FGCNet: Face Geometry-Constrained Network for Face Point Cloud Completion

Raw face point clouds obtained from scanning are often incomplete, resulting in a loss of structural details and posing challenges for many tasks, such as facial surgery navigation, face recognition, and face correspondence. Existing point clouds completion methods generally learn a deterministic partial-to-complete mapping, but they overlook the structure prior information in objects to be completed. To address this limitation, we propose a Face Geometry-Constrained Network (FGCNet) to predict complete face point clouds with structure prior. For fully leveraging the similarity of face structures, we first use the complete point clouds to learn a face structure prior, which embeds sets of reference structures in a low-dimensional latent space. Then, to improve the quality of the structure prior, we incorporate diffusion models with this prior to predict coarse reasonable face structure. In addition, we design a feature attention encoder to enhance feature discriminability and an integrating-splitting decoder to combine information from different features. Experiments on two public datasets (Bosphorus and FRGC V2) demonstrate that our method outperforms state-of-the-art methods with an average completion error of 0.0015.