Gallery-Distance Profiles for Open-Set Person Re-Identification with Frozen Encoders in Camera Networks

Open-set person re-identification (ReID) must determine whether a query identity is represented in the current gallery before returning a cross-camera match. We investigate whether the complete query-to-gallery distance distribution provides more reliable presence evidence than its nearest value when the encoder is frozen and target unknown identities are unavailable. The proposed Source-Episodic Distance-Statistics Evaluator (SDSE) projects each distance vector onto nine interpretable local, global, and candidate-conditioned statistics. Auxiliary identity episodes reproduce the gallery-presence hypotheses and fit a low-capacity evaluator; leave-one-out scores from the enrolled target gallery calibrate deployment without changing the encoder or accepted-query ranking. Across Market-1501, DukeMTMC-reID, and CUHK03, SDSE gives the strongest macro unknown-detection result among seven frozen-embedding scorers. A 20-split paired analysis confirms a 0.0095 gain in area under the receiver operating characteristic curve (AUROC) over nearest distance, with the clearest benefits under cross-camera-domain transfer. Controlled gallery perturbations preserve the relative advantage but show that sparse identity evidence limits absolute performance. Grouped-profile experiments further identify the combination of local and global distribution evidence as the main source of the gain. An auxiliary operating-point gate and a six-candidate encoder-selection experiment illustrate how the common profile interface can support deployment decisions, while their measured scope is kept separate from the central scoring contribution. The results establish SDSE as a lightweight gallery-relative decision layer for fixed-gallery open-set ReID rather than a substitute for representation learning.