Face frontalization with adaptive soft symmetry

Frontalization typically represents the process of generating frontal views of faces from single unconstrained images. Previous researches have suggested that face frontalization may considerably improve the performance of current face recognition systems in the wild. While many frontalization methods focuses on 3D reconstruction, approximating facial surface for each input image could be a hard problem and lead to facial misalignment. Therefore, more straightforward and simple approaches are proposed. A single unmodified 3D surface is used to approximate the shape of all input faces, and the frontal view is obtained through two specific steps, hard frontalization and soft symmetry. However, we find that large pose variations and landmark localization error may result in extreme artifacts in frontalized faces. To improve the performance of this strategy, we firstly proposed an adaptive method to take the posture of head into consideration, and use it to adjust the soft symmetry step. Then the quality of frontalization is assessed so we are able to adaptively compensate for the landmark localization error. Experiments on the Labeled Faces in the Wild (LFW) benchmark and CAS-PEAL-R1 dataset verify the effectivity of our methods to unconstrained conditions.