IMAGEimate - An End-to-End Pipeline to Create Realistic Animatable 3D Avatars from a Single Image Using Neural Networks

Abstract

Current advances in image based 3D human shape estimation and parametric human models enable creating realistic 3D virtual humans. We present a pipeline which takes advantage of these models and takes a single input image to create realistic 3D animatable avatars. The pipeline extracts shape and pose parameters from the input image and builds an implicit surface representation, which is then fitted onto a parametric human model. This fitted human model is animated to new and novel poses extracting pose parameters from a motion capture dataset. We extend the pipeline showcasing realism and interaction by texture painting it using Substance Painter and embedding it in an AR scene using Adobe Aero respectively.