一些文章

1、Proxy task 的设计

1.1 基于先验知识的设计

1.1.1 利用 Motion propagation 来设计 proxy task

​ Self-Supervised Learning via Conditional Motion Propagation (opens new window)

1.1.2 利用 Motion Prediction 来设计 proxy task

​ Dense Optical Flow Prediction From a Static Image (opens new window)

1.2 基于连贯性的设计

1.2.1 利用 Jigsaw Puzzles （物体的语义连贯性）来设计proxy task

​ Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles (opens new window)

1.2.2 利用视频中物体运动的时间连贯性来设计 proxy task

​ Shuffle and Learn: Unsupervised Learning using Temporal Order Verification (opens new window)

1.3 基于数据内部结构的设计

目前很火的基于contrastive learning的方法，包括NPID, MoCo, SimCLR等，我们可以将它们统一为instance discrimination [6]任务。如下图，这类任务通常对图片做各种变换，然后优化目标是同一张图片的不同变换在特征空间中尽量接近，不同图片在特征空间中尽量远离。

1.3.1 Momentum Contrast for Unsupervised Visual Representation Learning (MoCov1)

​ Momentum Contrast for Unsupervised Visual Representation Learning (opens new window)

1.3.2 Improved Baselines with Momentum Contrastive Learning (Mocov2)

​ Improved Baselines with Momentum Contrastive Learning (opens new window)

1.3.3 A Simple Framework for Contrastive Learning of Visual Representations (SimCLRv1)

​ A Simple Framework for Contrastive Learning of Visual Representations (opens new window)

1.3.4 Big Self-Supervised Models are Strong Semi-Supervised Learners (SimCLRv2)

​ Big Self-Supervised Models are Strong Semi-Supervised Learners (opens new window)

​ 这一篇本质上应该是算半监督学习

1.3.5 Bootstrap your own latent: A new approach to self-supervised Learning (BYOL)

​ Bootstrap your own latent: A new approach to self-supervised Learning (opens new window)

1.4 其他的设计

1.4.1利用 图像着色来设计 proxy task

​ Colorful Image Colorization (opens new window)

1.4.2 利用 Image Inpainting 来设计 proxy task

​ Context Encoders: Feature Learning by Inpainting (opens new window)

​

1.4.3 利用 Rotation Prediction 来设计 proxy task

​ Unsupervised Representation Learning by Predicting Image Rotations (opens new window)

​

1.4.4 利用 Instance Discrimination 来设计 proxy task

1.4.5利用 Counting 来设计 proxy task

1.4.6利用 Moving foreground segmentation 来设计 proxy task

1.4.7 利用Context Prediction 来设计 proxy task

​ Unsupervised Visual Representation Learning by Context Prediction (opens new window)

​

2、基于对比学习的文章

1.3.1 Momentum Contrast for Unsupervised Visual Representation Learning (MoCov1)

​ Momentum Contrast for Unsupervised Visual Representation Learning (opens new window)

1.3.2 Improved Baselines with Momentum Contrastive Learning (Mocov2)

​ Improved Baselines with Momentum Contrastive Learning (opens new window)

1.3.3 A Simple Framework for Contrastive Learning of Visual Representations (SimCLRv1)

​ A Simple Framework for Contrastive Learning of Visual Representations (opens new window)

1.3.4 Big Self-Supervised Models are Strong Semi-Supervised Learners (SimCLRv2)

​ Big Self-Supervised Models are Strong Semi-Supervised Learners (opens new window)

​ 这一篇本质上应该是算半监督学习

1.3.5 Bootstrap your own latent: A new approach to self-supervised Learning (BYOL)

​ Bootstrap your own latent: A new approach to self-supervised Learning (opens new window)

1.3.6 Data-Efficient Image Recognition with Contrastive Predictive Coding (CPCv2)

​ Data-Efficient Image Recognition with Contrastive Predictive Coding (opens new window)

1.3.7 Unsupervised Learning of Visual Featuresby Contrasting Cluster Assignments (SwAV)

​ Unsupervised Learning of Visual Featuresby Contrasting Cluster Assignments (opens new window)

1.3.8 Exploring Simple Siamese Representation Learning（SimSiam）

​ Exploring Simple Siamese Representation Learning (opens new window)

1.3.9 Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised VisualRepresentation Learning（PixPro, CVPR 2021）

​ paper: Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning (opens new window)

​ code: Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning, CVPR 2021 (opens new window)

1.3.10 AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations from Self-Trained Negative Adversaries（AdCo, CVPR 2021）

​ paper: AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations from Self-Trained Negative Adversaries (opens new window)

​ code: AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations from Self-Trained Negative Adversaries (opens new window)

Date	Training Paradigm	Method	Backbone	Label fraction	Top-1 Accuracy
2019-11	Semi Sup.	Noisy Student	EfficientNet(480M)	100+extra	88.4-
-	Sup.	-	ResNet50(24M)	100	76.5
2020-06	Self Sup	SwAV	ResNet50(24M)	0	75.3
2020-06	Self Sup	BYOL	ResNet50(24M)	0	74.3
2020-03	Self Sup	Moco v2	ResNet50(24M)	0	71.1
2020-02	Self Sup	SimCLR v1	ResNet50(24M)	0	69.3
2019-05	Self Sup	CPCv2	ResNet50(24M)	0	63.8
2019-11	Self Sup	Moco v1	ResNet50(24M)	0	60.6

3、参考文献

• 自监督学习的一些思考 (opens new window)

• 自监督学习（Self-Supervised Learning） 2018-2020年发展综述 (opens new window)

• 自监督学习(Self-Supervised Learning)综述_1 (opens new window)

• 自监督学习(Self-Supervised Learning)综述_2 (opens new window)

• Contrastive Self-Supervised Learning (opens new window)

• 「上帝视角」看对比自监督学习，SimCLR、CPC、AMDIM并无本质差异 (opens new window)

• Self-Supervised Image Classification on ImageNet (opens new window)

• Self-Supervised Image Classification on ImageNet (finetuned) (opens new window)

• Self-supervised Learning: Generative or Contrastive (opens new window)

• Self-supervised Learning: Generative or Contrastive 自监督学习2020综述 (opens new window)

• A curated list of awesome self-supervised methods (opens new window)

4、参考代码

• A python library for self-supervised learning on images. (opens new window)

• Self-Supervised Learning Toolbox and Benchmark (opens new window)