Gallery-Distance Profiles for Open-Set Person Re-Identification with Frozen Encoders in Camera Networks Authors: Jin Wang, Jiannong Cao, Changlin Song Published: 2026 ABSTRACT 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. 1. Introduction Person re-identification (ReID) associates a person observed by one camera with gallery images collected by other cameras. Standard benchmarks formulate this task as closed-set retrieval: every query is assumed to have a valid gallery match. Camera networks violate that assumption whenever non-enrolled people enter the scene. A nearest-neighbor system must nevertheless return an identity, so a visually plausible impostor can become a false match. The risk is amplified when deployment cameras differ from the representation-training environment in viewpoint, illumination, background, resolution, or detector quality [1,2]. This study asks a specific scientific question: when the ReID representation is fixed, can the full query-to-gallery distance distribution reveal gallery membership more reliably than the nearest distance alone? The target problem is gallery-relative open-set ReID. Each query must either be linked to a valid enrolled cross-camera identity or rejected because its identity is absent from the current gallery. The question is deliberately downstream of representation learning: the encoder and gallery ranking remain unchanged, and target unknown identities are unavailable for learning or calibration. Nearest-distance rejection is a natural starting point because nearest-neighbor evidence is effective for novelty detection in deep feature spaces [3]. Its limitation is statistical: one extreme observation stands in for hundreds or thousands of gallery comparisons. A difficult enrolled query can have a relatively large nearest distance yet retain a coherent identity-specific neighborhood. Conversely, a non-enrolled query can acquire one accidental close match while remaining inconsistent with both that candidate identity and the wider gallery. These cases suggest treating identity presence as a distributional hypothesis test rather than a one-distance threshold test. The Source-Episodic Distance-Statistics Evaluator (SDSE) answers this question through a fixed nine-dimensional profile. Lower-tail order statistics describe proximity and ambiguity; global location, spread, and upper-tail summaries locate the query within the complete gallery distribution; and a candidate-conditioned dispersion contrast measures the coherence of the nearest identity. A class-balanced logistic evaluator combines these signals without changing the encoder. Grouped-profile, nested-profile, and leave-one-feature-out experiments test which parts of the design carry the evidence, rather than assuming that nine is an optimal dimension. The protocol separates evaluator fitting, threshold calibration, and target evaluation. Auxiliary episodes provide known and pseudo-unknown distance profiles, with the two identity sets disjoint within each episode. A regularized logistic evaluator is fitted to these profiles. Leave-one-out scores from the enrolled target gallery then calibrate the rejection threshold, while official target queries are reserved for evaluation. A separate gallery internal resampling procedure is used to compare frozen encoders. Neither procedure uses target unknown labels. SDSE produces a continuous unknown-risk score. A separate Unknown-Candidate Gate (UCG) examines the binary operating point through gallery-distance energy and diagonal Mahalanobis deviation, while the fixed-dimensional profile also supplies a common interface for a gallery-only encoder-selection experiment. These two deployment analyses are evaluated independently of the central score: UCG does not alter threshold-free ranking, and candidate selection occurs before official target queries are observed. Our 2023 arXiv preprint introduced the nine-statistic representation and the candidate-positive/candidate-negative coefficient-of-variation (CV) contrast [4]. Table 1 distinguishes that representation from the present gallery-presence formulation, source-episodic evaluator, target-known calibration, official Rank-1, mean average precision (mAP), and open-set classification rate (OSCR) protocol, and deployment analyses. The contributions form a single claim–mechanism–evidence chain:We formulate frozen-encoder, gallery-relative open-set ReID as a distributional presence test and derive three observable evidence groups from the known/unknown mixture structure of query-to-gallery distances.We realize this principle as SDSE: an interpretable fixed-dimensional profile, a low-capacity evaluator fitted from identity-disjoint auxiliary episodes, and target-known-only calibration. The ReID representation is never adapted, so accepted-query ranking is preserved.We connect each design claim to paired evidence: seven post hoc scorers, 20 identity splits with bootstrap and non-parametric inference, grouped and nested profile analyses, and controlled perturbations of gallery size, samples per identity, imbalance, and query prevalence.We also use the profile for operating-point control and pre-query encoder selection. These analyses identify both the useful mechanism and its boundary: energy explains the observed UCG behavior, and