Data augmentation with Mobius transformations

Abstract Data augmentation has led to substantial improvements in the performance and generalization of deep models, remains a highly adaptable method evolving model architectures varying amounts data—in particular, extremely scarce available training data. In this paper, we present novel applying Möbius transformations augment input images during training. are bijective conformal maps that generalize image translation operate over complex inversion pixel space. As result, can on sample level preserve data labels. We show inclusion enables improved prior sample-level techniques such as cutout standard crop-and-flip transformations, most notably low regimes.