Face Synthesis and Recognition Using Disentangled Representation-Learning Wasserstein GAN

We propose the Disentangled Representation-learning Wasserstein GAN (DR-WGAN) trained on augmented data for face recognition and face synthesis across pose. We improve the state-of-the-art DR-GAN with the Wasserstein loss considered in the discriminator so that the generative and adversarial framework can be better trained. The improved training leads to better face disentanglement and synthesis. We also highlight the influences of imbalanced training data on the disentangled facial representation learning, and point out the difficulty of generating faces of extreme poses. We explore the recently proposed nonlinear 3D Morphable Model (3DMM) to augment the training data, and verify the contributions made by the learning on augmented data. Additionally, we also compare different data normalization schemes and reveal the benefit of using the group normalization. The proposed framework is verified through the experiments on benchmark databases, and compared with contemporary approaches for performance evaluation.