We present a deformation correction technique called deep body deformation (DBD) to improve realism in skeletal animated human characters. We observe despite heterogeneity of shape across the human body, deformations on specific body parts exhibit high coherence due to physiological constraints. We therefore design an autoencoder net-work to encode shape variations and learn deformations unique to each semantically meaningful body part. To enforce deformation consistency over the entire animation sequence, we further develop a temporal en-coding scheme based on LSTM. Compared with the state-of-the-art such as LBS and SMPL, shape deformations produced by our technique are more accurate and appear more natural and preserve rich details.
Get the paper Here
See our demo Here