Towards High-Resolution Face Pose Synthesis
Abstract: Synthesizing different views of a face image is a challenging task that can potentially help in several computer graphics and computer vision applications. In this work, we present a novel approach to address this task. We leverage the power of Generative Adversarial Networks (GANs) to synthesize face poses in a high-resolution and realistic fashion. We control the rotation of synthesized faces along the three axes of space (roll, pitch, yaw). We start by estimating the pose of each face in the training set and storing a vector containing the rotation angles. Then, we use the images along with the angles to train a conditioned version of a state-of-the-art GAN. Our experiments show image synthesis with a high-realistic finish, plus the absolute control of the pose of synthesized face images.
