Abstract

We introduce style augmentation, a new form of data augmentation based onrandom style transfer, for improving the robustness of convolutional neuralnetworks (CNN) over both classification and regression based tasks. Duringtraining, our style augmentation randomizes texture, contrast and color, whilepreserving shape and semantic content. This is accomplished by adapting anarbitrary style transfer network to perform style randomization, by samplinginput style embeddings from a multivariate normal distribution instead ofinferring them from a style image. In addition to standard classificationexperiments, we investigate the effect of style augmentation (and dataaugmentation generally) on domain transfer tasks. We find that dataaugmentation significantly improves robustness to domain shift, and can be usedas a simple, domain agnostic alternative to domain adaptation. Comparing styleaugmentation against a mix of seven traditional augmentation techniques, wefind that it can be readily combined with them to improve network performance.We validate the efficacy of our technique with domain transfer experiments inclassification and monocular depth estimation, illustrating consistentimprovements in generalization.

 

Introduction (beta)

None

 

Conclusion (beta)

None