Facial Recognition: Precision Up 10–25%, False Positives Down 30–60%
In identity verification, a false positive is not a rounding error. Most of the gain came from preprocessing and threshold work, not from a bigger model.
- Client
- Identity verification deployment
- Industry
- Computer vision
+10–25%
Precision over baseline
-30–60%
False positives
Production
Deployment-ready pipeline
Validated
Held-out evaluation
The problem
An off-the-shelf facial recognition model performed acceptably on benchmark data and unacceptably in the deployment environment. Real conditions — variable lighting, angles, image quality, occlusion from glasses and masks — produced a false positive rate that made the system unusable for identity verification.
In this application the two error types are not equivalent. A false negative inconveniences someone who then tries again. A false positive admits the wrong person, and the cost of that is not measured in user experience.
Where the gains came from
Not from a larger model. The first move was a preprocessing pipeline matched to the actual capture conditions: alignment, normalisation of lighting variance, and quality gating that rejects frames too poor to judge rather than guessing from them.
Then training strategy — augmentation reflecting the real distribution of angles and conditions rather than a generic set, and sampling that stopped the easy cases from dominating the loss.
Then threshold tuning, which is where most of the false positive reduction actually lives. The default operating point on a similarity score is a vendor's guess at your risk tolerance; setting it against your own cost matrix is close to free and is routinely skipped.
Evaluation
Evaluation ran on held-out data from the deployment environment, not on the public benchmark the base model was trained against. Benchmark numbers were treated as irrelevant to the decision, because they measure a different distribution.
We reported precision, recall and the confusion matrix at the chosen operating point rather than a single accuracy figure, because accuracy on an imbalanced verification task is a number that flatters everyone and informs no one.
Results
Precision improved 10–25% over the baseline model depending on the condition segment, and false positives fell 30–60% through the combination of quality gating and threshold tuning.
The system is deployment-ready for verification use, with an explicit operating point chosen against the client's risk tolerance rather than inherited from a default.
What we'd flag
The ranges reflect genuine variation across conditions. Well-lit frontal capture sits at the top; poor lighting and extreme angles sit at the bottom, and honest reporting means saying so rather than quoting the best segment.
Facial recognition in production carries obligations beyond accuracy — consent, retention, jurisdiction and bias evaluation across demographic groups. Those are not engineering afterthoughts, and any deployment plan that treats them as such will fail somewhere more expensive than a model metric.
