【论文阅读笔记】Large-scale Video Classification with Convolutional Neural Networks

最新推荐文章于 2020-09-24 16:42:10 发布
最新推荐文章于 2020-09-24 16:42:10 发布
阅读量2.7k 收藏 1
点赞数
分类专栏： 论文阅读笔记
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/bpp94_github/article/details/82425856
版权

  • 论文贡献：

1.采用多种方法扩展CNN到视频识别，并用Sports-1M数据集进行验证和测试。

2.提出了将输入处理为低分辨率流和高分辨率流的方法，在不影响精确度的前提下显著减少了CNN的训练时间。

3.验证模型能够扩展到UCF101数据集上，且显著提升了精确率。

  • 时间信息融合模型：

a）单帧模型（Single-frame）：用于得到静态图像对视频分类的贡献。CNN结构：C(96,11,3)-N-P-C(256,5,1)-N-P-C(384,3,1)-C(384,3,1)-C(256,3,1)-P-FC(4096)-FC(4096)。C（d,f,s）中d代表卷积核个数，f代表卷积核大小f X f，s代表步长。N代表normalization layers，P代表池化层，参数均为2 X 2。分类器为softmax。

b）后期融合模型（Late Fusion）：模型中包含两个单帧模型，参数共享。两个模型的输入使间隔为15的两帧，通过卷积层后对两个模型的输出使用两个全连接层进行融合。单帧模型无法检测到运动信息，但是通过第一个卷积层能够得到全局运动特征。

c）早期融合模型（Early Fusion）：通过整个时间窗即时得到像素级的组合信息。修改单帧模型的第一层卷积核为11X11X3XT，T表示时间范围参数，本文中使用T=10或者1/3秒左右。像素数据的早期连接融合使网络能够精确地检测到局部物体的运动方向和速度。

d）慢融合（Slow Fusion）：该模型是上述两个的模型的中和。模型高层能够逐步得到更多时间和空间信息。第一层卷积核的T=4，clip为10帧，步长为2。第二和第三卷积层T=2，步长为2。

                                  

  • 加速CNN：

Fovea and context streams：将178X178的图像整体缩小为89X89，以及截取中间89X89的原图像，分别输入到CNN。如下图所示，可以提速并且不影响模型准确率。

                                       

  • 结果：

 

 

