Progressive Random Convolutions for Single Domain Generalization

Single domain generalization aims to train a generalizable model with onlyone source domain to perform well on arbitrary unseen target domains. Imageaugmentation based on Random Convolutions (RandConv), consisting of oneconvolution layer randomly initialized for each mini-batch, enables the modelto learn generalizable visual representations by distorting local texturesdespite its simple and lightweight structure. However, RandConv has structurallimitations in that the generated image easily loses semantics as the kernelsize increases, and lacks the inherent diversity of a single convolutionoperation. To solve the problem, we propose a Progressive Random Convolution(Pro-RandConv) method that recursively stacks random convolution layers with asmall kernel size instead of increasing the kernel size. This progressiveapproach can not only mitigate semantic distortions by reducing the influenceof pixels away from the center in the theoretical receptive field, but alsocreate more effective virtual domains by gradually increasing the stylediversity. In addition, we develop a basic random convolution layer into arandom convolution block including deformable offsets and affine transformationto support texture and contrast diversification, both of which are alsorandomly initialized. Without complex generators or adversarial learning, wedemonstrate that our simple yet effective augmentation strategy outperformsstate-of-the-art methods on single domain generalization benchmarks.