Synthetic data augmentation for facial re-identification

Facial Re-identification datasets which facilitate the training of Deep Neural Networks (DNNs), tend to be high quality images of celebrities harvested from the internet. There is however a domain gap between these datasets, and the low quality samples used in real-world systems and scenarios such as surveillance footage. In this work we describe a novel process of data augmentation using synthetically generated images, which aids cross-domain generalisability, without the need to acquire large amounts of real data in the target domain. We also contribute a new dataset derived from this process: syn-Face. Our approach is validated by training with standard high quality datasets with synthetic augmentation and testing in 2 different realistic sets.