哈士奇的布偶
关注
  • 0
    点赞
  • 踩
  • 1
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
专栏目录
Large-Scale Video Classification with Convolutional Neural Networks
08-06
论文：Large-Scale Video Classification with Convolutional Neural Networks
【论文阅读】视频理解系列论文精读
三木今天学习了嘛の博客
10-28 1344
对视频理解中的一些经典文章做了精读，并简单记录
CNN（卷积神经网络）在视频动作分类中的应用
热门推荐
IT狗WeiWei
03-12 2万+
简介 Large-scale Video Classification with Convolutional Neural Networks Fusion Method Multi-resolution CNN Two-Stream Convolutional Networks for Action Recognition in Videos简介最近接触了一些卷积神经网络的只是以及其在视频动作分类中
【论文学习】Large-scale Video Classification with Convolutional Neural Networks
I am what i am
09-29 5016
Large-scale Video Classification with Convolutional Neural Networks原文地址粗略翻译摘要：卷积神经网络（CNN）是图像识别问题上一个非常强大的模型。受到这些结果的鼓舞，我们在大范围的视频分类中提供了广泛的实例评估，我们使用的数据集是100万个、属于487的类别的YouTube视频。为了利用局部时空信息，我们学习了多种方法来扩展CNN在
Large-scale Video Classification with Convolutional Neural Networks
xiaofei0801的博客
06-06 1862
1. Introduction        图像和视频在互联网上变得无所不在，这鼓励了开发可以分析其语义内容的不同应用程序的算法，包括搜索和摘要。近来，卷积神经网络（CNN）[15]已经被证明是一种有效的模型，用于理解图像内容，为图像识别，分割，检测和检索提供了最先进的结果[11,3,2， 20，9，18]。这些结果背后的关键因素是将网络扩展到数千万个参数和可以支持学习过程的大量标签数据集的技
Large-scale Video Classification with Convolutional Neural Networks(泛读)
小胜爱捣鼓
10-08 5403
一.文献名字和作者     Large-scale Video Classification with Convolutional Neural Networks, CVPR2014     二.阅读时间     2014年10月8日 三.文献的贡献点     文章的贡献点主要有三个：1.将CNN拓展，用于视频分类；     2.使用两种不同的分辨率的
行为识别论文笔记|SlowFusion|Large-scale Video Classification with Convolutional Neural Networks|
njuptalex的博客
09-24 907
Large-scale Video Classification with Convolutional Neural Networks 论文笔记 Andrej Karpathy et al. “Large-scale Video Classification with Convolutional Neural Networks” CVPR 2014 Motivation 缺少使用CNN进行视频分类的数据集和方法，所以推出了数据集 Sports-1M（1 million YouTube videos,
CVPR2020论文解析：视频分类Video Classification
吴建明wujianming_110117
03-21 3377
CVPR2020论文解析：视频分类Video Classification Rethinking Zero-shot Video Classification: End-to-end Training for Realistic Applications 论文链接：https://arxiv.org/pdf/2003.01455.pdf 摘要 深度学习（deep learning，DL）是在大型...
AlexNet - ImageNet Classification with Deep Convolutional Neural Networks 译文
07-12
AlexNet - ImageNet Classification with Deep Convolutional Neural Networks 译文
Large-Scale Software Architecture.pdf
09-25
Contents Preface xi Acknowledgments xvii 1 Introduction 1 1.1 What is Software Architecture 1 1.1.1 What software architecture is not 4 1.1.2 Attributes of software architecture 5 1.1.3 Definitions of other key architecture-related terms 7 1.1.4 Other types of architectures 8 1.2 Why Architect? 10 1.3 Architectural Viewpoint Summary 12 1.4 Other Software Architecture Approaches 16 1.4.1 The 4+1 Views 16 1.4.2 RM-ODP viewpoints 17 1.4.3 Bass architectural structures 1.4.4 Hofmeister software architecture views 1.5 Recommended Reading 2 RolesoftheSoftwareArchitect 21 2.1 Relationship to other key roles in development organization 25 Role: project management 25 Role: development team managers 25 Role: system architect/chief engineer 26 Role: chief software engineer 26 Role: hardware architect 27 Role: network architect 27 19 19 19 Role: technical leads of each release 28 Role: data architect 28 Role: systems engineering leads 28 Role: software systems engineering lead 29 2.2 Skills and Background for the Architect 29 2.3 Injecting Architecture Experience 31 2.4 Structuring the Architecture Team 32 2.5 Traps and Pitfalls Associated with the Role of Software Architect 33 2.5.1 Clear definition of leadership 34 2.5.2 Reporting structure for the software architect 34 2.5.3 Geographical location of software architect and technical leads 35 2.5.4 Architecture team size and composition 36 2.5.5 Software architect lifecycle participation 36 2.6 Recommended Reading 37 3 SoftwareArchitectureandtheDevelopment Process 39 3.1 Overview of Iterative Development 39 3.1.1 Overall process phases 40 3.1.2 Lifecycle stages 41 3.1.3 Architecture and agile processes 43 3.1.4 Start early, refine constantly 47 3.2 Requirements Management 48 3.2.1 Use cases and requirements engineering 48 3.2.2 Additional requirements that impact