The dense shape and pixel correspondence enable us to adopt a series of self-supervised constraints to fit the 3D face model accurately and learn the per-vertex correctives end-to-end. 3D3M builds ...