Complementarity at Low False-Accept Rates

Face verification at low false-accept rates depends on how models handle a small tail of hard negatives. At Carnegie Mellon University, I implemented the full IJB-C evaluation protocol and used complementary model errors to improve verification. Across 15.7 million template pairs, score-level ensembling surpassed both TopoFR and LVFace at four of five operating points.

Hard-negative analysis explained why fusion was useful: score correlation was Spearman ρ = −0.415, while false-accept sets had a Jaccard overlap of 0.178 at FAR = 10⁻⁴. I used this evidence to formulate a topology-aware training strategy combining late-stage hard-example reweighting with prototype regularization. The measured gains came from score-level ensembling; the training strategy was a proposal informed by that analysis.

Geometry-Aware Adaptive Inference

In a separate detection study, I developed a geometry-aware retry policy that recovered faces missed by the baseline detector. On the fixed 1,835-image robustness benchmark, successful detections increased from 1,727 to 1,777 at IoU ≥ 0.5: 50 recovered detections and approximately 46% fewer failures.

The approach also improved an external RetinaFace implementation’s detection success from 80.0% to 90.9%. Together, these experiments paired controlled evaluation with targeted inference changes, using Python, PyTorch, NumPy, and OpenCV.