Dual-Expert Landmark Localization in 3D Facial Point Clouds Under Controlled Synthetic Local Surface Loss for Rigid Initialization

Local surface loss can destabilize landmark-based rigid initialization in three-dimensional (3D) facial point clouds. We propose a dual-expert framework for localizing five anatomical landmarks under controlled synthetic surface loss. The Clean expert is optimized for peak-based localization on complete surfaces, whereas the Occlusion expert combines local coordinate regression, visibility estimation, heteroscedastic modeling, and a global structural prior. A model-output reliability gate removes dependence on the protocol-supplied surface-loss ratio. For scenes not classified as reliably complete, the Occlusion expert provides the default and fallback prediction, while a validation-selected Clean residual is applied only under landmark-wise agreement. On a subject-independent FaceScape test set, mean localization errors were 0.470±0.688, 1.159±1.093, and 2.216±2.169 mm at 0%, 30%, and 50% surface loss. The method significantly outperformed the Unified occlusion-aware and structure-robust (OASR) model and the Occlusion expert at 30% and 50% loss after Holm correction and achieved lower error than Unified OASR in 22 of 24 structured-corruption conditions. In a controlled large-pose stress test, initialization achieved 100% iterative closest point (ICP) success versus 98.82% for the two-stage stratified graph convolutional network (2S-SGCN) baseline and improved coarse alignment, whereas post-ICP accuracy did not differ significantly. These results support robust rigid initialization under controlled synthetic surface loss on FaceScape; generalization to real sensor-acquired point clouds remains to be established.