Efficient Unsupervised Clustering of Facial Geometry and Head Orientation Using 2D Landmarks

Facial geometry and head orientation play a crucial role in understanding human expressions and interactions. This study proposes an efficient, interpretable, and real-time framework for unsupervised clustering of facial geometry and head orientation using 2D geometric descriptors derived from dlib’s 68-point landmarks. The pipeline computes handcrafted descriptors, including lip aspect ratio, eye aspect ratio, normalized pupil position, and head pose angle, and leverages KMeans clustering to segment semantic facial states without supervision. Experiments conducted on FER and AffectNet datasets, demonstrate robust clustering performance, achieving Silhouette Scores of up to 0.721 for eye-state clustering and 0.567 for lip-state clustering. Comprehensive evaluation using the Davies–Bouldin Index and Calinski–Harabasz Score confirms the discriminative power of the proposed descriptors, while ablation studies highlight their individual contributions. Cross-dataset comparisons further reveal strong generalizability across modalities. With a lightweight design delivering 52 FPS on a CPU-only system and minimal memory footprint, the framework is well-suited for real-time applications, including fatigue monitoring, viseme synthesis, and human–computer interaction. Unlike prior geometric studies that focus on isolated modalities or deep clustering pipelines, this work presents a unified and interpretable geometry-based clustering paradigm that integrates four facial modalities—lips, eyes, pupils, and head orientation. The framework achieves comparable clustering quality to lightweight deep baselines while remaining fully transparent and CPU-real-time. This makes it well-suited for edge deployment and human–computer interaction scenarios. GitHub: https://vineetkumarrakesh.github.io/2D-Geo-Clustering