hanguangxin/notes-of-classic-detection-papers

Notes of Classic Detection Papers, including GoogLeNet, ResNet, DenseNet, Faster-RCNN, YOLOv1, SSD, RetinaNet, CornerNet, CenterNet (triplet), CenterNet (object as points), FCOS

11篇论文的笔记，总字数 3062 + 6153 + 1927 + 3092 + 5481 + 4203 + 2868 + 5381 + 2884 + 1975 + 2899 +3871 = 43796

我个人认为，需要到能给其他人讲论文的程度，这里的Notes可以满足您80~90%的需求

若有前后冲突，请以后为准。时间仓促，可能会有一些小瑕疵，欢迎批评和交流！

如果有用，还请给个Star~

To Myself

本文列出的论文是Detection方向入门必读的论文。

写作这些Notes的初衷，是希望能找到一个相对高效和稳定的方法，来对一个方向建立起基本的知识体系。也希望这个方法能推广应用到我对于其他的方向的学习上。

从零基础入门Detection，到现在完成这些工作，大概花了3个月的时间。

一方面，我觉得达到这样的程度，耗费的时间还是有点长的，不过这算是人生的第一次，也情有可原。吸取经验和教训就好了！

另一方面，经过对这些论文的精读，我能明显感觉到自己知识的深度和广度有了很大的提升，成长还是很明显的（当然还是很菜）。

总结一下，我认为自己在这个过程中有这些收货：

1. 从结果上看，我基本上算是建立了对Detection方向的基本理解和感知，在Computer Vision上，算是扣好了第一颗扣子。

2. 从方法上看，今后我要面对一个全新的方向，我也有了方法论，知道该如何下手。

3. 对于问题的细致思考和辩证思考、钻研，对于思维的提升有很大的帮助。要多问自己为什么。

   现在如果让我去讲这些论文，我相信自己能够讲得明白！

当然，也存在一些不足：

1. 在之后读论文的时候，应该对照着代码一起来。

   这样，对于论文中一些难以理解的点，可以直接去看代码实现，很多疑难问题就可以迎刃而解。

   而且，这对于代码能力的提高也有好处。

2. 从具体的流程上，也有必须要优化的地方。

   • 首先，入门一开始的时候，没有必要去一篇一篇地去做Overview。

     要了解整个方向，直接上综述就好了。

     所谓的第一轮可以大幅度地简化。

   • 其次，读论文的时候一定尽可能带上代码。

     所谓的后两轮也可以合并到一起。

   这样，原本的 “Overview + Paper + Code” 就可以变为 “1~2 Survey + Paper & Code”

How To Use the Notes

因GitHub对公式的支持不好，所以此处只是提供源文件

若您需要阅读，建议下载本仓库，以获得最好的阅读体验（如果不想下载可以点击链接到CSDN，但是格式会有点不舒服）

在您下载源文件到本地后，可以通过 [Detection Paper List.md](./Detection Paper List.md) 的链接快速打开对应的Notes源文件

在具体使用上，我推荐您：

1. 将博客作为参考，以博客的行文思路去理解论文（博客的行文思路来自于Andrew Ng关于 “Reading Research Paper” 的课程）。
2. 对于一些不明白的地方，去找博客的相应位置。

