Graph Neural Networks for 3D facial morphology: Assessing the effectiveness of anthropometric and automated landmark detection

This study investigates the potential of Graph Neural Networks (GNNs) for analyzing 3D facial morphology, leveraging facial landmarks as graph nodes to capture the intrinsic structure of 3D face scans. This research evaluates the effectiveness of three distinct approaches for defining graph vertices by associating them with: (1) a well-established set of anthropometric landmarks identified through tactile assessment, widely considered the gold standard in facial anthropometry; (2) automatically detected 3D facial keypoints estimated using advanced algorithms; and (3) geometry-based random point cloud sub-sampling via farthest point sampling (FPS). To evaluate the effectiveness of GNNs and facial landmarks in capturing and representing meaningful morphological patterns, the study employs two benchmark tasks: gender classification and age regression. Extensive experiments across various GNN architectures and three datasets — each presenting diverse and challenging conditions — demonstrate that semantically meaningful landmarks, whether anthropometric or automatically detected, consistently outperform non-semantic random samples in both tasks and across all datasets. These results highlight the crucial role of semantic contextualization in graph-based facial analysis. Notably, models utilizing automatically detected facial keypoints achieved performance comparable to those based on manually annotated anthropometric landmarks, offering a scalable and cost-effective alternative without compromising accuracy. These findings support the integration of automated GNN-based methodologies into a wide range of applications, including clinical diagnosis, forensic analysis, and biometric recognition. • Learning 3D facial features to classify gender and age. • Comparing anthropometric vs. automatic landmarks and FPS sub-sampling. • Characterizing graph nodes with local and global geometric features. • Comparing GCN, GAT, DynamicEdgeConv, and Transformer models.