Rank-1 Identity Consistency: Gallery Membership Classification for Operational 1:N Contexts

Gabriella Pangelinan, Aman Bhatta, Michael C. King, Kevin Bowyer

Keywords: Face recognition, face identification, demographic bias, fairness, open-set identification

TL;DR: Classifying 1:N search results as in- or out-of-gallery using rank-1 identity agreement across multiple matchers outperforms score-based thresholds in accuracy and fairness, especially under degraded image conditions.

Abstract:

In real-world deployments, 1:N face identification systems face a fundamental question: is a probe image's identity enrolled in the gallery or not? We propose the first approach to using consistency of rank-1 identity across multiple matchers as a method to classify the result of 1:N search as in-gallery (ING) or out-of-gallery (OOG). Our "1-consistency" method classifies a probe as ING if all matchers return the same rank-1 identity, and OOG otherwise. We compare its performance to two threshold-based methods: score-thresholding (using raw similarity scores) and gap-thresholding (using the score difference between rank-1 and rank-2 identities).

We evaluate these methods across 12 experimental configurations that systematically vary image quality, gallery enrollment structure, and demographic composition. On average, 1-consistency achieves the highest ING recall (92.0% vs. 72.1% score, 79.6% gap), highest overall accuracy (92.8% vs. 81.4% score, 89.6% gap), and lowest demographic disparities (4.6 pp vs. 10.5 pp score, 8.5 pp gap). Under degraded-probe conditions—the most operationally relevant scenario—1-consistency achieves win margins averaging 16.1 pp for ING recall and 6.1 pp for overall accuracy. With this combination of quality robustness and demographic fairness, 1-consistency is well-suited for real-world contexts like law enforcement investigations—and has the potential to reduce wrongful arrests and enable better allocation of investigative efforts.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 8