Overview

As a Graduate Researcher at Carnegie Mellon University (February–May 2026), I worked on face-recognition evaluation, score-level fusion, and detector robustness. The work included two separate evaluation tracks: verification of face templates, and successful face detection under challenging image conditions.

Face-recognition evaluation

I implemented the full IJB-C evaluation protocol for two face-recognition systems, processing approximately 469,000 image records, 23,000 templates, and 15.7 million template pairs. I reproduced the results of published model checkpoints at low false-accept-rate operating points.

I then implemented score-level fusion. The combined scores improved the true accept rate at four of five false-accept-rate operating points. Analysis of score correlations and overlapping false accepts showed complementary error patterns concentrated among hard negatives.

On the hard-negative tail, the score correlation was Spearman ρ = −0.415, and the false-accept sets had a Jaccard overlap of 0.178 at FAR = 10⁻⁴. Based on this analysis, I proposed topology-aware training fusion through late-stage boundary refinement and prototype-level topology regularization. This was a training proposal, separate from the implemented score-level fusion results.

These results describe verification performance at fixed false-accept rates; they are not a general classification accuracy or a claim that every operating point improved.

Detector robustness

I built a traceable benchmark with six evaluation slices and 1,885 unique images, then developed a geometry-aware retry policy for face detection.

On a frozen pool of 1,835 images, the policy increased successful detections from 1,727 to 1,777, recovering 50 additional detections at an intersection-over-union (IoU) threshold of at least 0.5. The improvement was also reproduced with an independent detector implementation.

The full benchmark and the frozen evaluation pool are different sets. Successful-detection counts are reported separately from recognition true accept rates and should not be read as average precision (AP).

Scope

The recognition and detection experiments were evaluated separately. I have not established that the detector retry policy improves an end-to-end detector-to-recognizer system.

Tools used in this work include Python, PyTorch, NumPy, and OpenCV. No face images, evaluation records, or research attachments are distributed on this website.