model-selection evidence is confined to the tested six-candidate bank. We formulate frozen-encoder, gallery-relative open-set ReID as a distributional presence test and derive three observable evidence groups from the known/unknown mixture structure of query-to-gallery distances. We formulate frozen-encoder, gallery-relative open-set ReID as a distributional presence test and derive three observable evidence groups from the known/unknown mixture structure of query-to-gallery distances. We realize this principle as SDSE: an interpretable fixed-dimensional profile, a low-capacity evaluator fitted from identity-disjoint auxiliary episodes, and target-known-only calibration. The ReID representation is never adapted, so accepted-query ranking is preserved. We realize this principle as SDSE: an interpretable fixed-dimensional profile, a low-capacity evaluator fitted from identity-disjoint auxiliary episodes, and target-known-only calibration. The ReID representation is never adapted, so accepted-query ranking is preserved. We connect each design claim to paired evidence: seven post hoc scorers, 20 identity splits with bootstrap and non-parametric inference, grouped and nested profile analyses, and controlled perturbations of gallery size, samples per identity, imbalance, and query prevalence. We connect each design claim to paired evidence: seven post hoc scorers, 20 identity splits with bootstrap and non-parametric inference, grouped and nested profile analyses, and controlled perturbations of gallery size, samples per identity, imbalance, and query prevalence. We also use the profile for operating-point control and pre-query encoder selection. These analyses identify both the useful mechanism and its boundary: energy explains the observed UCG behavior, and model-selection evidence is confined to the tested six-candidate bank. We also use the profile for operating-point control and pre-query encoder selection. These analyses identify both the useful mechanism and its boundary: energy explains the observed UCG behavior, and model-selection evidence is confined to the tested six-candidate bank. Figure 1 gives the component-level view of the gallery-conditioned score and auxiliary gate, while Figure 2 places these components in the complete selection, calibration, and deployment pipeline. Figure 3 separates the three uses of data: evaluator learning, gate-weight selection, and target-known calibration. Together, they make explicit which quantities are learned, which are calibrated, and which remain frozen. 2. Related Work 2.1. Closed-Set and Open-World Person Re-Identification Closed-set ReID learns an embedding that brings cross-camera images of the same person together while separating different identities. Cross-domain methods improve transfer through image translation, pseudo-labeling, clustering, feature alignment, or domain-general representation learning [1,2,5,6]. OSNet, used in our main experiments, aggregates multiple receptive scales in an efficient feature extractor [7]. These approaches improve representation and ranking, but they normally retain the assumption that every query has a gallery match. A transferable encoder can, therefore, reduce retrieval errors without defining when no enrolled identity is valid. Liao et al. formalized open-set person ReID as two coupled subproblems: detecting whether the probe identity is present in the gallery and, if accepted, identifying the corresponding person [8]. Their protocol trained general metric learners on independent identities and showed that closed-set retrieval accuracy did not translate into strong low-false-accept open-set performance. This detection–identification coupling is a central distinction between open-set ReID and sample-wise novelty detection. Subsequent work has mainly strengthened either the representation or the gallery model. In watch-list ReID, Adversarial PersonNet jointly trains a generator, feature extractor, and target discriminator against target-like impostors [9]. DistributionNet represents each image by a feature distribution and learns uncertainty-aware embeddings; its open-world evaluation reports true-target rate at controlled false-target rates [10]. Both approaches alter the learned representation for open-world verification. A different line treats the gallery itself as dynamic: Casao et al. cluster incoming unknown samples and incrementally create or update identity models [11], while recent dynamic benchmarks vary identities, cameras, and time [12]. Existing open-world ReID research, therefore, spans three settings: detection and identification against a static watch list, representation learning that is robust to impostors or uncertainty, and online discovery with gallery expansion. We study a complementary deployment setting in which the encoder and enrolled gallery are fixed, target unknown labels are unavailable, and absent identities must be rejected without changing the ranking of accepted queries. We also consider the preceding deployment decision of choosing a frozen encoder. Table 2 summarizes the differences. Domain shift and identity novelty are related but distinct. A known query may be difficult because its camera conditions differ from those seen during representation training; an unknown query may resemble an enrolled person even without a domain change. SDSE does not predict domain membership. It estimates whether the query identity is absent from the current gallery, and its evaluation retains the official cross-camera correctness requirement for known identities. 