Mask-Region Landmark Prediction in 3D Masked Face Point Clouds Using Non-Mask Facial Landmarks

Due to the impact of the COVID-19 pandemic, the $8^{\text {th }}$ Korean human body measurement survey conducted by the Size Korea Center between 2020 and 2024 includes a subset of 3D scans in which subjects are wearing face masks. Such data limit the utility of the resulting 3D scans for subsequent point-cloud-based learning and analysis. This study addresses this issue by predicting facial landmarks in the mask region of 3D masked face point clouds using only facial landmarks outside the mask region. To this end, we propose a simple two-stage model that first applies positional encoding to non-mask facial landmarks and then uses an MLP regressor to estimate the missing landmarks in the mask region. The model is trained on an unmasked face dataset and qualitatively evaluated on a masked face dataset from the same survey. As a result, we observed both cases in which mask-region landmarks were accurately reconstructed and failure cases where large variations in facial shape led to unstable predictions. Based on these observations, we analyzed the limitations and underlying causes of the our setting and discussed possible directions for future research.