Evaluation of YOLO, SAM2, and U-Net Methods for Facial Attribute Segmentation Across Classes and Viewing Angles

Facial attribute segmentation is challenging because facial regions differ in size, shape, texture, and visibility, especially under non-frontal viewing conditions. This study evaluates the segmentation of four facial attribute classes: hair, eyebrows, mustache, and beard. A dataset of 5000 manually annotated FFHQ facial images was used to train and compare approaches within the same evaluation framework: YOLO segmentation, YOLO detection combined with SAM2 segmentation, and U-Net. The methods were evaluated on annotated test set, per facial attribute, and on an additional controlled phantom-based dataset acquired at five viewing angles. The hybrid YOLO detection and SAM2 segmentation pipeline achieved the best overall performance, with micro IoU of 0.820 and a micro Dice score of 0.893 on the annotated test set. Hair and beard are segmented more reliably than eyebrows and mustache, while segmentation accuracy decreased as the viewing angle increased. These results show that facial attribute segmentation performance depends on the selected method, target class, and acquisition viewpoint. The findings provide a basis for selecting and further improving segmentation methods for future registration and medical robotic applications.