Data-driven modeling of subtle eye region deformations

Eyes are extremely important in face-to-face communication due to their critical role in conveying emotions, mental state and non-verbal cues. The eye region has a wide variety of possible shapes and deformations, which play a key role in the modeling of faithful digital double. In this paper, we present DREAM (Data-dRiven Eye Area Modeling) the first data-driven eye region model that covers all the possible eye poses, allowing to recreate precise identities, gaze directions, eyelid shapes and aperture, and eye region deformations. We set up an innovative pipeline with eyelash removal, automatic landmarking and mesh registration to generate high quality training data. Comparative results show that DREAM outperforms existing models and allows to cover the modeling of a larger range of morphologies, gaze directions and eyelids apertures. We also present a new scan fitting methodology specific to eye region. We eventually show that DREAM can be used to extract a high-quality, animation-ready eye region rig from a single textured scan.