Deep face recognition has achieved near-saturated performance on several controlled benchmarks, yet its robustness in unconstrained real-world scenarios remains a critical challenge. This paper presents a comparative analysis of three backbone architectures,a classification-based SE-ResNet-50 and two metric-learning-based models, IR-ResNet-50 and IR-SE-ResNet-50 trained with ArcFace loss on the large-scale VGGFace2 dataset and evaluated on the challenging IJB-C verification benchmark. All models are trained from scratch on a 4,605-identity subset of VGGFace2, achieving strong in-domain performance with validation accuracies in the 94-96% range and verification AUC of approximately 0.955 on VGGFace2 validation pairs. However, when evaluated on the IJB-C 1:1 verification protocol, the same models exhibit a dramatic drop in performance. IR-ResNet-50 with ArcFace achieves only 1.10% true accept rate (TAR) at false accept rate (FAR) 10-3 and 0.36% at FAR 10-4. IR-SE-ResNet-50 with alignment and MTCNN-based cropping marginally improves TAR to 1.75% at FAR 10-3 and 0.57% at FAR 10-4, while SE-ResNet-50 trained with softmax classification surprisingly performs best among the three, reaching 28.02% TAR at FAR 10-3 and 11.99% at FAR 10-4. These results highlight a severe generalization gap between curated training data and unconstrained benchmarks and suggest that strong margin-based losses alone are insufficient to guarantee cross-domain robustness. We further analyze failure modes related to pose, alignment mismatch, and template noise, and motivate a Pose-Aware Feature Rectification Network (PAFR-Net) as a future direction to improve pose robustness by learning pose-conditioned residual corrections in feature space.