2.2. Open-Set Recognition and Adjacent Novelty Detection Open-set recognition assigns known classes while reserving a rejection outcome for unseen classes. OpenMax fits class-specific extreme-value models to a trained classifier’s activation vectors, whereas PROSER changes classifier training by reserving placeholders for prospective unknown classes [13,14]. These are important representation-and-classifier methods, but neither is a drop-in scorer for a non-parametric ReID gallery: OpenMax requires a fixed activation vector and class-wise calibration, and PROSER requires retraining the classifier and feature space. Treating every enrolled identity as a newly trained parametric class would change the frozen-encoder condition studied here and confound representation changes with post hoc scoring. The directly comparable family is novelty detection on a fixed reference embedding set. Mahalanobis scoring estimates global deviation from enrolled features, deep nearest-neighbor detection uses local reference distances, and local outlier factor compares local density [3,15,16]. Energy provides a smooth aggregate of distances [17]. The experiments compare nearest distance, mean 5-nearest-neighbor distance, local outlier factor, shrinkage Mahalanobis distance, energy, and unit-normalized cosine distance on identical frozen embeddings and identity splits. This design isolates the unknown-scoring rule while leaving gallery retrieval unchanged. Unlike fixed-class classification, ReID assigns identities through a replaceable gallery. One identity may have several cross-camera samples, and retrieval rather than a parametric class head determines the predicted identity. A query is, therefore, “known” only relative to the current gallery. The nearest distance retains one extreme observation; SDSE tests whether additional order, distributional, and identity-conditioned evidence improves that rule. Energy and Mahalanobis scores summarize different forms of incompatibility. UCG searches their weight on auxiliary pseudo-open-set splits and calibrates the threshold from target known-gallery samples. The search is informative only when it can suppress an unhelpful cue; the observed selections favor energy, locating the operating-point effect in that component rather than in general two-cue complementarity. Three related terms describe different system functions. Unknown-person detection asks whether a query identity is absent from the current gallery. Open-set person ReID combines that decision with the correct identification of accepted enrolled queries under a fixed gallery. Open-world person ReID is broader and may include discovery, temporal association, continual adaptation, and enrollment of previously unknown people. SDSE provides the first function as a score and is evaluated end-to-end on the second; it does not implement the dynamic functions of a complete open-world system. Evaluation must, therefore, distinguish between score separation, retrieval, and a selected operating point. Table 3 summarizes the reported quantities. AUROC and area under the precision–recall curve (AUPR) measure threshold-free unknown-person detection; the false-positive rate at 95% unknown recall (FPR@95) evaluates one point on the receiver operating characteristic curve. Rank-1 and mAP measure known-person retrieval. OSCR couples detection and retrieval: a known query contributes only if it is both accepted and correctly identified. Known acceptance and unknown rejection describe one calibrated threshold rather than score ranking. 2.3. Episodic Learning and Unsupervised Model Selection Episodic learning varies support and query label sets so that training resembles the intended test-time task [18,19]. PEELER uses novel-class episodes to adapt a few-shot recognizer for open-set decisions [20]. SDSE uses episodic learning in a more restricted way: auxiliary identities generate known and pseudo-unknown distance-profile tasks, but the ReID representation is fixed. Only a low-capacity evaluator is learned over the nine statistics available from any new gallery. Selecting a model without target labels is difficult because its deployment performance cannot be measured directly. MetaOD transfers performance information from historical outlier datasets and uses dataset meta-features to choose an unsupervised detector [21]. AutoMS instead selects a novelty detector from clean reference data while controlling false discovery through jackknife prediction [22]. Both methods address generic detector or hyperparameter selection; they do not include cross-camera identification correctness or a replaceable identity gallery. Our exploratory selector uses the enrolled ReID gallery itself: cross-camera identities are divided into support samples, pseudo-known queries, and withheld pseudo-unknown queries, producing an internal ranking by the same AUROC or OSCR criterion used at deployment. Real target unknowns and official benchmark queries are not used. Equal identity and episode budgets provide a balanced comparison, and all candidates remain frozen. Its validation is limited to the six candidates described in Section 4.6; broader architectural and source-domain generality remains untested. Table 4 locates SDSE relative to these neighboring settings. Its role lies at the decision and deployment-selection stages, downstream of representation learning. 