂波不变性、鲁棒性强的特征是核心问题。
贡献：本文（1）引入了二值化的行人分割轮廓图作为额外输入，并与彩色图像合并为四通道的新输入，然后设计一种基于分割轮廓的对比注意模型来学习背景无关的行人特征。（2）在]]></description><category></category></item><item><title><![CDATA[Beyond Scalar Neuron]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/110006766</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/110006766</guid><author>soarflight_ing</author><pubDate>Mon, 23 Nov 2020 20:25:37 +0800</pubDate><description><![CDATA[Beyond Scalar Neuron

论文：Beyond Scalar Neuron:Adopting V ector-Neuron Capsules for Long-Term Person Re-Identification
作者：悉尼科技大学
代码:code

摘要
背景：目前的人员再识别(reid)工作主要集中在一个人不太可能改变衣服的短期情景。然而，在长期重新定位的情况下，一个人有很大的机会改变衣服。一个复杂的再识别系统应该考虑到这样的变化。
主要贡献：
①本文提出引了一个名为Celeb-r]]></description><category></category></item><item><title><![CDATA[CBAM & BAM & DANet]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/109994964</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/109994964</guid><author>soarflight_ing</author><pubDate>Mon, 23 Nov 2020 14:18:32 +0800</pubDate><description><![CDATA[注意力模块集合（CBAM &amp; BAM &amp; DANet）

论文：Convolutional Block Attention Module &amp; Bottleneck Attention Module ,2018  &amp; Dual Attention Network for Scene Segmentation, cvpr,2019
代码：CBAM&amp;BAMcode     DANet code

BAM
介绍
文中把重心放在了Attention对于一般深度神经网络的影响上，]]></description><category></category></item><item><title><![CDATA[Spindle Net]]></title><link>https://blog.youkuaiyun.com/soarflight_ing/article/details/109900899</link><guid>https://blog.youkuaiyun.com/soarflight_ing/article/details/109900899</guid><author>soarflight_ing</author><pubDate>Sat, 21 Nov 2020 16:21:10 +0800</pubDate><description><![CDATA[Spindle Net

论文：Spindle Net:Person Re-Identification with Human Body Region Guide Feature Decomposition and Fusion,cvpr,2017
论文地址：链接
代码：SpindleNet

摘要
背景：由检测算和姿势变化引起的行人身体不对准的问题会为不同图像间的特征匹配造成严重的影响。
贡献：本文提出了一种基于人体区域引导的多阶段特征分解和树形结构的竞争特征融合的新型卷积神经网络-SpindleNet，]]></description><category></category></item></channel></rss>