FaceBench: A Locked-Pipeline Comparative Evaluation of Five Off-the-Shelf Face Recognizers across Pose, Age, and Resolution

Abstract

 Fair comparison of pretrained face recognition models is often confounded by vendor-specific detectors, incompatible crop geometries, and accuracy-only reporting at an unstated decision threshold. This paper presents FaceBench, a reproducible evaluation framework, and reports a locked comparative study of five off-the-shelf recognizers — FaceNet, Dlib, InsightFace Buffalo-L, AdaFace, and MagFace — under one shared RetinaFace/SCRFD detect-and-crop protocol (Baseline B: bounding-box margin 0.35, cosine operating point 0.40). Primary evidence spans pairwise verification on four high-quality (HQ) benchmarks — LFW, CFP-FP, CPLFW, and AgeDB-30 — covering frontal, cross-pose, and cross-age conditions, together with dedicated uncached embedding throughput, pairwise McNemar significance tests against the leading model, LFW closed-set Rank-1 identification as a function of gallery size N ∈ {10, 100, 500, 1000}, and an accuracy–efficiency synthesis. Buffalo-L attains the highest mean verification accuracy across the four HQ sets (0.9045) at 27.63 frames/s on the reported workstation; MagFace and AdaFace are the fastest embedders (97.60 and 92.83 frames/s) but measurably weaker on hard-pose accuracy at the shared cosine threshold; Dlib retains a high ROC-AUC and low equal error rate (EER) yet its fixed-threshold accuracy collapses to chance level because the fixed cosine threshold of 0.40 is poorly calibrated for Dlib's similarity distribution. Every accuracy gap against Buffalo-L is statistically significant under McNemar's test (all p < 0.001; closest is CPLFW versus FaceNet, p = 4.2×10⁻⁵). A supplementary TinyFace Rank-1 probe (Gallery_Match versus Probe, N ≤ 500) is reported separately from the HQ means and is explicitly not official TinyFace mean average precision. Source code and experiment manifests are released under an MIT license; face images are used under each dataset's original terms and are not redistributed.