MachineLearning-Papers_Survey

機械学習関連の論文 Survey 用レポジトリです。
論文まとめ記事は、Issues に記載しています。進捗は、Projects ページ で管理しています。

■ 構成

◎ 基礎系（基礎モデル）

CNN

• [ResNet] Deep Residual Learning for Image Recognition
• Spatial Transformer Networks

GCN

• Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
• Semi-Supervised Classification with Graph Convolutional Networks
• [R-GCN] Relational Graph Convolutional Network

RNN

• [RNN] Recursive Neural Network
• [LSTM] long short-term memory
• [GRU] gated recurrent unit

Transformer

• [Transformer] Attention Is All You Need
• [Vision Transformer] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• TransGAN: Two Transformers Can Make One Strong GAN

VAE

• [VAE] Auto-Encoding Variational Bayes
• [VQ-VAE] Neural Discrete Representation Learning
• β-VAE: LEARNING BASIC VISUAL CONCEPTS WITH A CONSTRAINED VARIATIONAL FRAMEWORK

GANs

• [GAN] Generative Adversarial Networks
• [DCGAN] Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• [cGAN] Conditional Generative Adversarial Nets
• [WGAN] Wasserstein GAN
• [WGAN-gp] improved Training of Wasserstein GANs
• SAGAN [Self-Attention Generative Adversarial Networks]
• InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets
• [RSGAN,RGAN,RaGAN] The relativistic discriminator: a key element missing from standard GAN
• GAN-Tree: An Incrementally Learned Hierarchical Generative Framework for Multi-Modal Data Distributions
• A U-Net Based Discriminator for Generative Adversarial Networks

Flow-based generative model

• NICE: NON-LINEAR INDEPENDENT COMPONENTS ESTIMATION
• Real NVP [Density estimation using Real NVP]
• Glow [Generative Flow with Invertible 1×1 Convolutions]
• i-ResNets [Invertible residual networks]
• Residual Flows for Invertible Generative Modeling

Autoregressive Models

• [PixelRNN, PixelCNN] Pixel Recurrent Neural Networks

meta-learning, few-shot learning

• MAML:Model Agnostic Meta-Learning for Fast Adaption

Neural-ODE

• [Neural-ODE] Neural Ordinary Differential Equations
• Augmented Neural ODEs

Neural Processes

• Conditional Neural Processes
• Neural Processes

◎ アプリケーション系（CV）

Image Classification

• xxx

Semantic Segmentation

• U-Net: Convolutional Networks for Biomedical Image Segmentation
• [PSPNet] Pyramid Scene Parsing Network
• Pyramid Attention Network for Semantic Segmentation
• [DeepLab v3+] Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
• Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks
• Hypercolumns for Object Segmentation and Fine-grained Localization
• Tversky loss function for image segmentation using 3D fully convolutional deep networks
• Boundary loss for highly unbalanced segmentation
• Human Parsing
  • [JPPNet] Look into Person: Joint Body Parsing & Pose Estimation Network and A New Benchmark
  • [CE2P] Devil in the Details: Towards Accurate Single and Multiple Human Parsing
  • Graphonomy: Universal Human Parsing via Graph Transfer Learning
  • Hierarchical Human Parsing with Typed Part-Relation Reasoning
  • [CorrPM] Correlating Edge, Pose with Parsing

Object Detection

• Fast R-CNN
• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
• Focal Loss for Dense Object Detection

Instance Segmentation

• Mask R-CNN
• Human Parsing
  • Parsing R-CNN for Instance-Level Human Analysis

Pose Estimation

• DensePose: Dense Human Pose Estimation in the Wild

Image Registration / geometric matching

• Convolutional neural network architecture for geometric matching

image-to-image

• [pix2pix] Image-to-Image Translation with Conditional Adversarial Networks
• [pix2pix-HD] High-Resolution_Image_Synthesis_and_Semantic_Manipulation_with_Conditional_GANs
• [CycleGAN] Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks
• [StarGAN] StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation
• [SPADE] Semantic Image Synthesis with Spatially-Adaptive Normalization
• [Neural Collage] Spatially Controllable Image Synthesis with Internal Representation Collaging
• Recapture as You Want
• Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation
• [Impersonator++] Liquid Warping GAN with Attention: A Unified Framework for Human Image Synthesis
• Focal Frequency Loss for Generative Models
• few-shot learning
  • SinGAN: Learning a Generative Model from a Single Natural Image
  • [DeepSIM] Deep Single Image Manipulation

noize-to-image

• [PGGAN] Progressive Growing of GANs for Improved Quality, Stability, and Variation
• [StyleGAN] A Style-Based Generator Architecture for Generative Adversarial Networks
• [StyleGAN2] Analyzing and Improving the Image Quality of StyleGAN
• few-shot learning
  • Training Generative Adversarial Networks with Limited Data / StyleGAN2 with adaptive discriminator augmentation (ADA)
  • Data-Efficient GANs with DiffAugment
  • [Lightweight GAN] Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis

