For certain test datasets, modern facial recognition systems are able to surpass human-level performance. However, the ability of such systems to generalise to in-the-wild images is largely limited by the availability of training data. In-the-wild images may contain subjects at large poses, or pictured under extreme lighting conditions or with extreme expressions. While images displaying these characteristics do exist in some training datasets, their distribution is typically unbalanced with the conditions in most images being fairly well controlled. The goal of this thesis is to harness the power of deep generative models (e.g. GANs) to synthesise balanced training datasets containing extreme but realistic variations. The main challenge is to introduce such variation whilst preserving the identity.

This research is being conducted in collaboration with IDEMIA, a global leader in trusted identity technologies.