Face Frontalization with Inpainting Method

Kaiyu Guo,Juyong Zhang
DOI: https://doi.org/10.1145/3373509.3373572
2019-01-01
Abstract:Problems caused by large profile views are always holdback for face recognition and face reconstruction, such as texture missing problem. In this paper, we propose a method to generate frontal faces from faces of different poses in the wild. Different from previous works for face frontalization which only use GANs or only render a rotated 3D face model to an image, we combine face reconstruction and generative adversarial model, which can preserve the shape information from 3D space and refine the texture through learned distribution from GAN. The task of our GAN is to inpaint the incomplete frontal images which are rendered from the reconstructed 3D model. As the results show, our inpainting method has a better performance than the Poisson editing method on face verification.
What problem does this paper attempt to address?