Text-Based Motion Synthesis with a Hierarchical Two-Stream RNN

We present a learning-based method for generating animated 3D pose sequences depicting multiple sequential or superimposed actions provided in long, compositional sentences. We propose a hierarchical two-stream sequential model to explore a finer joint-level mapping between natural language sentences and the corresponding 3D pose sequences of the motions. We learn two manifold representations of the motion –- one each for the upper body and the lower body movements. We evaluate our proposed model on the publicly available KIT Motion-Language Dataset containing 3D pose data with human-annotated sentences. Experimental results show that our model advances the state-of-the-art on text-based motion synthesis in objective evaluations by a margin of 50%.