architecture 49 3.2.3 Requirements tracing 49 3.3 Management of the Technology Roadmap 50 3.3.1 External software products 50 3.3.2 Software technology management traps and pitfalls 53 3.3.3 Organizational technology roadmap 54 3.4 Effective Technical Meetings 55 3.4.1 Informal technical meetings 55 3.4.2 Peer reviews and inspections 56 3.4.3 Design reviews 57 3.4.4 Design communication meetings 57 3.4.5 Management meetings 57 3.4.6 Vendor presentations 58 3.4.7 Distributed technical meetings 58 3.5 Traps and Pitfalls of the Software Architecture Process Activities 59 The out-of-touch architect 59 Analysis paralysis 60 Design for reuse 60 Use cases 60 Schedule 60 3.6 Computer-Aided Software Engineering (CASE) Tools 61 3.7 Recommended Reading 62 vi Contents 4 ExampleSystemOverview 63 4.1 System Overview 64 4.2 Overview of System Interfaces 64 4.3 Constraints 67 4.4 Major Operational Requirements and Software Requirements 67 5 UMLQuickTour 69 5.1 UML Diagram Summary 69 5.2 General Diagramming Conventions 72 5.2.1 General UML features: stereotypes, tagged values, multi-instance 73 5.2.2 View labels 74 5.3 The Diagrams 75 5.3.1 Component instance diagrams 75 5.3.2 Class and subsystem diagrams 76 5.3.3 Interaction (sequence and collaboration) diagrams 77 5.3.4 Deployment diagrams 79 5.3.5 Statechart diagrams 80 5.3.6 Activity diagrams 81 5.4 Managing Complexity 81 5.4.1 Use Case focused modeling 82 5.4.2 Element focused modeling 82 5.4.3 Level of detail 83 5.4.4 Controlling the number of models 83 5.4.5 Use Supplemental Textural Information 85 5.5 Recommended Reading 85 6 SystemContext andDomainAnalysis 87 6.1 Conceptual Diagrams 87 6.2 Context Viewpoint 89 6.3 Domain Analysis Techniques 94 6.3.1 A formal analysis technique 95 6.3.2 Other techniques for finding domain entities 98 6.3.3 Analysis shortcuts 100 6.4 Analysis Viewpoints 101 6.4.1 Analysis Interaction Viewpoint 101 6.4.2 Analysis Focused Viewpoint 103 6.4.3 Analysis Overall Viewpoint 105 6.4.4 Candidate subsystem identification 107 6.5 Recommended Reading 108 7 Component DesignandModeling 111 7.1 Overview 111 7.1.1 Component-based development 111 7.1.2 Terminology 112 7.1.3 Communication and interfaces 115 7.1.4 Finding components 115 7.1.5 Qualities of component design 116 Contents vii 7.2 Component Viewpoint 116 7.2.1 Component communication 117 7.2.2 Component interfaces 118 7.2.3 Message-based component modeling 121 7.2.4 Combining interfaces and messaging 124 7.2.5 Comparison of interfaces and messaging 127 7.2.6 Mechanism and performance annotations 128 7.3 Component Interaction Viewpoint 131 7.3.1 Component to Component Interactions 131 7.4 Component State Modeling 133 7.5 Modeling Highly Configurable Component Architectures 137 7.6 Recommended Reading 137 8 SubsystemDesign 139 8.1 Terminology 139 8.2 Modeling Subsystems, Interfaces, and Layers 141 8.2.1 Subsystem Interface Dependency Viewpoint 141 8.2.2 Enhancing the Subsystem Dependency Views with layers 8.2.3 Top-level Dependencies 144 8.2.4 The Layered Subsystem Viewpoint 146 8.3 Mapping Subsystems and Layers to Implementation 151 8.3.1 Subsystems, layers, and build trees 151 8.3.2 Subsystems and components 8.4 Recommended Reading 154 9 TransactionandDataDesign 155 9.1 Logical Data Architecture 155 9.1.1 Logical data model stability 157 9.1.2 Effects of the stable logical data model 158 9.2 Logical Data Viewpoint 159 9.2.1 Logical Data View example 160 9.2.2 Logical Data View for messaging 9.3 Data Model Design – Other Considerations 9.3.1 Data models and layers 164 9.3.2 Data models and reflection 165 9.3.3 Mapping objects to relational database 166 9.4 Transaction Design 169 9.4.1 Transaction concepts 170 9.4.2 Modeling transaction dynamics 171 9.4.3 Transactions and interface design 173 9.5 Recommended Reading 174 10ProcessandDeployment Design 177 10.1 Physical Data Viewpoint 178 10.1.1 Modeling other storage attributes 179 10.1.2 Detailed physical storage modeling 181 10.2 Process Viewpoint 183 viii Contents 16 153 3 16 143 3 TEAMFLY Team-Fly ? 