As face recognition systems gain wider deployment and the amount of prospective data available for enrollment increases, the need to manage the size of the gallery of enrolled identities has emerged as a design consideration. Gallery editing offers the opportunity to exclude exemplars that are not useful in matching, either because of their strong similarity to other exemplars, or because they represent poor quality and redundant face extractions. This work assesses two simple yet effective approaches to gallery size reduction. One method streamlines facial galleries by leveraging Mean-Aware Cosine similarity and One-Sided Selection followed by sample mean estimation to create a reduced and representative subset of feature vectors in the gallery. The other approach trains a multiclass SVM on the gallery samples and retains only those samples selected as support vectors. The proposed methods were evaluated on the PaSC and IJB-S video datasets developed for the NIST MBGC program and the IARPA JANUS program, respectively. Experimental results demonstrate improvement of TAR@FAR=1% from 86.65% to 91.83% on the PaSC video dataset. The Ranks 1 and 5 correct match percentages on the IJB-S dataset (Surveillance-to-Surveillance task) rise from 23.53% to 67.16% and from 36.87% to 74.41%, respectively.
Paper
Full text
Data-driven approaches for gallery editing in face recognition
OpenAlex · Face recognition and analysis · 2025
Abstract
As face recognition systems gain wider deployment and the amount of prospective data available for enrollment increases, the need to manage the size of the gallery of enrolled identities has emerged as a design consideration. Gallery editing offers the opportunity to exclude exemplars that are not useful in matching, either because of their strong similarity to other exemplars, or because they represent poor quality and redundant face extractions. This work assesses two simple yet effective approaches to gallery size reduction. One method streamlines facial galleries by leveraging Mean-Aware Cosine similarity and One-Sided Selection followed by sample mean estimation to create a reduced and representative subset of feature vectors in the gallery. The other approach trains a multiclass SVM on the gallery samples and retains only those samples selected as support vectors. The proposed methods were evaluated on the PaSC and IJB-S video datasets developed for the NIST MBGC program and the IARPA JANUS program, respectively. Experimental results demonstrate improvement of TAR@FAR=1% from 86.65% to 91.83% on the PaSC video dataset. The Ranks 1 and 5 correct match percentages on the IJB-S dataset (Surveillance-to-Surveillance task) rise from 23.53% to 67.16% and from 36.87% to 74.41%, respectively.