Generating full-view face images from a single-view image

Lei Zhong, Changmin Bai, Jianfeng Li

Abstract: In the fields of face identification and virtual reality, it is important to generate a face image at arbitrary views from a single-view face image. In this work, a generative adversarial network is used to design an end-to-end network that can generate full-view face images with only an arbitrary single-view image input. We aim to tackle the problem of the existing databases containing the ground-truth only in specific views, with no data available for network training in arbitrary views. We propose to share the generator weights for specific views and arbitrary views during training and use optical flow interpolation between two adjacent specific views to constrain the consistency of facial structure in continuous views. Finally, it is proved using a public database that the proposed method is significantly better than the existing network for the generation of the full-view face images, particularly under the continuous rotation view.