A curated list and survey of awesome Vision Transformers.

English | 简体中文

A curated list and survey of awesome Vision Transformers.

You can use mind mapping software to open the mind mapping source file. You can also download the mind mapping HD pictures if you just want to browse them.

Contents

• Survey

  • Image Classification
    • Attention-based
      • Training Strategy
      • Model Improvements
        • Tokenization Module
        • Position Encoding Module
        • Attention Module
        • FFN Module
        • Normalization Module Location
        • Classification Prediction Head Module
        • Others
    • MLP-based
    • ConvMixer-based
    • General Architecture Analysis
    • Others
  • Object Detection
  • Semantic-Segmentation

• Papers

  • Transformer Original Paper
  • ViT Original Paper
  • Image Classification
    • 2020
    • 2021
    • 2022
  • Object Detection
  • Semantic-Segmentation

Survey

Only typical algorithms are listed in each category.

Image Classification

Chinese Blogs

• Vision Transformer 必读系列之图像分类综述(一)：概述

Attention-based

Training Strategy

• [DeiT] Training data-efficient image transformers & distillation through attention (ICML 2021-2020.12) [Paper]
• [Token Labeling] All Tokens Matter: Token Labeling for Training Better Vision Transformers (2021.4) [Paper]

Model Improvements

Tokenization Module

Image to Token：

• Non-overlapping Patch Embedding

  • [ViT] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021-2020.10) [Paper]
  • [TNT] Transformer in Transformer (NeurIPS 2021-2021.3) [Paper]
  • [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]

• Overlapping Patch Embedding

  • [T2T-ViT] Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet (2021.1) [Paper]

  • [ResT] ResT: An Efficient Transformer for Visual Recognition (2021.5) [Paper]

  • [PVTv2] PVTv2: Improved Baselines with Pyramid Vision Transformer (2021.6) [Paper]

  • [ViTAE] ViTAE: Vision Transformer Advanced by Exploring Intrinsic Inductive Bias (2021.6) [Paper]

  • [PS-ViT] Vision Transformer with Progressive Sampling (2021.8) [Paper]

Token to Token：

• Fixed sampling window tokenization
  • [ViT] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021-2020.10) [Paper]
  • [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]
• Dynamic sampling tokenization
  • [PS-ViT] Vision Transformer with Progressive Sampling (2021.8) [Paper]
  • [TokenLearner] TokenLearner: What Can 8 Learned Tokens Do for Images and Videos? (2021.6) [Paper]

Position Encoding Module

Explicit position encoding：

• Absolute position encoding
  • [Transformer] Attention is All You Need] (NIPS 2017-2017.06) [Paper]
  • [ViT] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021-2020.10) [Paper]
  • [PVT] Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions (2021.2) [Paper]
• Relative position encoding
  • [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]
  • [Swin Transformer V2] Swin Transformer V2: Scaling Up Capacity and Resolution (2021.11) [Paper]
  • [Imporved MViT] Improved Multiscale Vision Transformers for Classification and Detection (2021.12) [Paper]

Implicit position encoding：

• [CPVT] Conditional Positional Encodings for Vision Transformers (2021.2) [Paper]
• [CSWin Transformer] CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows (2021.07) [Paper]
• [PVTv2] PVTv2: Improved Baselines with Pyramid Vision Transformer (2021.6) [Paper]
• [ResT] ResT: An Efficient Transformer for Visual Recognition (2021.5) [Paper]

Attention Module

Include only global attention：

• Multi-Head attention module

  • [ViT] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021-2020.10) [Paper]

• Reduce global attention computation

  • [PVT] Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions (2021.2) [Paper]

  • [PVTv2] PVTv2: Improved Baselines with Pyramid Vision Transformer (2021.6) [Paper]

  • [Twins] Twins: Revisiting the Design of Spatial Attention in Vision Transformers (2021.4) [Paper]

  • [P2T] P2T: Pyramid Pooling Transformer for Scene Understanding (2021.6) [Paper]

  • [ResT] ResT: An Efficient Transformer for Visual Recognition (2021.5) [Paper]

  • [MViT] Multiscale Vision Transformers (2021.4) [Paper]

  • [Imporved MViT] Improved Multiscale Vision Transformers for Classification and Detection (2021.12) [Paper]

