Our motivation is straightforward: since the vanilla classifier trained with SGD will inadvertently focus more on those domain-specific features, the weights of the trained classifier can be considered as an indicator of those features.(作者的想法:既然classifier不可避免地会关注domain-specific feature,那么关注了这些bias的classifier就可以作为bias的一种indicator。此外,单个classifier无法定位所有的bias feature,因此提出使用多个分类器共同组成一个专门学习bias的分支。)
For images, there are several factors typically correlated to domain-specific features, such as the background contexts [1], the texture of the objects [19], and high-frequency patterns that are almost invisible to the human eye [55].(图像中domain-specific特征通常有多个因素,包括背景,纹理和人眼不可见的高频信号,因此作者认为只用一个classifier来指示bias是不全面的)
Based on the proposed biased-branch, we have an indicator to those domain-specific features. A follow-up question is how to suppress those domain-specific features in favor of focusing more on those desired domain-generalized features.(多头分类器学习完毕后,现在相当于有了对 domain-specific feautre 的指示器,接下来就是去除这些 bias。)
However, if we optimize the whole network (including the feature extractor f, biased-branch classifier $g{bias}$, and general-branch classifier $g{gen}$ simultaneously at the beginning, there is no guarantee that the classifier $g{gen}$ will pay more attention to those domain-general features. To address this issue, we introduce a two-stage learning mechanism to enable the interaction between the two branches.(若直接对3个损失函数进行联合优化,无法保证 $g{gen}$ 能更多地关注域泛化特征,因此作者提出两阶段训练策略)
As pointed out in [55], the low-frequency component (LFC) is much more generalizable than high-frequency component (HFC), i.e., LFC typically represents those domain-generalized (semantic) features, and HFC denotes those domain-specific (superficial) features.(低频通常对应域泛化特征,高频通常对应域特定特征)
MAD (CVPR 2023)
Modality-Agnostic Debiasing for Single Domain Generalization, CVPR 2023. [paper]
Institution & Author
Preliminaries
Domain Adaptation (DA)
Domain Generalization (DG)
Single Domain Generalization (SDG)
Motivation
目前做SDG的方法大多是设计各种数据增强方法,将单一域的数据增强到不同域,以提高模型泛化性。但是,作者认为这些方法都是 modality-specific 的,并且大多只适用于图像领域。也就是说,这些方法关注的是数据增强方法,将单一图像域增强到多种图像域。但是若输入的是点云数据,点云数据具有和图像不同的domain shift规则(图像的domain shift一般是纹理、结构,而点云数据一般是几何结构和位置),那么这些数据增强方法是不适用于点云数据的。因此,作者从网络结构设计的角度来解决SDG问题,实现和数据模态无关的SDG方法。
Contribution
Method
Module
Pipeline
疑问:
Experiments