Automatic Detection of Peripheral Facial Landmarks Using 3D Facial Data

Facial landmarking research has traditionally focused on midface landmarks, largely overlooking peripheral facial features that are crucial in specialized medical applications. This study addresses this gap by developing a novel model to detect 12 peripheral landmarks, which are valuable in clinical scenarios such as obstructive sleep apnea detection and facial palsy assessment. Using the FRGC v2 dataset of 3D facial scans, we manually annotated a subset of 900 scans with peripheral landmarks following a rigorous pre-processing protocol. Building upon a high-performance deep learning model initially designed for midface landmark detection, we restructured its architecture to predict peripheral landmarks by modifying its output layers and training it from scratch using 2D projections of 3D scans. We conducted a comprehensive set of experiments, systematically testing different input configurations by varying both the number of images per scan (16, 32, and 64) and the image types (mesh-only, RGB, and depth maps). The model demonstrated its potential by achieving a mean error of 10.72 mm, highlighting the promise of deep learning techniques in addressing the complex challenge of peripheral landmark detection. These findings underscore the model's capability to support niche medical applications and open pathways for future advancements that could refine its accuracy, ultimately contributing to improved diagnostic tools in healthcare.