10.2.1 Processes and components 186 10.2.2 Process and component management 186 10.2.3 Process State Viewpoint 189 10.3 Deployment Viewpoint 193 10.3.1 Scalable node design 194 10.3.2 Backup/archive design 199 10.4 Recommended Reading 199 11ArchitectureTechniques 201 11.1 Architecture Development Techniques 201 11.1.1 Commonality and variability analysis 202 11.1.2 Design for change 203 11.1.3 Generative programming techniques 204 11.1.4 Building a skeleton system 205 11.1.5 Prototyping 206 11.1.6 Interface development – Design by Contract 206 11.1.7 Architectural description languages 208 11.1.8 Architecture evaluation 208 11.2 Software Partitioning Strategies – Separation of Concerns 208 11.2.1 Functional decomposition 209 11.2.2 Isolate donfiguration data 210 11.2.3 Isolate hardware-specific components 210 11.2.4 Isolate time-critical components 211 11.2.5 Separate domain implementation model from human interface 211 11.2.6 Separate domain implementation model from implementation technology 211 11.2.7 Separate main function from monitoring 212 11.2.8 Separate fault recovery processing 212 11.2.9 Adaptation of external interfaces 213 11.3 Software Changeability and Dependency Management 213 11.3.1 The stable dependencies principle (SDP) 214 11.3.2 Acyclic dependencies principle 215 11.3.3 Interface Separation Principle 216 11.4 Using Architectural Patterns 216 11.5 Integration Strategies 218 11.5.1 Data-only integration 219 11.5.2 Executable integration 220 11.6 Establishing Architecture to Support Development 221 11.6.1 Configuration and change management 221 11.6.2 Build management 222 11.6.3 Continuous integration 222 11.6.4 Anticipate multi-language development 223 11.6.5 Anticipate tactical development (scripting) 224 11.7 Recommended Reading 225 12ApplyingtheViewpoints 227 12.1 Bottom-Up Architecture Development 227 12.2 Top-Down Architecture Development 229 Contents ix 12.3 Message Protocol and Interface Development 231 12.4 Reengineering Existing Systems 233 12.5 Documenting the Architecture 233 12.6 Conclusions 235 12.6.1 Becoming an architect 235 12.6.2 State of the Practice 237 12.6.3 Looking forward 238 12.6.4 Final thoughts 240 12.7 Recommended Reading 241 Appendix:SummaryofArchitectural Viewpoints 243 Bibliography 251 Index 257
imagenet-classification-with-deep-convolutional-neural-networks原版和翻译..rar
06-09
imagenet-classification-with-deep-convolutional-neural-networks的原版和人工翻译word版本
4824-imagenet-classification-with-deep-convolutional-neural-networks.rar
07-09
ALEXNET相关学习资料，防止因为网络原因下不下来，备份了一份，侵删 如果发现积分不为0，可以私信我一下，一般一周内能调回0
NIPS2012-ImageNet Classification with Deep Convolutional Neural Networks
05-20
图像分类的一个里程碑，另外一个经典的CNN网络！
【论文阅读笔记】Beyond Short Snippets: Deep Networks for Video Classification
bbp94_github的博客
09-03 3527
主要目的： 视频分类 测试数据集： Sports-1M、UCF-101 方法概况： 使用在imageNet上预训练过的CNN（AlexNet或者GoogleLeNet）提取帧级特征，再将帧级特征和提取到的光流特征输入到池化框架或者LSTM进行训练，得到分类结果。 主要贡献： 1.提出采用CNN来得到视频级的全局描述，并且证明增大帧数能够显著提高分类性能。 2...
【论文阅读笔记】Learning Spatio-Temporal Representation with Pseudo-3D Residual Networks
bbp94_github的博客
10-16 2346
代码地址：https://github.com/ZhaofanQiu/pseudo-3d-residual-networks 主要贡献：以经济且有效的方式构建了仿3D卷积神经网络模型：P3D ResNet 出发点：3D 卷积神经网络能够同时学习到视频的时间和空间特征，但是存在所需内存大和计算成本高的问题，导致难以构建更深层的3D CNN，因此提出了用二维空间卷积和一维时间卷积来模拟3D卷积的想...
【论文阅读笔记】Visual Sentiment Prediction Based on AutomaticDiscovery of Affective Regions
bbp94_github的博客
09-10 814
主题：视觉情感分析 贡献： a）提出了一个深度框架，用于自动发现图像的情感区域，这些区域很可能会引发重要的情绪信息。该框架不依赖于图像中的对象类别，也不需要对边界框注释，比现有方法更加通用。 b）使用CNN构建视觉情感预测模型，该模型利用来自全局图像和局部图像区域的整体和局部信息。实验证明，最终得到的表达特征有助于视觉情感分类，在情感数据集的表现优于前人的方法。 c）实验结果表明，本文提出...
【论文阅读笔记】Learning Spatiotemporal Features with 3D Convolutional Networks
bbp94_github的博客
09-06 703
论文主要贡献： 1.验证了3D卷积神经网络能够同时捕捉外观和运动信息，是优秀的特征提取器。 2.在特定的构架内，3X3X3的卷积核在所有层中都表现出最好的效果。 3.3D卷积神经网络提取出的特征结合线性模型分类器所得到的结果在6个数据集（4个识别任务）中都有突出的性能。 好的视频特征描述应该包含如下特征： a）具有通用性，对任意类型的视频都能够有好的描述效果。 b）特征足够紧凑，紧...
SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS 代码
最新发布
07-27
很抱歉，根据提供的引用内容，我无法提供关于"SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS"代码的信息。引用的内容主要介绍了图上结点分类的半监督问题以及相关的研究方法和改进。如果您需要...

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值