Working-paper summary · September 2026
Evaluating Local One-to-Many Identification
A design and evaluation study of local identification, covering gallery size, image quality, label noise and the choice between automatic decisions and human review.
The research question
Matching a person against an enrolled gallery raises a different question from checking a claimed identity. There are many opportunities for another candidate to score highly, and a useful system also has to decide when it cannot make a reliable choice. This study examines those decisions alongside the recognition model.
Make the error definitions explicit
The evaluation separates failure to acquire a usable face, failure to return the correct enrolled identity and acceptance of an incorrect identity. It distinguishes settings where unregistered people may appear from those assumed to contain only enrolled people. Comparisons use matched operating rules, paired outcomes and uncertainty intervals suited to the trial structure.
Treat the data as something to test
Public datasets can overlap in identity even when their labels differ. The study audits those overlaps and records corrected scoring alongside earlier results. It also examines difficult lighting, small faces, enrolment coverage and repeated observations, with detection failures included in the evaluation.
What the current evidence supports
The work includes a functioning local prototype, public-dataset experiments and simulated station replays. Some more complex approaches did not improve the comparison, and adaptation remains an active research question. These experiments help refine the system and its next study; they do not estimate its accuracy in an untested workplace. This page summarises a working draft rather than reproducing its private deployment brief.