<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[qq_40760171的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/qq_40760171</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; qq_40760171]]></copyright><item><title><![CDATA[论文阅读：Weakly-supervised Action Localization with Background Modeling]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/108320498</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/108320498</guid><author>qq_40760171</author><pubDate>Wed, 14 Oct 2020 09:40:26 +0800</pubDate><description><![CDATA[Weakly-supervised Action Localization with Background Modeling摘要Localization from Weak Supervision AssumeWeak Supervision定位

摘要

使用一个attention model以提取前景和背景以显示建模外观。
结合bottom-up，class-agnostic attention modules与top-down class-specific activation map。并应用自监督。]]></description><category></category></item><item><title><![CDATA[论文阅读：Weakly-Supervised Action Localization with Expectation-Maximization Multi-Instance Learning]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/108330509</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/108330509</guid><author>qq_40760171</author><pubDate>Wed, 14 Oct 2020 09:39:17 +0800</pubDate><description><![CDATA[Weakly-Supervised Action Localization with Expectation-Maximization Multi-Instance LearningMethodE StepM Step
Method

E Step

更新φφφ 通过最小化KL(qφ(z∣X)∣∣pθ(z∣X,yc))KL(q_φ(z|X) || p_θ(z|X, y_c))KL(qφ​(z∣X)∣∣pθ​(z∣X,yc​))。
这里假设posterior pθ(zt∣Xt,yc)p_θ(z_t|X_t, ]]></description><category></category></item><item><title><![CDATA[论文阅读：Temporal Segment Networks: Towards Good Practices for Deep Action Recognition]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/101462705</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/101462705</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:04:38 +0800</pubDate><description><![CDATA[作者贡献
本文旨在发现为视频中的动作识别设计有效的ConvNet架构的原理，并在有限的训练样本下学习这些模型。
第一个贡献是时域网（TSN）：这是一种基于视频的动作识别的新颖框架。这是基于远程时态结构建模的思想。它结合了稀疏的时间采样策略和视频级别的监督，可以使用整个动作视频进行有效的，高效的学习。
另一个贡献是作者研究了在时间段网络的帮助下学习视频数据上的ConvNet的一系列良好实践。
该方法...]]></description><category></category></item><item><title><![CDATA[论文阅读：Convolutional Neural Networks for Human Action Recognition]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/101433944</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/101433944</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:04:19 +0800</pubDate><description><![CDATA[3D Convolutional Neural Networks for Human Action Recognition作者贡献3D 卷积3D卷积体系结构模型正则化（没懂。。。）模型实现实验对TRECVID数据的动作识别对KTH数据的动作识别5 结论
作者贡献

应用3D卷积运算从视频数据中提取空间和时间特征以进行动作识别。这些3D特征提取器可在空间和时间维度上运行，从而捕获视频流中的运动信息。...]]></description><category></category></item><item><title><![CDATA[论文阅读：Two-Stream Convolutional Networks for Action Recognitionin Videos]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/101536430</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/101536430</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:03:59 +0800</pubDate><description><![CDATA[Two-StreamConvolutionalNetworks forActionRecognitioninVideos作者贡献双流网络空间流时间流时间ConvNet架构与先前表示的关系多任务学习实施细节ConvNets配置。训练测试在ImageNet ILSVRC-2012上进行预训练。多GPU训练使用来自OpenCV工具箱的[2]的现成GPU实现来计算光流。
作者贡献

提出了一个包含空间和时...]]></description><category></category></item><item><title><![CDATA[论文阅读：Convolutional Two-Stream Network Fusion for Video Action Recognition]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/101909133</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/101909133</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:03:43 +0800</pubDate><description><![CDATA[Convolutional Two-Stream Network Fusion for Video Action Recognition结论其他网络的缺点工作方法空间融合融合方法融合方法讨论在哪融合网络时间融合架构实现细节评估数据集和实验方案如何在空间上融合两个流？在哪里将两条流在空间上融合？从深层模型转变为更深层模型如何在时间上融合两个流？与最新技术的比较
结论

