Deep Human Digitization from Minimal Input – DeepHD

The goal of this project is to develop a new generation of deep learning based methods that can create accurate digital 3D replicas of clothed people from a single color image captured with a consumer grade camera. These replicas can take the form of an explicit colored 3D reconstruction of the subject, or an implicit function encoding their 3D shape and appearance, that can be subsequently tessellated into an explicit 3D model or rendered from novel view-points. This research has promissing applications in human 3D content generation and immersive video-conferencing. We propose original solutions to the existing methods' limitations, namely their memory and computational inefficiency, lack of generalization and heavy dependence on expensive full 3D supervision.