A Benchmark and a Data-Driven Solution for Facial Relighting
Facial relighting is a problem that requires very high quality results due to its use in post-production as well as the high sensitivity of humans to artifacts and deformations on generated facial imagery. This problem has been examined in the literature for more than two decades. However, the quantitative evaluation of the state-of-the-art methods has been very limited due to the lack of a standard dataset. In this paper, we publicly release a large dataset covering many different lighting conditions, head poses and hundreds of subjects for the problem of facial relighting. We quantitatively and qualitatively evaluate the state-of-the art techniques and discuss their shortcomings. We also propose a generative network that takes a facial image as input and produces high quality results for target lighting configurations. Our architecture includes specific design decisions to ensure high-quality facial texture and realistic shadows in the results.