可以在卷积层融合时空网络，而不...]]></description><category></category></item><item><title><![CDATA[论文阅读：Temporal Structure Mining for Weakly Supervised Action Detection]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/103352344</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/103352344</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:03:30 +0800</pubDate><description><![CDATA[Temporal Structure Mining for Weakly Supervised Action Detection作者贡献与SMS比较方法定义时间结构挖掘阶段过滤器学习最大路径发现实验数据集和评估方法主干网络和双流融合实现细节消融实验和比较结果
作者贡献
提出了时间结构挖掘(TSM)方法
图2显示了提出的时间结构挖掘(TSM)方法的体系结构。如图所示，一个未修剪的视频被分割成多个段，...]]></description><category></category></item><item><title><![CDATA[论文阅读：UntrimmedNets for Weakly Supervised Action Recognition and Detection]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102662289</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102662289</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:02:58 +0800</pubDate><description><![CDATA[UntrimmedNets for Weakly Supervised Action Recognition and Detection作者贡献从未修剪的视频中学习剪辑采样UntrimmedNets特征提取模块分类模块选择模块视频预测训练行为识别与检测行为识别行为检测
作者贡献
问题：

为每个行为实例标注时间持续时间花费高且耗时。
这些时间的注释可能是主观的，在不用人之间不一致。

目标：从未裁...]]></description><category></category></item><item><title><![CDATA[论文阅读：Hide-and-Seek: Forcing a Network to be Meticulous for Weakly-supervised Object......]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102722856</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102722856</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:02:49 +0800</pubDate><description><![CDATA[Hide-and-Seek: Forcing a Network to be Meticulous for Weakly-supervised Object and Action Localization作者贡献方法弱监督对象定位隐藏随机图片块设置隐藏像素值Object localization network architectureWeakly-supervised action locali...]]></description><category></category></item><item><title><![CDATA[论文阅读：Two-Phase Learning for Weakly Supervised Object Localization]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102790446</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102790446</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:02:37 +0800</pubDate><description><![CDATA[Two-Phase Learning for Weakly Supervised Object Localization作者贡献Two-Phase Learning数据集Baseline结构First Phase LearningInference Conditional Feedback
作者贡献
问题：由于只有图像级别的注释，弱监督语义分割和定位只关注于图像最重要部分
解决：two-phase...]]></description><category></category></item><item><title><![CDATA[论文阅读：Attention-based Dropout Layer for Weakly Supervised Object Localization]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102941758</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102941758</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:02:27 +0800</pubDate><description><![CDATA[Attention-based Dropout Layer for Weakly Supervised Object Localization作者贡献ADL: Attention-based Dropout Layer实验
作者贡献
问题：网络的学习只能学习图像最具有辨别力的部分。
解决：提出了基于attention的dropout层，基于两个关键的组成部分：drop mask 和 importa...]]></description><category></category></item><item><title><![CDATA[论文阅读：Temporal Action Localization by Structured Maximal Sums]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/103336799</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/103336799</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:02:01 +0800</pubDate><description><![CDATA[Temporal Action Localization by Structured Maximal Sums概述作者贡献结构化预测定位结构化最大和
概述
将行为定位作为对任意长度时间窗口的结构化预测，其中每个窗口被评分为逐帧分类分数的总和。模型将每个行为的开始，中间和结束分类为单独的组件，允许模拟每个动作的时间演变，并利用此结构中存在的信息时间依赖关系。 在这个框架中，通过搜索结构化最大和来定位...]]></description><category></category></item><item><title><![CDATA[论文阅读：W-TALC: Weakly-supervised Temporal Activity Localization and Classiﬁcation]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/103705046</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/103705046</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:01:51 +0800</pubDate><description><![CDATA[W-TALC: Weakly-supervised Temporal Activity Localization and Classiﬁcation框架概述主要贡献方法特征提取弱监督层kkk-max多实例学习Co-Activity Similarity优化分类和定位实验
框架概述
W−TALCW-TALCW−TALC框架如图所示：


该方法利用现有的双流网络作为特征提取器。帧输入的数量取决于所使...]]></description><category></category></item><item><title><![CDATA[论文阅读：TURN TAP: Temporal Unit Regression Network for Temporal Action Proposals]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/106299116</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/106299116</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:01:16 +0800</pubDate><description><![CDATA[TURN TAP: Temporal Unit Regression Network for Temporal Action Proposals摘要

