A Deep Side-View to Side-View Face Recognition System

Pose variation in unconstrained settings challenges face recognition, especially for side-view poses, which are affected by asymmetry, self-occlusion, and limited facial features. While deep learning has improved extreme pose handling, most methods focus on frontalization, with datasets favoring frontal views. As a result, generalization remains a challenge, with side-view poses facing difficulties in detection, alignment, and feature extraction, an issue often overlooked in the literature. This study addresses the unique challenges of side-view-to-side-view recognition. We build a recognition system leveraging transfer learning to achieve promising performance on the CMU Multi-PIE and UT-Door datasets. Next, we conduct cross-dataset evaluations, comparing deep features with Local Binary Patterns (LBP), an untrained algorithm. Results show that LBP exhibits strong robustness to cross-dataset variations and can outperform deep learning-based methods, underscoring the ongoing limitations in generalization.