Conditional MoCoGAN for Zero-Shot Video Generation

Abstract

We propose a conditional generative adversarial network (GAN) model for zero-shot video generation. In this paper we deal with a zero-shot conditional generation setting, i.e., generating conditional unseen videos from training samples with missing classes. The task is an extension of conditional data generation. The key idea is to learn disentangled representations in the latent space of GANs. To do this, our model is based on the motion and content decomposed GAN (MoCoGAN) and the auxiliary classifier GAN (ACGAN). Using the two GANs, we build three models to find better disentangled representations and to generate good quality videos. In experiments with colored MNIST and the Weizmann action database, we show the effectiveness of our proposed models.