Scalable Human Avatar Creation

Human avatar creation is required in many applications. Commercial solutions have made tremendous progress in recent years, such as MetaHuman from Unreal. Yet, creating these avatars still needs significant artist efforts. Recent advances in deep learning, mainly neural implicit representations, have enabled controllable human avatar generation from different sensory inputs. However, to generate realistic cloth deformations from novel input poses, watertight meshes or dense full-body scans are usually needed as inputs. Our group has focused on advancing the frontier of realistic human avatar creation from affordable commodity sensors. Our work MetaAvatar is the first approach that created generalizable and animatable neural signed distance fields (SDFs) representing clothed human avatars from monocular RGB sequences. We achieved this by using meta-learning to learn an initialization of a hypernetwork that predicts the parameters of neural SDFs. We demonstrated that our meta-learned clothed human body models are very robust, being the first to generate avatars with realistic dynamic cloth deformations given as few as eight monocular depth frames.