3D Face Geometry Enhancement: A Self-Supervised Framework

In the rapidly advancing realm of digitalization, 3D face modeling technologies are becoming increasingly vital across various domains. However, existing monocular 3D face reconstruction methods, particularly those based on PCAderived 3D Morphable Models(3DMMs), face challenges in accurately capturing fine-grained facial details due to their inherent smoothness. This limitation affects the uniqueness, accuracy of appearance, and emotional expression of digital identity. To address these challenges, our study employs a self-supervised learning framework leveraging the HIFI3D++ model to achieve high-fidelity 3D face reconstruction from monocular images. We introduce anatomically driven constraints and a facial shape denoising network with a Hybrid Normal Attention module to enhance geometric detail and suppress texture noise. The proposed method improves the quality of the facial shape and appearance details reconstruction, overcoming the misalignment issues seen in previous models. This study's unique contribution lies in its novel approach to balancing shape and appearance in 3D face reconstruction, realize reasonable high-frequency details for digital human modeling.