HUMAN MOTION GENERATION BY STOCHASTIC CONDITIONING OF DEEP RECURRENT NETWORKS ON POSE MANIFOLDS

Human motion generation is a stochastic process. The 3D motion generation task requires efficient regulation of stochasticity and a controlled approach for error-accumulation. Current generation approaches either fail to check error-amplitude or to preserve the signal. In this paper, we present a stochastic approach for 3D human motion generation. To this end, we design a fully differentiable, end-to-end, block-based autoregressive recurrent neural network (RNN) architecture. The proposed model incorporates variable auto-conditioning length along with probabilistic variational inference on the RNN hidden-state, to regulate stochasticity. We separately train an auto-encoder to bound skeletons on a known manifold of valid-poses. We extensively test the proposed approach on publicly available Motion-Capture benchmarks. The quantitative and qualitative evaluations indicate the superiority of the proposed approach in comparison to state-of-the-art on long-term motion generation while achieving comparable performance on short-term prediction task.