CV技术指南.pdf

【实例截图】

【核心代码】

目录
计算机视觉入门路线.........................................................................................................................7
必学的内容.................................................................................................................................7
可以学一学的内容.....................................................................................................................8
NMS(非极大值抑制)总结................................................................................................................ 11
标准 NMS..................................................................................................................................11
Soft NMS....................................................................................................................................13
DIoU NMS..................................................................................................................................15
Pytorch 代码调试工具--torchsnooper.............................................................................................17
池化技术总结...................................................................................................................................21
池化的作用...............................................................................................................................21
池化回传梯度...........................................................................................................................21
最大池化与平均池化的使用场景...........................................................................................22
空间金字塔池化.......................................................................................................................23
RoI 池化.....................................................................................................................................23
其他类型的池化.......................................................................................................................24
CNN 可视化技术总结（一）--特征图可视化................................................................................25
CNN 可视化方法.......................................................................................................................26
直接可视化...............................................................................................................................26
反卷积网络( deconvnet )......................................................................................................... 26
反池化 Unpooling.............................................................................................................27
修正 Rectification..............................................................................................................27
Filtering..............................................................................................................................28
反卷积网络特征可视化结果...........................................................................................28
导向反向传播...................................................................................................................29
CNN 可视化技术总结（二）--卷积核可视化................................................................................31
卷积核可视化的原理...............................................................................................................31
实现代码...................................................................................................................................31
可视化效果图...........................................................................................................................34
CNN 可视化技术总结（三）--类可视化........................................................................................36
本文档来源于公众号《CV 技术指南》，欢迎用于学术分享，禁止用于商业行为
3
CAM（Class Activation Map）.................................................................................................36
Grad-CAM..................................................................................................................................37
CNN 可视化技术总结（四）--可视化工具与项目........................................................................40
CNN-Explainer...........................................................................................................................40
一些可视化特征图、卷积核、热力图的代码。..................................................................42
结构可视化工具.......................................................................................................................42
网络结构手动画图工具...........................................................................................................44
CNN 结构演变总结（一）经典模型..............................................................................................46
LeNet......................................................................................................................................... 46
AlexNet(2012)............................................................................................................................47
NiN(2014)..................................................................................................................................48
VGG(2014).................................................................................................................................48
GoogLeNet(2014)......................................................................................................................49
Inception_v2 和 Inception_v3..................................................................................................51
Inception_v4, Inception_ResNet_v1 和 v2...............................................................................52
ResNet(2015).............................................................................................................................53
WRN(2017)................................................................................................................................53
DenseNet(2018)........................................................................................................................54
CNN 结构演变总结（二）轻量化模型..........................................................................................57
Xception(2017)..........................................................................................................................57
MobileNet_v1(2017).................................................................................................................58
MobileNet_v2(2018).................................................................................................................59
MobileNet_v3(2019).................................................................................................................60
ShuffleNet_v1(2018).................................................................................................................61
ShuffleNet_v2(2018).................................................................................................................63
SqueezeNet(2017).....................................................................................................................64
PELEE(2019)...............................................................................................................................65
CSPNet(2020)............................................................................................................................ 66
CNN 结构演变总结（三）设计原则..............................................................................................68
本文档来源于公众号《CV 技术指南》，欢迎用于学术分享，禁止用于商业行为
4
提升模型的表示能力的结构或方式.......................................................................................68
模型的设计原则.......................................................................................................................70
轻量化模型设计原则...............................................................................................................71
数据增强方法总结...........................................................................................................................74
数据增强的作用.......................................................................................................................74
数据增强的分类.......................................................................................................................74
常用方法...................................................................................................................................75
Cutout(2017).............................................................................................................................75
Random Erasing(2017)..............................................................................................................76
Mixup(2018)..............................................................................................................................77
Hide-and-Seek(2018).................................................................................................................77
CutMix(2019).............................................................................................................................78
GridMask(2020).........................................................................................................................79
FenceMask(2020)......................................................................................................................81
KeepAugment(2020).................................................................................................................81
其它的数据增强方法...............................................................................................................83
多样本数据增强方法...............................................................................................................83
CV 方向的高效阅读英文文献方法总结.........................................................................................85
为何一定要摆脱翻译软件.......................................................................................................85
如何摆脱翻译软件...................................................................................................................86
论文如何做笔记.......................................................................................................................86
特征金字塔技术总结.......................................................................................................................88
两种构建方式...........................................................................................................................88
ASPP(2017)................................................................................................................................88
FPN(2017)..................................................................................................................................89
PANet(2018).............................................................................................................................. 90
RFB(2018)..................................................................................................................................91
ASFF(2019).................................................................................................................................92
FPT(2020)...................................................................................................................................93
本文档来源于公众号《CV 技术指南》，欢迎用于学术分享，禁止用于商业行为
5
YOLOF(2021)..............................................................................................................................94
其它改进的特征金字塔...........................................................................................................95
参考论文...................................................................................................................................96
论文创新的常见思路总结...............................................................................................................98
思路...........................................................................................................................................98
注意力机制技术总结.....................................................................................................................101
STN(2015)................................................................................................................................101
OPAM(2017)............................................................................................................................102
Residual Attention(2017)........................................................................................................103
BAM(2018).............................................................................................................................. 104
CBAM(2018)............................................................................................................................104
Non-Local(2018)......................................................................................................................105
PAN(2018)................................................................................................................................106
Squeeze-and-Excitation(2018)................................................................................................107
CCNet(2019)............................................................................................................................108
GCNet(2019)............................................................................................................................109
DANet(2019)............................................................................................................................110
Coordinate Attention(2021)....................................................................................................111
参考论文.................................................................................................................................112
归一化方法总结.............................................................................................................................113
LRN(2012)................................................................................................................................113
Batch Normalization(2015).....................................................................................................113
BN、LN、IN 和 GN 的区别与联系........................................................................................115
Instance Normalization(2016).................................................................................................116
Layer Normalization (2016).....................................................................................................116
Group Normalization(2018)....................................................................................................117
Weights Normalization(2016).................................................................................................118
Batch Renormalization(2017)................................................................................................. 118
Cross-GPU BN(2018)............................................................................................................... 119
本文档来源于公众号《CV 技术指南》，欢迎用于学术分享，禁止用于商业行为
6
FRN(2019)................................................................................................................................120
Cross-Iteration BN(2020).........................................................................................................121
参考论文.................................................................................................................................123
高效阅读英文文献的方法总结（二）........................................................................................ 124
损失函数技术总结.........................................................................................................................128
损失函数分类与应用场景.....................................................................................................128
L1 loss......................................................................................................................................128
L2 loss......................................................................................................................................129
Negative Log-Likelihood Loss..................................................................................................129
Cross-Entropy Loss..................................................................................................................130
Hinge Embedding Loss............................................................................................................131
Margin Ranking Loss............................................................................................................... 132
Triplet Margin Loss..................................................................................................................132
KL Divergence Loss..................................................................................................................133
欠拟合与过拟合技术总结.............................................................................................................134
欠拟合与过拟合的概念.........................................................................................................134
欠拟合产生的原因与解决方法.............................................................................................135
过拟合产生的原因与解决方法.............................................................................................136
计算机视觉专业术语总结.............................................................................................................137
backbone、head、neck 和 fine-tune....................................................................................137
Preprocess 和 Postprocess......................................................................................................138
先验知识.................................................................................................................................139
embedding.............................................................................................................................. 140
feature map.............................................................................................................................140
池化.........................................................................................................................................141
语义信息.................................................................................................................................141
总结.........................................................................................................................................141