2020年，計算機視覺領域會有哪些新的研究方向值得提前探索？

相關報道：

Computer Vision After 5 Years - CVPR Workshop

五年後的計算機視覺會是什麼樣？和CV先驅們一同暢想（上） | CVPR2019?
mp.weixin.qq.com
2. 深度學習和計算機視覺領域在 2018 年還有哪些能突破 / 興起或者全新的發展方向？
深度學習和計算機視覺領域在 2018 年還有哪些能突破 / 興起或者全新的發展方向？?
www.zhihu.com
3. 深度學習領域有哪些瓶頸？
深度學習領域有哪些瓶頸？?
www.zhihu.com

比較「紅」和值得探索的方向，比如，自監督/半監督/弱監督/無監督、少樣本/單樣本/零樣本、可解釋性、對抗樣本/安全、多模態、3D等等~

個人對GAN比較感興趣，CVPR 2020上GAN的論文超110+篇之多，可見其之火。事實上，GAN也在積極向前面那些「紅」方向靠攏，畢竟在數據、靈活性、可控性/可解釋性、更廣泛的應用性上，GAN也急需改善、值得進一步探索，以做到更好。

CVPR 2020 整理到的GAN論文（117篇）涉及的方向（僅供參考，未經允許不得轉載）：

1 虛擬試衣、時尚-妝容編輯
2 姿勢轉換、人像合成
3 卡通化、上色、風格遷移
4 文本-圖像生成
5 主動學習
6 強化學習
7 自監督、半監督、無監督學習
8 少樣本、單樣本、零樣本學習
9 圖表示學習
10 域適應-知識遷移
11 遊戲生成
12 醫學圖像
13 多樣性生成
14 圖像恢復、超分去霧去雨
15 圖像修復
16 反射/反光、陰影去除與生成
17 圖像分解、分離
18 注意力機制
19 紋理生成
20 解耦表徵（可控/可解釋性等）
21 聚類
22 語義分割
23 行人重識別
24 GAN結合對抗樣本/攻擊
25 人臉方面（可控生成等等）
26 圖像轉換
27 訓練改進（網路/訓練策略等等）
28 語義生成
29 NAS
30 圖像配准
31 缺失數據處理
32 視頻生成
33 隱私保護
34 模型壓縮
35 3D等其它

詳細版：CVPR 2020最全GAN論文簡介匯總；更多GAN閱讀：生活不止有枸杞，還有深度學習GAN和遠方

清單版：

1 虛擬試衣、時尚-妝容編輯

Fashion Editing with Adversarial Parsing Learning
PSGAN: Pose and Expression Robust Spatial-Aware GAN for Customizable Makeup Transfer
Image Based Virtual Try-on Network from Unpaired Data
Towards Photo-Realistic Virtual Try-On by Adaptively Generating?Preserving Image Content

2 姿勢轉換、人像合成

MISC: Multi-condition Injection and Spatially-adaptive Compositing for Conditional Person Image Synthesis
Deep Image Spatial Transformation for Person Image Generation
Controllable Person Image Synthesis with Attribute-Decomposed GAN
Wish You Were Here: Context-Aware Human Generation

3卡通化、上色、風格遷移

Learning to Cartoonize Using White-box Cartoon Representations
PuppeteerGAN: Arbitrary Portrait Animation with Semantic-aware Appearance Transformation
Unpaired Portrait Drawing Generation via Asymmetric Cycle Mapping
Reference-Based Sketch Image Colorization using Augmented-Self Reference and Dense Semantic Correspondence
Two-Stage Peer-Regularized Feature Recombination for Arbitrary Image Style Transfer

4 文本-圖像生成

ManiGAN: Text-Guided Image Manipulation
CookGAN: Causality based Text-to-Image Synthesis
RiFeGAN: Rich Feature Generation for Text-to-Image Synthesis from Prior Knowledge

5主動學習

State-Relabeling Adversarial Active Learning

6 強化學習

RL-CycleGAN: Reinforcement Learning Aware Simulation-To-Real

7 自監督、半監督、無監督學習

Regularizing Discriminative Capability of CGANs for Semi-Supervised Generative Learning
A Self-supervised Approach for Adversarial Robustness
Transformation GAN for Unsupervised Image Synthesis and Representation Learning
Unsupervised Learning for Intrinsic Image Decomposition from a Single Image
Self-Supervised Viewpoint Learning From Image Collections
Diverse Image Generation via Self-Conditioned GANs
Self-supervised Domain-aware Generative Network for Generalized Zero-shot Learning
PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models
Adversarial Robustness: From Self-Supervised Pre-Training to Fine-Tuning
Self-supervised Domain-aware Generative Network for Generalized Zero-shot Learning