• Generalized linear attention

  • [T2T-ViT] Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet (2021.1) [Paper]

Introduce extra local attention：

• Local window mode

  • [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]
  • [Swin Transformer V2] Swin Transformer V2: Scaling Up Capacity and Resolution (2021.11) [Paper]
  • [Imporved MViT] Improved Multiscale Vision Transformers for Classification and Detection (2021.12) [Paper]
  • [Twins] Twins: Revisiting the Design of Spatial Attention in Vision Transformers (2021.4) [Paper]
  • [GG-Transformer] Glance-and-Gaze Vision Transformer (2021.6) [Paper]
  • [Shuffle Transformer] Shuffle Transformer: Rethinking Spatial Shuffle for Vision Transformer (2021.6) [Paper]
  • [MSG-Transformer] MSG-Transformer: Exchanging Local Spatial Information by Manipulating Messenger Tokens (2021.5) [Paper]
  • [CSWin Transformer] CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows (2021.07) [Paper]

• Introduce convolutional local inductive bias

  • [ViTAE] ViTAE: Vision Transformer Advanced by Exploring Intrinsic Inductive Bias (2021.6) [Paper]
  • [ELSA] ELSA: Enhanced Local Self-Attention for Vision Transformer (2021.12) [Paper]

• Sparse attention

  • [Sparse Transformer] Sparse Transformer: Concentrated Attention Through Explicit Selection [Paper]

FFN Module

Improve performance with Conv's local information extraction capability：

• [LocalViT] LocalViT: Bringing Locality to Vision Transformers (2021.4) [Paper]
• [CeiT] Incorporating Convolution Designs into Visual Transformers (2021.3) [Paper]

Normalization Module Location

• Pre Normalization

  • [ViT] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021-2020.10) [Paper]
  • [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]

• Post Normalization

  • [Swin Transformer V2] Swin Transformer V2: Scaling Up Capacity and Resolution (2021.11) [Paper]

Classification Prediction Head Module

• Class Tokens

  • [ViT] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021-2020.10) [Paper]
  • [CeiT] Incorporating Convolution Designs into Visual Transformers (2021.3) [Paper]

• Avgerage Pooling

  • [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]
  • [CPVT] Conditional Positional Encodings for Vision Transformers (2021.2) [Paper]
  • [ResT] ResT: An Efficient Transformer for Visual Recognition (2021.5) [Paper]

Others

(1) How to output multi-scale feature map

• Patch merging

  • [PVT] Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions (2021.2) [Paper]
  • [Twins] Twins: Revisiting the Design of Spatial Attention in Vision Transformers (2021.4) [Paper]
  • [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]
  • [ResT] ResT: An Efficient Transformer for Visual Recognition (2021.5) [Paper]
  • [CSWin Transformer] CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows (2021.07) [Paper]
  • [MetaFormer] MetaFormer is Actually What You Need for Vision (2021.11) [Paper]

• Pooling attention

  • [MViT] Multiscale Vision Transformers (2021.4) [Paper][Imporved MViT]

  • [Imporved MViT] Improved Multiscale Vision Transformers for Classification and Detection (2021.12) [Paper]

• Dilation convolution

  • [ViTAE] ViTAE: Vision Transformer Advanced by Exploring Intrinsic Inductive Bias (2021.6) [Paper]

(2) How to train a deeper Transformer

• [Cait] Going deeper with Image Transformers (2021.3) [Paper]
• [DeepViT] DeepViT: Towards Deeper Vision Transformer (2021.3) [Paper]

MLP-based

• [MLP-Mixer] MLP-Mixer: An all-MLP Architecture for Vision (2021.5) [Paper]

• [ResMLP] ResMLP: Feedforward networks for image classification with data-efficient training (CVPR2021-2021.5) [Paper]

• [gMLP] Pay Attention to MLPs (2021.5) [Paper]

• [CycleMLP] CycleMLP: A MLP-like Architecture for Dense Prediction (2021.7) [Paper]

ConvMixer-based

• [ConvMixer] Patches Are All You Need [Paper]

General Architecture Analysis

• Demystifying Local Vision Transformer: Sparse Connectivity, Weight Sharing, and Dynamic Weight (2021.6) [Paper]
• A Battle of Network Structures: An Empirical Study of CNN, Transformer, and MLP (2021.8) [Paper]
• [MetaFormer] MetaFormer is Actually What You Need for Vision (2021.11) [Paper]
• [ConvNeXt] A ConvNet for the 2020s (2022.01) [Paper]

