Development Update

Putting Face Recognition to the Test

What our first on-device experiments taught us about matching faces, privacy, and why we are still evaluating age estimation.

5 min readDevelopment journal

This week we moved from talking about facial recognition to testing it on real photographs. Our question is simple: can Octui make a large collection easier to explore while keeping the people in those pictures in control?

What we tried

We built an isolated Android evaluation for Luxand FaceSDK and ran it on a Samsung phone. We tested face detection, face-to-face comparisons, recurring-person grouping, and age estimates. We also measured how long analysis took, rather than relying on a promising demo.

What the first results actually tell us

Our initial 24-photo diagnostic set was useful for uncovering problems, but it is far too small to declare a winner. A strict grouping threshold avoided many mistaken matches but missed a large number of genuine matches. Lowering that threshold helped recall, though it also increased the risk of grouping different people together.

We also caught an unusual face-boundary result that needs additional handling before it can be trusted in the app. This is why testing matters: small technical details can become confusing experiences for real families.

Age estimation remains an open question

Our first age estimates were inconsistent. That does not mean age-aware organization is off the roadmap. We still want to evaluate InsightFace more thoroughly, including photographs taken many years apart, before making a decision. The same goes for recognizing a person from childhood into adulthood: it's a goal, not a capability we've validated yet.

What matters more than a clever demo

False matches can undermine trust. We want Octui to suggest connections and let you review them, not silently rewrite the story of your family. On-device processing remains a priority because photos should not have to leave your device just to become easier to organize.

What's next

We'll expand the evaluation with a larger, carefully checked photo set, compare the recognition tools under similar conditions, and keep improving the basic import and project experience. We're preparing for a small friends-and-family beta, not claiming that these experimental AI features are ready to ship.

Thanks for following the build. We'll keep sharing what worked, what surprised us, and what still needs work.

Join the Octui early-access list → · Read our guide to old printed photos →