The latest research progress of domain generalization from CVPR 2022

Domain Adaptation（DA： Domain adaption ）,Domain Generalization（DG： Domain generalization ） It has always been a hot research direction of major summit meetings .DA Suppose we have a labeled training set （ Source domain ）, At this time, we want the model to perform well on another data set （ Target domain ）, Use the unlabeled data of the target domain , Improving the adaptability of the model between domains is DA What is emphasized . On this basis ,DG Further weakening the hypothesis , We only have data from multiple source domains , I have no idea what the target domain is , How to improve model generalization at this time ？ The core is how to use the rich information brought by multiple source domains . Four articles are selected in this paper CVPR 2022 Domain generalization related articles to study the latest progress .

 

BIRM

Paper title ：Bayesian Invariant Risk Minimization

Thesis link ：

https://openaccess.thecvf.com/content/CVPR2022/papers/Lin_Bayesian_Invariant_Risk_Minimization_CVPR_2022_paper.pdf

1.1 Motivation

Generalization under distribution shift is an open challenge of machine learning . Constant risk minimization （IRM） This problem is solved by extracting invariant features . although IRM It has a complete theoretical system , But the performance in depth model is not very good , This paper holds that this failure is mainly The depth model is easy to over fit Caused by the , It is theoretically verified that when the model is over fitted ,IRM Degenerate into traditional ERM. This article will Bayesian  inference introduce IRM Put forward Bayesian Invariant Risk Min-imization（BIRM） To alleviate this problem to a certain extent and achieved good results .

1.2 Background

Here is a brief introduction OOD Problems and IRM, So-called OOD The question can be written in the following form

That is to find the best classifier ,encoder Parameters    Make the model have better performance in the worst domain ,i.e., . there    Is domain    Negative log likelihood of data in ：

Invariant  Risk  Minimization（IRM）. IRM To solve the following problems ：

That is, he wants to learn a encoder Parameters  , This encoder For all classifier parameters    Both are optimal at the same time . To achieve this goal ,encoder Need to get rid of spurious feature. But this optimization form bi-level Of , Very difficult to solve , So he put forward an approximate target.

DARLING

Paper title ：Towards Unsupervised Domain Generalization

Thesis link ：

https://arxiv.org/abs/2107.06219

2.1 Motivation

current DG Most algorithms are pre trained on large data sets backbone, Then the algorithm is developed on the downstream data set finetune, Last in unseen Test on the data set of . But the pre - training data set introduces the idea that bias, such as imagenet Most of the images come from the real world , Therefore, its pre trained model is not real data on downstream tasks （ Hand drawn image , Cartoon image ） Performance will be poor . This paper considers a new setting, The model First, perform unsupervised pre training on multiple source domains , And then traditional DG The process of , Erase the tradition pretrain Introduced bias.

2.2 Method

The method name in this article is Domain-Aware  Representation  Learn-ING（DARLING）, This method tries to use self supervised learning technology to improve the multi domain pre training method, so as to improve the generalization performance . The traditional self supervised learning loss is ：

there    It is in self-monitoring anchor and positive features ,  and    It's two separate encoder,  It's a temperature parameter . But this method can not build the confidence of the model domain , because    It is actually different in different domains . Every domain    The conditional distribution of can be written ：

The model architecture is as follows , there    It's below similarity predictor Output . That is to say, it has become each domain Compare the weighted form of learning loss ：

 

PCL

Paper title ：PCL: Proxy-based Contrastive Learning for Domain Generalization

Thesis link ：

https://openaccess.thecvf.com/content/CVPR2022/papers/Yao_PCL_Proxy-Based_Contrastive_Learning_for_Domain_Generalization_CVPR_2022_paper.pdf

3.1 Motivation

This article is also related to contrastive learning A related article , A simple method is to pull the positive sample pairs of different fields closer , At the same time, push other negative sample pairs further . This paper finds that It is not good to directly adopt this kind of supervised comparative learning , This paper considers that there is a significant distribution gap between domains , It makes the distance between positive sample pairs close directly, but hinders the generalization of the model . Therefore, this paper proposes a new prototype based approach （proxy） Comparative learning method .

3.2 Method: Proxy-based Contrastive Learning

Let's look at it first , What is the difference between prototype based self supervised learning and traditional self supervised learning . As shown in the figure below ：

PACS Dataset is a typical domain generalization benchmark , It includes four areas ： art 、 comic 、 Photos and sketches , There are seven categories in each field .DG From multiple source domains （ Like art 、 Photo 、 Sketch ） Training models , And in the target area （ Like comics ） Test on . In the training phase , The target data set cannot be accessed .

Traditional comparison based losses （ for example , Monitor and compare losses ） Using its sample to sample relationship , Among them, samples from different domains of the same class can be regarded as positive . We think , Optimizing some difficult sample pairs may deteriorate the generalization ability of the model . This article calls it Positive alignment problem （ positive alignment problem）. Because between the domains gap It could be very big , Therefore, directly aligning positive samples from different domains may be harmful to the model . This paper proposes a prototype based contrast loss to solve this problem . The specific differences between the two are as follows ：

 

Style Neophile

Paper title ：Style Neophile: Constantly Seeking Novel Styles for Domain Generalization

Thesis link ：

https://openaccess.thecvf.com/content/CVPR2022/papers/Kang_Style_Neophile_Constantly_Seeking_Novel_Styles_for_Domain_Generalization_CVPR_2022_paper.pdf

4.1 Motivation

At present, most DG Methods all mention learning one domain-invariant Characteristics of , Such methods usually assume that each domain has a different image style , then enforce The characteristics of images of the same class under different styles are as similar as possible . and , These methods are limited to a limited set of styles （ The number of fields we get is finite ）, Because they get enhanced patterns from a set of fixed training images or by interpolating training data . This paper presents a new method , Can generate more styles of data , These styles haven't even appeared in the training set .

4.2 Method

This paper describes the so-called image style by using the mean and variance of the image , For images feature map , We have  .

The overall framework of this paper is as follows , Let's introduce each part in turn .

summary

Four articles are selected in this paper CVPR And DG Related articles , They are from cause and effect invariance , Pre training mode , Self supervised learning +DG And data enhancement , Overcome the shortcomings of the past . But for now , There is still no milestone approach that can be applied in the vast majority of OOD Of benchmark Show overwhelming effect on , Better, faster and stronger methods are still lacking at present .