Others

Object Detection

Semantic Segmentation

⬆ back to top

Papers

Transformer Original Paper

• [Transformer] Attention is All You Need] (NIPS 2017-2017.06) [Paper]

ViT Original Paper

• [ViT] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021-2020.10) [Paper]

Image Classification

2020

• [DeiT] Training data-efficient image transformers & distillation through attention (ICML 2021-2020.12) [Paper]
• [Sparse Transformer] Sparse Transformer: Concentrated Attention Through Explicit Selection [Paper]

2021

• [T2T-ViT] Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet (2021.1) [Paper]

• [PVT] Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions (2021.2) [Paper]

• [CPVT] Conditional Positional Encodings for Vision Transformers (2021.2) [Paper]

• [TNT] Transformer in Transformer (NeurIPS 2021-2021.3) [Paper]

• [Cait] Going deeper with Image Transformers (2021.3) [Paper]

• [DeepViT] DeepViT: Towards Deeper Vision Transformer (2021.3) [Paper]

• [Swin Transformer] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (ICCV2021-2021.3) [Paper]

• [CeiT] Incorporating Convolution Designs into Visual Transformers (2021.3) [Paper]

• [LocalViT] LocalViT: Bringing Locality to Vision Transformers (2021.4) [Paper]

• [MViT] Multiscale Vision Transformers (2021.4) [Paper]

• [Twins] Twins: Revisiting the Design of Spatial Attention in Vision Transformers (2021.4) [Paper]

• [Token Labeling] All Tokens Matter: Token Labeling for Training Better Vision Transformers (2021.4) [Paper]

• [ResT] ResT: An Efficient Transformer for Visual Recognition (2021.5) [Paper]

• [MLP-Mixer] MLP-Mixer: An all-MLP Architecture for Vision (2021.5) [Paper]

• [ResMLP] ResMLP: Feedforward networks for image classification with data-efficient training (CVPR2021-2021.5) [Paper]

• [gMLP] Pay Attention to MLPs (2021.5) [Paper]

• [MSG-Transformer] MSG-Transformer: Exchanging Local Spatial Information by Manipulating Messenger Tokens (2021.5) [Paper]

• [PVTv2] PVTv2: Improved Baselines with Pyramid Vision Transformer (2021.6) [Paper]

• [TokenLearner] TokenLearner: What Can 8 Learned Tokens Do for Images and Videos? (2021.6) [Paper]

• Demystifying Local Vision Transformer: Sparse Connectivity, Weight Sharing, and Dynamic Weight (2021.6) [Paper]

• [P2T] P2T: Pyramid Pooling Transformer for Scene Understanding (2021.6) [Paper]

• [GG-Transformer] Glance-and-Gaze Vision Transformer (2021.6) [Paper]

• [Shuffle Transformer] Shuffle Transformer: Rethinking Spatial Shuffle for Vision Transformer (2021.6) [Paper]

• [ViTAE] ViTAE: Vision Transformer Advanced by Exploring Intrinsic Inductive Bias (2021.6) [Paper]

• [CycleMLP] CycleMLP: A MLP-like Architecture for Dense Prediction (2021.7) [Paper]

• [CSWin Transformer] CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows (2021.07) [Paper]

• [PS-ViT] Vision Transformer with Progressive Sampling (2021.8) [Paper]

• A Battle of Network Structures: An Empirical Study of CNN, Transformer, and MLP (2021.8) [Paper]

• [Swin Transformer V2] Swin Transformer V2: Scaling Up Capacity and Resolution (2021.11) [Paper]

• [MetaFormer] MetaFormer is Actually What You Need for Vision (2021.11) [Paper]

• [Imporved MViT] Improved Multiscale Vision Transformers for Classification and Detection (2021.12) [Paper]

• [ELSA] ELSA: Enhanced Local Self-Attention for Vision Transformer (2021.12) [Paper]

• [ConvMixer] Patches Are All You Need [Paper]

2022

• [ConvNeXt] A ConvNet for the 2020s (2022.01) [Paper]

Object Detection

Semantic Segmentation

⬆ back to top

Stay tuned and PRs are welcomed!