摘要

]]></description><category></category></item><item><title><![CDATA[论文阅读：Self-Erasing Network for Integral Object Attention]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102904796</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102904796</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:01:02 +0800</pubDate><description><![CDATA[Self-ErasingNetworkforIntegralObjectAttention作者贡献Self-Erasing Network思想
作者贡献
问题：随着训练迭代，对抗性擦除可能会导致注意区域逐渐扩展到非目标的区域。
解决：提出了Self-Erasing NetWork。
Self-Erasing Network
思想
对于给定的初始attention map，将图片在空间上分成三个区域...]]></description><category></category></item><item><title><![CDATA[论文阅读：AutoLoc: Weakly-supervised Temporal Action Localization in Untrimmed Videos]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102696925</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102696925</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:00:40 +0800</pubDate><description><![CDATA[AutoLoc: Weakly-supervised Temporal Action Localization in Untrimmed Videos作者贡献Outer-Inner-Contrastive LossForwardBackwardAutoLoc输入数据准备和特征提取Classiﬁcation BranchLocalization Branch
作者贡献

开发了一个新的弱监督TAL(...]]></description><category></category></item><item><title><![CDATA[论文阅读：Temporal Action Detection with Structured Segment Networks]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/106274984</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/106274984</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:00:25 +0800</pubDate><description><![CDATA[Temporal Action Detection with Structured Segment Networks摘要方法三阶段结构
摘要
问题：在未裁剪的视频中检测行为。
解决：

Structured Segment Networks(SSN): 可通过结构化的时间金字塔对每个行为实例的时间结构进行建模。 在金字塔的顶部，进一步使用了一个分解的判别模型，该模型包含两个分类器，分别用于对行为进行分类和确定完整性。 这使框架可以有效地将积极的proposals与背景的或不完整的proposals区分开，从]]></description><category></category></item><item><title><![CDATA[论文阅读：Action Recognition from Single Timestamp Supervisionin Untrimmed Videos]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102766171</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102766171</guid><author>qq_40760171</author><pubDate>Wed, 15 Jul 2020 11:00:17 +0800</pubDate><description><![CDATA[Action Recognition from Single Timestamp Supervision in Untrimmed Videos作者贡献从单个时间戳监督中识别采样分布初始化模型更新分布参数
作者贡献

提出了一种从单个时间戳初始化的采样分布，以选择相关的帧来训练行为识别分类。（因为时间戳的粗糙定位和每个行为有不同的时长，所以最初的抽样分布可能与实际情况不完全相符，如下图第一行）
提...]]></description><category></category></item><item><title><![CDATA[MatConvNet：6.代码(四)cnn_train.m 注释]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102372732</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102372732</guid><author>qq_40760171</author><pubDate>Mon, 21 Oct 2019 10:33:40 +0800</pubDate><description><![CDATA[原文链接：https://blog.youkuaiyun.com/qq_20259459/article/details/54602056
博主博客地址：http://blog.youkuaiyun.com/qq_20259459
作者邮箱（ jinweizhi93@gmai.com ）
cnn_train.m
%调用cnn_train：
% [ net, info ] = cnn_train(net, imdb, @g...]]></description><category></category></item><item><title><![CDATA[MatConvNet：5.代码(三)cnn_mnist_experiments.m 注释]]></title><link>https://blog.youkuaiyun.com/qq_40760171/article/details/102372530</link><guid>https://blog.youkuaiyun.com/qq_40760171/article/details/102372530</guid><author>qq_40760171</author><pubDate>Mon, 21 Oct 2019 10:33:12 +0800</pubDate><description><![CDATA[原文链接：https://blog.youkuaiyun.com/qq_20259459/article/details/54600368
博主博客地址：http://blog.youkuaiyun.com/qq_20259459
作者邮箱（ jinweizhi93@gmai.com ）
cnn_mnist_experiments.m
%% 原Code需要修改，不然会有错误发生，修改后代码如下：
%调用cnn_mnist...]]></description><category></category></item></channel></rss>