Generative Adversarial Graph Convolutional Networks for Human Action Synthesis

Synthesising the spatial and temporal dynamics of the human body skeletonremains a challenging task, not only in terms of the quality of the generatedshapes, but also of their diversity, particularly to synthesise realistic bodymovements of a specific action (action conditioning). In this paper, we proposeKinetic-GAN, a novel architecture that leverages the benefits of GenerativeAdversarial Networks and Graph Convolutional Networks to synthesise thekinetics of the human body. The proposed adversarial architecture can conditionup to 120 different actions over local and global body movements whileimproving sample quality and diversity through latent space disentanglement andstochastic variations. Our experiments were carried out in three well-knowndatasets, where Kinetic-GAN notably surpasses the state-of-the-art methods interms of distribution quality metrics while having the ability to synthesisemore than one order of magnitude regarding the number of different actions. Ourcode and models are publicly available athttps://github.com/DegardinBruno/Kinetic-GAN.