3DCMM: 3D Comprehensive Morphable Models for Accurate Head Completion

Jie Zhang, Yan Luximon, Lei Zhu, Ping Li

Abstract: 3D head completion aims at recovering accurate 3D full-head geometry from 2D face images or 3D face scans. Previous 3D shape reconstruction studies primarily focused on the facial region, but ignored the scalp region. Moreover, as critical foundations in 3D head completion, powerful 3D head morphable models, however, are scarce. In this paper, we construct 3D comprehensive morphable models (3DCMM) of human faces and scalps, and develop a novel 3DCMM-based stepwise 3D full-head creation pipeline: reconstructing face regions firstly, and then completing scalp regions. Firstly, large-scale 3D heads from 2,528 identities were parameterized to construct powerful 3DCMM as our foundations. Then, a 3DCMM-based supervised converting method was presented to predict an accurate scalp region from a facial region and produce full-head geometry. Extensive experiments and comparisons demonstrated that our 3DCMM possesses better quality and descriptive power. Benefiting from this, our model-based 3D head completion method has higher accuracy than model-based fitting method.