Learning Animable 3D Face Model from Natural Scene Images

Subjects: Information Science and Systems Science >> Basic Disciplines of Information Science and Systems Science submitted time 2024-01-06

Cuilian Jiang

Abstract： Although the current 3D face reconstruction methods based on a single image can recover fine geometric details, these methods have limitations. The faces generated by some methods can't be really animated because they don't model how wrinkles change with expressions. Other methods are trained on high-quality facial scanning, and cannot be well extended to images of natural scenes. The method used in the report can return to the details of three-dimensional face shapes and animations, which are specific to individuals but can change with expressions. The model of this method can be trained to generate a UV displacement map from a low-dimensional potential representation composed of person-specific detail parameters and general expression parameters, while the regression quantity can be trained to predict details, shapes, expressions, postures and lighting parameters from a single image. In order to achieve this, this method introduces a new loss of detail consistency, which separates people-specific details from wrinkles that depend on expressions. This unwrapping makes it possible to synthesize realistic personal specific wrinkles by controlling expression parameters while keeping personal specific details unchanged. This method is learned from images of natural scenes, and there is no paired 3D data supervision.
 

Peer Review Status:Awaiting Review

 Hits 791  Downloads 237  Comment 0