MedoidFormer: A Medoid-Based Transformer for Anthropometric Landmark Localization in Respirator Fit Test Panel Development

Developing a respirator fit test panel (RFTP) is crucial for evaluating the fit of respiratory protective equipment (RPE) across diverse facial structures. An RFTP comprises multiple headform cells representing various face shapes. Current methods assign individuals to these cells using manual facial measurements with calipers and measuring tapes, which are costly, time-consuming, and prone to human error. Landmark localization offers a promising alternative but faces challenges such as occlusions, texture variability, and high computational demands. To address these limitations, we propose MedoidFormer, a novel transformer-based method that localizes 3-D facial landmarks directly from point clouds, without 2-D projections or textures. MedoidFormer introduces medoids as representative points that capture local geometric structures in the point cloud. Our method comprises four key components: medoid generation, neighbor attention, feature aggregation, and landmark generation. First, medoid generation downsamples the point cloud into medoids. Neighbor attention refines the medoids’ features by modeling their interactions. Feature aggregation combines the refined features while preserving detailed facial geometry. Landmark generation leverages the refined features to predict 3-D landmark coordinates, which are used to compute anthropometric dimensions for assigning individuals to headform cells in RFTP development. Furthermore, we introduce two new datasets: the anthropometric landmark dataset (ALD) and UOW3D-36/10. Both datasets consist of high-quality 3-D face models annotated with 18 anthropometric landmarks. MedoidFormer outperforms existing methods, achieving a mean Euclidean distance error (MEDE) of 0.57 mm on the FaceScape dataset for landmark localization, and accuracies of 79.31 % and 77.14 % on the UOW3D-36 and UOW3D-10 datasets, respectively, for headform cell assignment.