Idea of Writing

• Topic

  论文的题目

• Motivation

  包含论文要解决的问题、论文的Idea等

• Technique

  论文的模型、顶层的技术等

• Key Element

  一些底层的技术、关键概念、细节等

• Math

  所有的数学内容，包括数学表示、损失函数等

• Use Yourself

  对于论文一些要点的个人理解、一些在今后可能有用的idea等

• Relativity

  相关的论文、博客等

Sections of Each Notes

Overview

Topic	Motivation	Technique	Key Element	Use Yourself / Math
GoogLeNet	sparse structure & dense matrix	Inception	Inception path 1x1 conv	Architecture data augmentation model ensemble Test Time Augmentation Inception
ResNet	problems to slove identity mapping	residual learning residual block ResNet Architecture	shortcut connection feature concatenate	ResNet Architecture residual block shortcut connection ResNet V2
DenseNet	Problem to Solve Modifications	DenseNet Architecture Advantages	Dense Block Transition Layers Growth Rate Bottleneck Structure	Bottleneck Structure Feature Reuse Transition Layers
Faster RCNN	Problem to Solve Proposal with CNN	Architecture Region Proposal Network RoI Pooling Anchor	anchor/proposals/bbox Pyramids Positive & Negative Label Sampling Strategy 4-Step Alternating Training	Architecture Region Proposal Network RoI Pooling Anchor
YOLO v1	Problem to Solve Detection as Regression	YOLO v1 Architecture	Grid Partition Model Output “Responsible” NMS Getting Detections Error Type Analysis Data Augmentation	Classification & Detection Explicit & Implicit Grid Handle Small Object Multi-Task Output Weighted Loss NMS Data Augmentation
SSD	Problem to Solve Contributions	SSD Architecture Pros & Cons	Higher Speed What Is Resample Low & High Level Feature Small Object Difficulties Data Flow Anchor & GtBox Matching Foreground & Background Hard Negative Mining NMS Data Augmentation Testing (Inferencing) Step Performance Analysis Model Analysis	Convolution For Speed Feature Pyramids Fashion Positive & Negative Imbalance Zoom In & Zoom Out
RetinaNet	Problem to Solve Negative ==> Easy	Focal Loss RetinaNet	Class Imbalance Feature Pyramid Network ResNet-FPN Backbone Classify & Regress FCN Post Processing Anchor Design Anchor & GT Matching prior $\pi$ Initialization Ablation Experiments	Data Analysis Feature Pyramid Network More Scale $\not= $ Better
CornerNet	Problem to Solve KeyPoints (Anchor-Free)	CornerNet Stacked Hourglass Network Prediction Module Corner Pooling	Why CornerNet Better? Why Corner Pooling Works Grouping Corners Getting Bounding Box Data Augmentation Ablation Experiments	Loss Function Corner Pooling Math
CenterNet (triplets)	Problem to Solve Idea Intuition	CenterNet Architecture Center Pooling Cascade Corner Pooling Central Region Exploration	Baseline：CornerNet Generating BBox Training Inferencing Ablation Experiment Error Analysis Metric AP & AR & FD Small & Medium & Large	Central Region Loss Function
CenterNet (Object as Points)	Problem to Solve Idea	CenterNet Architecture	Center Point & Anchor Getting Ground-Truth Model Output Data Augmentation Inference TTA Compared with SOTA Additional Experiments	Loss Function KeyPoint Loss $\text{L}k$ Offset Loss $\text{L}{off}$ Size Loss $\text{L}_{size}$
[FCOS](./[paper reading] FCOS.md)	Idea Contribution	FCOS Architecture Center-ness Multi-Level FPN Prediction	Prediction Head Training Sample & Label Model Output Feature Pyramid Inference Ablation Study FCN & Detection FCOS $vs.$ YOLO v1	Symbol Definition Loss Function Center-ness Remap of Feature & Image
[Cascade R-CNN](./[paper reading] Cascade R-CNN.md)	Problem to Solve Motivation	Architecture Advantages	Degradation of High IoU Quality Match Cascade Regression Cascade Detection Iterative BB Regression Integral Loss Training Strategy Inference Strategy implement Details Generalization Capacity Ablation Experiments Comparison with SOTA	Loss Function of Stages Standard BBox Regression Iterative BB Regression Classification

Details

[GoogLeNet](./[paper reading] GoogLeNet.md)

• Basic Concept
  • 结构的稀疏连接
    • 层间连接的稀疏结构
    • 特征连接的稀疏结构
  • 稀疏/密集分布的特征集
• motivation
  • problem to solve
    • sparse structure (of network)
    • dense matrix (of features)
  • sparse structure & dense matrix
• technique
  • Inception
• key elements
  • Inception path
  • 1x1 conv
• use yourself
  • Architecture
  • data augmentation
  • model ensemble
  • Test Time Augmentation
• blogs
• modifications
  • Inception V2

