APA: domain generalization using frequency based augmentation

Domain shift remains a major challenge in computer vision, where models trained on source domain(s) perform poorly on unseen target domains. One way to address this issue is by using data augmentation techniques, which generate synthetic images of the source images to mimic potential characteristics of the target domain. In this work, we introduce Amplitude-Phase Augmentation (APA), a frequency-domain augmentation technique. APA generates cross-domain images by multiplying the amplitude spectra of image pairs while preserving their original phase information. This approach exposes models to a broader spectrum of frequency-based variations, simulating diverse domain styles. We demonstrate that on two different datasets and with two different backbones, APA outperforms the baseline model, which is trained with original images, and also achieves competitive results compared to other models trained on the same backbones and datasets. Code available at https://github.com/sina-nuel/APA.