Inpainting

• [Deepfillv2] Free-Form Image Inpainting with Gated Convolution
• Pluralistic Image Completion
• Boundless: Generative Adversarial Networks for Image Extension

Person Image Generation

• Pose Guided Person Image Generation
• Disentangled Person Image Generation
• [Soft-Gated Warping-GAN] Soft-Gated Warping-GAN for Pose-Guided Person Image Synthesis

recommendation

• ViBE: Dressing for Diverse Body Shapes

text-to-image

• [StackGAN] Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Network

Optical Flow

• FlowNet: Learning Optical Flow with Convolutional Networks
• View Synthesis by Appearance Flow

3D Reconstruction

• param-to-3D / parametric 3D models
  • SMPL: A skinned multi-person linear model
  • [CAPE] Learning to Dress 3D People in Generative Clothing
  • TailorNet: Predicting Clothing in 3D as a Function of Human Pose, Shape and Garment Style
  • SIZER: A Dataset and Model for Parsing 3D Clothing and Learning Size Sensitive 3D Clothing
• image-to-3D / image-based 3D Reconstruction
  • none-parametric 3D models
    • Mesh R-CNN
    • Occupancy Networks: Learning 3D Reconstruction in Function Space
    • PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization
    • PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization
    • NormalGAN: Learning Detailed 3D Human from a Single RGB-D Image
    • using templete mesh
      • 3D Virtual Garment Modeling from RGB Images
  • parametric 3D models
    • [HMR] End-to-end Recovery of Human Shape and Pose
    • Multi-Garment Net: Learning to Dress 3D People from Images
    • Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single Images
    • BCNet: Learning Body and Cloth Shape from A Single Image
    • ExPose: Monocular Expressive Body Regression through Body-Driven Attention
    • I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image
• video-to-3D
  • TexMesh: Reconstructing Detailed Human Texture and Geometry from RGB-D Video
• texture mapping
  • 360-Degree Textures of People in Clothing from a Single Image
  • Learning to Transfer Texture from Clothing Images to 3D Humans
• camera localization
  • PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization
  • Geometry-Aware Learning of Maps for Camera Localization

Others

• xxx

◎ 理論系

xxx

• xxx

■ 機械学習系の論文調査や論文の読み方

• 論文調査

  • その論文の発展版を探すには、Google Scholer でその論文の被引用論文を探すのが手っ取り早い。
  • 基本的には、その論文の被引用数が多いものが注目の論文。但し最新の論文は当然被引用数は少なくなる。
  • 著者の所属組織も優れた論文であるか判断材料になり得る。例えば、GAFA や NVIDIA, Adobe などの論文は、優れた論文である可能性が高い
  • 基本的には、最新の論文ほど優れた実験結果を出しているので、Google Scholer で出来るだけ公開日時の最新のものを探す。
  • 論文の公式実装がないものは、自前実装の再現コストが高いので、公式実装があるものを優先的に探す。
  • 論文の公式実装の有無は、papers with code で確認できる。
  • github 上に論文公式実装がある場合は、スター数も優れた論文であるかの判断材料になり得る。
  • 探している分野の論文が１つ見つかれば、その論文を papers with code で検索する。その論文の "Tasks" が割り振られている場合は、その Tasks の名前からその分野の最新論文を調べることが出来る。

• 論文の読み方

  • 基本的に、「Abstract」→「Introduction」→「Conclusion」→「何をしたかの詳細」→「Experiments」→「Related Work」項目の順に読むのが効率的。Related Work は最悪飛ばしても良い。
  • Abstract や Introduction 内の "In this paper ...", "In this work ..." という文面や、Introduction の最後にある "the contributions of this paper ... as follows." という文面には注目。
  • その論文の未解決課題や問題点は、論文中の Future work や Conculution に述べられていることが多い。或いは、その論文を引用している論文の Related Work に述べられていることが多い。

■ 論文要約フォーマット（要約バージョン）

layout: post

title:  "論文タイトル"

date:   YYYY-MM-DD

categories: CV NLP Others

## 1. どんなもの？

## 2. 先行研究と比べてどこがすごいの？

## 3. 技術や手法の"キモ"はどこにある？

![Figure 1]({{ site.baseurl }}/assets/img/(cv, nlp, others)/(title)/figure1.png)

## 4. どうやって有効だと検証した？

## 5. 議論はあるか？

## 6. 次に読むべき論文はあるか？

### 論文情報・リンク

* [著者，"タイトル，" ジャーナル名，voluem，no.，ページ，年](論文リンク)

• 例）先端技術とメディア表現1 #FTMA15 from Yoichi Ochiai

■ 参考サイト

◎ 論文サイト

• arXiv
• Google Scholer
• Microsoft Academic

◎ 便利サイト

• DeepL
• papers with code
  • 論文実装の有無を確認できる。
• Hyper Collocation
• arXiv Vanity

◎ その他参考サイト

• arXivTimes
• ymym3412/acl-papers
• shunk031/paper-survey