[ResNet](./[paper reading] ResNet.md)

• summary
• motivation
  • problems to slove
    • network degradation
    • vanishing/exploding gradient
  • Relationship of degradation & gradient vanishing
  • identity mapping
• technique
  • residual learning
  • residual block
  • ResNet Architecture
• key element
  • shortcut connection
  • feature concatenate
• math
  • forward/backward propagation
  • residual learning
    • address degradation
    • address gradient vanishing
• use yourself
• articles
  • ResNet V2
  • Shattered Gradients (Gradient correlation)
  • Ensemble-like behavior
  • Hard For Identity Mapping
• blogs

[DenseNet](./[paper reading] DenseNet.md)

• Motivation
  • Problem to Solve
  • Modifications
    • Skip Connection
    • Feature Concatenate
• Technique
  • DenseNet Architecture
  • Advantages
    • Parameter Efficiency & Model Compactness
    • Feature Reuse & Collective Knowledge
    • Implicit Deep Supervision
    • Diversified Depth
• Key Element
  • Dense Block
  • Transition Layers
    • Components
    • Compression
  • Growth Rate
  • Bottleneck Structure
• Math
• Use Yourself
  • Bottleneck Structure
  • Transition Layers
  • Feature Reuse
• Blogs

[Faster RCNN](./[paper reading] Faster RCNN.md)

• Motivation
  • Problem to Solve
  • Proposal with CNN
• Technique
  • Architecture
  • Region Proposal Network
  • RoI Pooling
  • Anchor
• Key Element
  • anchor/proposals/bbox
  • Pyramids
  • Positive & Negative Sample
  • Sampling Strategy
  • 4-Step Alternating Training
• Math
  • Loss Function
  • Coordinates Parametrization
• Use Yourself
  • Architecture
  • Region Proposal Network
  • RoI Pooling
  • Anchor
• Articles
  • R-CNN
  • YOLO V1
• Blogs

[YOLO v1](./[paper reading] YOLO v1.md)

• Motivation
  • Problem to Solve
  • Detection as Regression
• Technique
  • YOLO v1 Architecture
• Key Element
  • Grid Partition
  • Model Output
  • “Responsible”
  • NMS
  • Getting Detections
  • Error Type Analysis
  • Handle Small Object
  • Data Augmentation
• Use Yourself
  • Classification & Detection
  • Explicit & Implicit Grid
  • Multi-Task Output
  • Loss & Sample & box
  • NMS
  • Data Augmentation
• Math
  • Activation Function
  • Loss Function
• Articles
  • R-CNN Based
  • Deep MultiBox
  • OverFeat
  • MultiGrasp
• Blogs

[SSD](./[paper reading] SSD.md)

• Motivation
  • Problem to Solve
  • Contributions
• Technique
  • SSD Architecture
  • Pros & Cons
• Key Element
  • Higher Speed
  • What Is Resample
  • Low & High Level Feature
  • Small Object Difficulties
  • Data Flow
  • Anchor & GtBox Matching
  • Foreground & Background
  • Hard Negative Mining
  • NMS
  • Data Augmentation
  • Testing (Inferencing)
  • Performance Analysis
  • Model Analysis
• Use Yourself
  • Convolution For Speed
  • Feature Pyramids Fashion
  • Positive & Negative Imbalance
  • Zoom In & Zoom Out
• Math
  • Loss Function
  • Multi-Level & Anchor
  • Layer Output & Filters
  • Model Output
• Blogs

[RetinaNet](./[paper reading] RetinaNet.md)

