Predicting Blendshapes of Virtual Humans for Low-Delay Remote Rendering using LSTM

This paper proposes a novel framework to reduce the perceptual delay in rendering the facial expressions of virtual humans. The proposed method reduces the delay by predicting the future blend coefficients that represent facial expressions. The prediction accuracy of the proposed method is improved by 27% on average over a conventional LSTM. We also subjectively confirmed that the proposed method achieves natural facial expressions.