Impact of Facial Variability on Shape-Only Landmark-Based Face Verification

Abstract

Facial landmarks are commonly used in face recognition for alignment, pose normalization, and quality assessment, but they can also serve as an interpretable representation of facial shape. This paper studies shape-only 1:1 face verification using dense 2D facial landmarks, without texture descriptors, deep face embeddings, or identity-specific neural network training. After Procrustes normalization, three geometric representations are evaluated: pairwise distances, distance ratios, and aligned land-mark coordinates. Robustness is assessed by applying controlled landmark-level variability, including random coordinate noise, regional perturbations, and dropout. Experiments on AgeDB-30 and CFP-FP use 7,463 valid images and 13,898 verification pairs, showing limited absolute discriminative performance, with the best AUC of 0.5706 and EER of 0.4315. Landmark instability affects verification in a region-dependent manner: high dropout causes the strongest average degradation ( ∆EER = 0.0247, ∆AUC=−0.0235), while mouth and eye perturbations produce the strongest regional degradation. Statistical testing confirms that most perturbation conditions produce significant changes relative to baseline. The study provides an interpretable benchmark for analyzing how facial shape variability influences landmark-based biometric verification.