• Motivation
  • Problem to Solve
  • Negative ==> Easy
• Technique
  • Focal Loss
  • RetinaNet
• Key Element
  • Class Imbalance
  • Feature Pyramid Network
  • ResNet-FPN Backbone
  • Classify & Regress FCN
  • Post Processing
  • Anchor Design
  • Anchor & GT Matching
  • prior $\pi$ Initialization
  • Ablation Experiments
• Math
  • Cross Entropy Math
    • Standard Cross Entropy
    • Balanced Cross Entropy
  • Focal Loss Math
• Use Yourself
  • Data Analysis
  • Feature Pyramid Network
  • More Scale $\not= $ Better
• Related Work
  • Two-Stage Method
  • One-Stage Method
• Related Articles
• Blogs

[CornerNet](./[paper reading] CornerNet.md)

• Motivation
  • Problem to Solve
  • KeyPoints (Anchor-Free)
• Technique
  • CornerNet
  • Stacked Hourglass Network
  • Prediction Module
  • Corner Pooling
• Key Element
  • Why CornerNet Better?
  • Why Corner Pooling Works
  • Grouping Corners
  • Getting Bounding Box
  • Data Augmentation
  • Ablation Experiments
• Math
  • Loss Function
  • Corner Pooling Math
• Use Yourself
  • Network Design
  • Intuition & Interpretability
  • Divide Task
  • TTA
• Related Work
  • Two-Stage
  • One-Stage
• Blogs

[CenterNet (triple)](./[paper reading] CenterNet (Triplets).md)

• Motivation
  • Problem to Solve
  • Idea
  • Intuition
• Technique
  • CenterNet Architecture
  • Center Pooling
  • Cascade Corner Pooling
  • Central Region Exploration
• Key Element
  • Baseline：CornerNet
  • Generating BBox
  • Training
  • Inferencing
  • Ablation Experiment
  • Error Analysis
  • Metric AP & AR & FD
  • Small & Medium & Large
• Math
  • Central Region
  • Loss Function
• Use Yourself
• Related Work
  • Anchor-Based Method
  • KeyPoint-Based Method
  • Two-Stage Method
  • One-stage Method
• Problems

[CenterNet (Object as Points)](./[paper reading] CenterNet (Object as Points).md)

• Motivation
  • Problem to Solve
  • Idea
• Technique
  • CenterNet Architecture
• Key Element
  • Center Point & Anchor
  • Getting Ground-Truth
  • Model Output
  • Data Augmentation
  • Inference
  • TTA
  • Compared with SOTA
  • Additional Experiments
• Math
  • Loss Function
  • KeyPoint Loss $\text{L}_k$
  • Offset Loss $\text{L}_{off}$
  • Size Loss $\text{L}_{size}$
• Use Yourself
• Related work
  • Anchor-Based Method
  • KeyPoint-Based Method

[FCOS](./[paper reading] FCOS.md)

• Motivation
  • Idea
  • Contribution
• Techniques
  • FCOS Architecture
  • Center-ness
  • Multi-Level FPN Prediction
• Key Elements
  • Prediction Head
  • Training Sample & Label
  • Model Output
  • Feature Pyramid
  • Inference
  • Ablation Study
  • FCN & Detection
  • FCOS $vs.$ YOLO v1
• Math
  • Symbol Definition
  • Loss Function
  • Center-ness
  • Remap of Feature & Image
• Use Yourself
• Related Work
  • Drawbacks of Anchor
  • DenseBox-Based
  • Anchor-Based Detector
  • YOLO v1
  • CornerNet

[Cascade R-CNN](#./[paper reading] Cascade R-CNN.md)

• Motivations
  • Problem to Solve
  • Motivation
• Techniques
  • Architecture
  • Advantages
• Key Elements
  • Degradation of High IoU
  • Quality Match
  • Cascade Regression
  • Cascade Detection
  • Iterative BB Regression
  • Integral Loss
  • Training Strategy
  • Inference Strategy
  • implement Details
  • Generalization Capacity
  • Ablation Experiments
  • Comparison with SOTA
• Math
  • Loss Function of Stages
  • Standard BBox Regression
  • Iterative BB Regression
  • Classification
• Use Yourself
• Related Work