论文简记| ICLR2021 under review《Domain Generalization with Mixstyle》

1 写在前面

未经允许,不得转载,谢谢~

很久不写blog,有些不习惯了。也不知道是真的因为忙的没有时间写了,还是自己懒起来了。可能赶DDL的时候真的是前者,但是大部分时候又是后者。

希望今年能够稍微多写一些。但是可能不会每篇内容都写的那么仔细了,就挑重要的记录一些。希望大家见谅,以及感谢还在看的小伙伴们。

2 introduction

  1. the importance of representation learning
  2. when the training and the test data are sampled from different distributions, severe performance degradation is expected.
  3. domain generalization (DG): multiple source domains containing the same visual classes are available for model training, the goal of DG is to enable the trained model to generalize well to any unseen domains.
  4. the difference between DG (domain generalization) and DA (domain adaptation): DA assumes some labeled or unlabeled target images are available during training and one single source domain is given generally.
  5. A straightforward solution: expose a model with a large variety of source domains. However, it is costly and even impossible.
  6. This paper: a novel MixStyle model based on probabilistically mixing instance-level feature statistics of training samples across source domains. Inspired by the style transfer, that the bottom layers will maintain the style information. Thus, replacing the style while preserving the semantic content will result in images of mixed new styles.
    7.easy to implement.

3 方法

主要是两种方法(style transfer和mixup)的一种集合。但又不是简单的使用了这两种方法,而是进行了巧妙的结合。

3.1 style transfer

IN vs AdaIN

3.2 mixup

原始的mixup对两张图片x1和x2以\lambda的比例融合,并对他们的类别标签也进行融合。
x = \lambda x_1 + (1-\lambda) x_2
y = \lambda y_1 + (1-\lambda) y_2

3.3 mixstyle

mixstyle混合的是两个图像的style信息,给定一个包含来自两个源数据集的图像input batch x,首先获取到reference batch \tilde{x} 。获取方式有两种,(a)batch中图像所属的domain已知:先交换不同domain数据,再各自shuffle,以保证被mix的数据对来自不同的source domain(b) batch中图像所属的domain不可知:随机。

最后的mixstyle:


4 实验

实验部分主要验证了两件事情:

  1. 比其他的DG方法好;
  2. 比mixup系列的方法好;

实验部分做的还是很充分的,验证了各个部分的作用。

推荐阅读更多精彩内容