8 少樣本、單樣本、零樣本學習

Adversarial Feature Hallucination Networks for Few-Shot Learning
One-Shot Domain Adaptation For Face Generation
Semi-supervised Learning for Few-shot Image-to-Image Translation
Self-supervised Domain-aware Generative Network for Generalized Zero-shot Learning

9 圖表示學習

Distribution-induced Bidirectional Generative Adversarial Network for Graph Representation Learning
Semantic Image Manipulation Using Scene Graphs

10 域適應、知識遷移

MineGAN: effective knowledge transfer from GANs to target domains with few images
Progressive Adversarial Networks for Fine-Grained Domain Adaptation
Gradually Vanishing Bridge for Adversarial Domain Adaptation
Domain Adaptation for Image Dehazing

11 遊戲生成

Learning to Simulate Dynamic Environments with GameGAN

12 醫學圖像

Synthetic Learning: Learn From Distributed Asynchronized Discriminator GAN Without Sharing Medical Image Data
Augmenting Colonoscopy using Extended and Directional CycleGAN for Lossy Image Translation

13 多樣性生成

Diverse Image Generation via Self-Conditioned GANs
UCTGAN: Diverse Image Inpainting based on Unsupervised Cross-Space Translation
MaskGAN: Towards Diverse and Interactive Facial Image Manipulation
StarGAN v2: Diverse Image Synthesis for Multiple Domains
Semantic Pyramid for Image Generation

14 圖像恢復、超分去霧去雨

Learning Invariant Representation for Unsupervised Image Restoration
Domain Adaptation for Image Dehazing
EventSR: From Asynchronous Events to Image Reconstruction, Restoration, and Super-Resolution via End-to-End Adversarial Learning
Learning Texture Transformer Network for Image Super-Resolution
PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models
Unpaired Image Super-Resolution using Pseudo-Supervision
Contextual Residual Aggregation for Ultra High-Resolution Image Inpainting
Nested Scale-Editing for Conditional Image Synthesis

15 圖像修復

UCTGAN: Diverse Image Inpainting based on Unsupervised Cross-Space Translation
Contextual Residual Aggregation for Ultra High-Resolution Image Inpainting
Semantic Pyramid for Image Generation

16 反射/反光、陰影去除與生成

Learning to Shadow Hand-drawn Sketches
Deep Adversarial Decomposition: A Unified Framework for Separating Superimposed Images
Single Image Reflection Removal through Cascaded Refinement
Single Image Reflection Removal with Physically-Based Training Images
Reflection Scene Separation From a Single Image
ARShadowGAN: Shadow Generative Adversarial Network for Augmented Reality in Single Light Scenes
BEDSR-Net: A Deep Shadow Removal Network from a Single Document Image

17 圖像分解、分離

Unsupervised Learning for Intrinsic Image Decomposition from a Single Image
Deep Adversarial Decomposition: A Unified Framework for Separating Superimposed Images

18 注意力機制

Attentive Normalization for Conditional Image Generation
S2A: Wasserstein GAN with Spatio-Spectral Laplacian Attention for Multi-Spectral Band Synthesis
Your Local GAN: Designing Two Dimensional Local Attention Mechanisms for Generative Models
End-to-End Adversarial-Attention Network for Multi-Modal Clustering
UCTGAN: Diverse Image Inpainting based on Unsupervised Cross-Space Translation
VIBE: Video Inference for Human Body Pose and Shape Estimation
DOA-GAN: Dual-Order Attentive Generative Adversarial Network for Image Copy-move Forgery Detection and Localization
Deep Image Spatial Transformation for Person Image Generation

19 紋理生成

Adversarial Texture Optimization from RGB-D Scans

20 解耦表徵

Semantic Pyramid for Image Generation
MixNMatch: Multifactor Disentanglement and Encoding for Conditional Image Generation
Image Processing Using Multi-Code GAN Prior
Nested Scale-Editing for Conditional Image Synthesis
G3AN: Disentangling Appearance and Motion for Video Generation
Neural Head Reenactment with Latent Pose Descriptors
Guided Variational Autoencoder for Disentanglement Learning
Disentangled Image Generation Through Structured Noise Injection

21 聚類

End-to-End Adversarial-Attention Network for Multi-Modal Clustering

22 語義分割

Learning Texture Invariant Representation for Domain Adaptation of Semantic Segmentation

23 行人重識別

Cross-modality Person re-identification with Shared-Specific Feature Transfer
Transferable, Controllable, and Inconspicuous Adversarial Attacks on Person Re-identification With Deep Mis-Ranking

24 GAN結合對抗攻擊/樣本

PhysGAN: Generating Physical-World-Resilient Adversarial Examples for Autonomous Driving
Adversarial Robustness: From Self-Supervised Pre-Training to Fine-Tuning
Adversarial Camouflage: Hiding Physical-World Attacks with Natural Styles
A Self-supervised Approach for Adversarial Robustness
DaST: Data-free Substitute Training for Adversarial Attacks
LG-GAN: Label Guided Adversarial Network for Flexible Targeted Attack of Point Cloud-based Deep Networks
Transferable, Controllable, and Inconspicuous Adversarial Attacks on Person Re-identification With Deep Mis-Ranking

