Face Off / Computer vision · Team coursework

When should a model say ‘unknown’?

A recognition score is only part of the story. Compare correct identifications with unknown people mistakenly accepted at two reported operating points.

My contributionOwned testing and comparison design, predictions, and metric analysis, rather than every model implementation.

Two reported operating points · Report and presentation available

Conceptual artwork

Explore the project

Explore the operating points

When should a model say ‘unknown’?

Recognizing more known faces also admitted more unknown faces at these reported settings. Read both outcomes together.

Reported threshold 9.914608

Known probes correctly identified

70.69%

Unknown probes falsely accepted

0.33%

Separate populations · both bars use a 0–100% scale

At this operating point, logistic regression identifies fewer known probes and falsely accepts fewer unknown probes.

FaceNet embeddings · held-out evaluation · report pages 5–6
Inspect a reported operating point	Known probes correctly identified	Unknown probes falsely accepted
Logistic regression	70.69%	0.33%
Cosine distance	89.36%	1.33%

These are separately selected operating points, not a controlled comparison at the same false-acceptance rate. They do not establish a universally better decision rule. The PCA experiment uses a different protocol and is excluded from this comparison.

Read the measured results ↗

Logistic regression operating point: 70.69% identified / 0.33% falsely accepted
Cosine distance operating point: 89.36% identified / 1.33% falsely accepted

Follow the evidence

Written report Presentation

Scope & limitations

These are two separately selected operating points. They do not establish which rule is better at a matched false-acceptance rate.
The report's intended 'approximately 1%' false-acceptance constraint is not a strict bound: 1.33% exceeds 1%.
The PCA experiment uses different splits and a different minimum image count. Its results are never combined with the FaceNet rows.
Labeled Faces in the Wild skews toward public figures. The study did not measure demographic fairness.
A team project. Sanjay owned the testing and comparison design, predictions, and metric analysis, not every model implementation.

The engineering decision

Evaluate unknown identities separately. Attach results to their operating point and keep different evaluation protocols out of a misleading common ranking.

What the evidence means

The reported points do not establish superiority at equal false-acceptance rates. The study did not measure demographic fairness, and PCA used a different evaluation protocol.

Implementation notes

An evaluation study of classical and learned face representations, designed around both successful identification and the acceptance of unknown people. The two reported FaceNet decision rules are separate operating points, not a ranking: logistic regression identifies 70.69% of known probes while falsely accepting 0.33% of unknowns, and cosine distance identifies 89.36% while falsely accepting 1.33%. Neither result establishes superiority at a matched false-acceptance rate.

Team: Edward Kim · Priyanshi Singh · Murabet Sarsur · Sanjay Manivasagam