V-DMGNN-GAN

Variational seq2seq DMGNN-based GAN (V-DMGNN-GAN) model for human skeletal motion prediction.

Abstract: For the task of modelling and predicting human motion trajectories, existing generative models suffer from occlusions and mislabeling that are commonplace in human motion datasets. We tackle this problem by building a network that can accurately infer trajectories of skeletal joints missing from input, by building upon previous work on deterministic multi-scale motion prediction. We build a GAN-based architecture consisting of a variational auto-encoder (VAE) as our generator network to learn the distribution of human motion and a recurrent neural network-based discriminator for explicitly penalizing the unrealistic motion that is generated from our generator.