25 人臉方面

Intuitive, Interactive Beard and Hair Synthesis with Generative Models
Disentangled Image Generation Through Structured Noise Injection
Adversarial Latent Autoencoders
StyleRig: Rigging StyleGAN for 3D Control over Portrait Images
Deep 3D Portrait from a Single Image
Disentangled and Controllable Face Image Generation via 3D Imitative-Contrastive Learning
MaskGAN: Towards Diverse and Interactive Facial Image Manipulation
Cascade EF-GAN: Progressive Facial Expression Editing with Local Focuses
Interpreting the Latent Space of GANs for Semantic Face Editing
One-Shot Domain Adaptation For Face Generation
Cross-Spectral Face Hallucination via Disentangling Independent Factors
Copy and Paste GAN: Face Hallucination from Shaded Thumbnails
The GAN that Warped: Semantic Attribute Editing with Unpaired Data
Neural Head Reenactment with Latent Pose Descriptors
CIAGAN: Conditional Identity Anonymization Generative Adversarial Networks
Rotate-and-Render: Unsupervised Photorealistic Face Rotationfrom Single-View Images

26 圖像轉換

Fine-grained Image-to-Image Transformation towards Visual Recognition
Reusing Discriminators for Encoding: Towards Unsupervised Image-to-Image Translation
Domain Adaptive Image-to-image Translation
DUNIT: Detection-based Unsupervised Image-to-Image Translation
High-Resolution Daytime Translation Without Domain Labels
Semi-supervised Learning for Few-shot Image-to-Image Translation
Deformation-aware Unpaired Image Translation for Pose Estimation on Laboratory Animals
Cross-domain Correspondence Learning for Exemplar-based Image Translation
Unsupervised Multi-Modal Image Registration via Geometry Preserving Image-to-Image Translation
Augmenting Colonoscopy using Extended and Directional CycleGAN for Lossy Image Translation
StarGAN v2: Diverse Image Synthesis for Multiple Domains

27 訓練改進

Analyzing and Improving the Image Quality of StyleGAN
A Characteristic Function Approach to Deep Implicit Generative Modeling
Alleviation of Gradient Exploding in GANs: Fake Can Be Real
Watch your Up-Convolution: CNN Based Generative Deep Neural Networks are Failing to Reproduce Spectral Distributions
A U-Net Based Discriminator for Generative Adversarial Networks
MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks
Noise Robust Generative Adversarial Networks
On Positive-Unlabeled Classification in GAN
AdversarialNAS: Adversarial Neural Architecture Search for GANs
Adversarial Latent Autoencoders
Your Local GAN: Designing Two Dimensional Local Attention Mechanisms for Generative Models

28 語義生成

SEAN: Image Synthesis with Semantic Region-Adaptive Normalization
Local Class-Specific and Global Image-Level Generative Adversarial Networks for Semantic-Guided Scene Generation
SketchyCOCO: Image Generation from Freehand Scene Sketches
BachGAN: High-Resolution Image Synthesis from Salient Object Layout
Panoptic-based Image Synthesis
Semantic Image Manipulation Using Scene Graphs
Semantically Multi-modal Image Synthesis
Wish You Were Here: Context-Aware Human Generation

29 NAS

AdversarialNAS: Adversarial Neural Architecture Search for GANs

30 圖像配准

Unsupervised Multi-Modal Image Registration via Geometry Preserving Image-to-Image Translation

31 缺失數據處理

GAMIN: Generative Adversarial Multiple Imputation Network for Highly Missing Data

32 視頻生成

G3AN: Disentangling Appearance and Motion for Video Generation

33 隱私保護

DOA-GAN: Dual-Order Attentive Generative Adversarial Network for Image Copy-move Forgery Detection and Localization
CIAGAN: Conditional Identity Anonymization Generative Adversarial Networks

34 模型壓縮

GAN Compression: Efficient Architectures for Interactive Conditional GANs

35 3D等其它

StereoGAN: Bridging Synthetic-to-Real Domain Gap by Joint Optimization of Domain Translation and Stereo Matching
SharinGAN: Combining Synthetic and Real Data for Unsupervised Geometry Estimation
Learning Generative Models of Shape Handles
SynSin: End-to-end View Synthesis from a Single Image
VIBE: Video Inference for Human Body Pose and Shape Estimation
Towards Unsupervised Learning of Generative Models for 3D Controllable Image Synthesis
LG-GAN: Label Guided Adversarial Network for Flexible Targeted Attack of Point Cloud-based Deep Networks

更多相關閱讀：