3. Method 3.1. Problem Formulation Let a frozen encoder, fm, from candidate model m map an image, x, to its exported embedding, z=fm(x)∈Rdm. The main Euclidean-distance experiments use these exported vectors without an additional unit-norm transform. When cosine distance is evaluated, we instead use z¯=z/(∥z∥2+ϵ). The target gallery is(1)Gk={(gi,yi,ci)}i=1N, where yi is an enrolled identity, and ci is a camera label. Evaluation queries comprise known queries Qk, whose identities occur in Gk, and unknown queries Qu, whose identities do not. Known and unknown are, therefore, defined relative to the deployed gallery, rather than as fixed semantic classes: changing Gk can change the status of the same identity. For query q,(2)di(m)=∥fm(q)−fm(gi)∥2,Dq(m)={di(m)}i=1N.For a fixed encoder, the query-level problem is to rank valid cross-camera gallery samples for accepted known queries and reject unknown queries. When several pretrained encoders are available, a separate model-level problem is to select one before Qk or Qu is observed, using only auxiliary identities and the enrolled target gallery. This formulation separates open-set inference from deployment-time encoder selection while giving both stages the same gallery-distance interface. For unit-normalized vectors, squared Euclidean distance and cosine distance satisfy ∥z¯q−z¯i∥22=2(1−z¯q⊤z¯i) and, therefore, have identical rankings and threshold-free metrics. Nearest-Euclidean distance and SDSE operate on the exported vectors, whereas the cosine baseline normalizes them internally; the two baselines, therefore, test different distance geometries rather than duplicate one another. The distance vector defines the empirical cumulative distribution function(3)F^q(m)(t)=1N∑i=1NIdi(m)≤t.Nearest distance retains only the first order statistic. SDSE uses finite projections that describe the lower tail, location, spread, upper tail, and candidate-conditioned structure. 3.2. Statistical Rationale: Gallery Presence as a Distributional Test The open-set decision can be written as a gallery-presence test,(4)H0:yq∈YG,H1:yq∉YG, where YG is the set of enrolled identities. Let Fm,qsame and Fm,qdiff denote the query-conditioned distance distributions for same-identity and different-identity gallery samples. If a known query of identity y has ny gallery samples, its empirical distance distribution can be viewed approximately as(5)F^m,qk(t)≈πyFm,qsame(t)+(1−πy)Fm,qdiff(t),πy=nyN, whereas an unknown query has no same-identity component,(6)F^m,qu(t)≈Fm,qdiff(t).This is a task-level statistical rationale rather than an independent-and-identically-distributed generative assumption: all distances share the query, and camera, pose, and gallery composition create dependence. The important structural difference is the possible presence of a coherent same-identity component under H0 and its absence under H1. The nearest distance is a noisy extreme of this empirical distribution. It can be dominated by one accidental impostor match and tends to change with gallery size even when the underlying representation is fixed. SDSE, therefore, combines three evidence groups, summarized in Table 5. The choice is structural rather than optimal: the mixture view motivates what information may be useful, but it does not prove that these nine projections are sufficient, minimal, or uniformly superior. A regularized logistic evaluator estimates their joint relevance from auxiliary episodes. Source episodes create H0 and H1 examples by withholding identities, allowing the evaluator to learn without target unknowns. Leave-one-out gallery scores provide the accepted-known reference distribution for threshold calibration. The same fixed-dimensional profile also permits equal-budget pseudo episodes to compare candidate encoders after candidate-specific standardization rather than treating raw distance scales as comparable. Equations (5) and (6) do not guarantee separability: representation failure, gallery imbalance, or domain shift can remove the expected difference. The paired experiments and the nested profile-size analysis test whether the proposed evidence is useful in practice. 3.3. Nine-Statistic Gallery Profile Figure 4 maps the distributional rationale in Table 5 to the nine computed quantities. Sort the distances as d(1)≤⋯≤d(N). The profile is(7)ϕm(q)=[d(1),d(2),μd,σd,med(d),d(N),CV(d),Δd,SoftCV(d)], where CV(d)=σd/(μd+ϵ) is the coefficient of variation, and Δd=d(2)−d(1). The first six terms retain absolute location and spread information, CV describes relative dispersion, and the gap indicates how strongly the nearest sample is separated from its closest competitor. Let y^(q) be the identity of the nearest gallery sample, and define(8)D+={di:yi=y^(q)},D−={di:yi≠y^(q)}.The ninth component is(9)SoftCV(d)=CV(D+)−CV(D−)CV(D+)+CV(D−)+ϵ.For nonnegative distances, SoftCV lies in [−1,1] up to the numerical stabilizer. Its sign indicates whether the nearest candidate identity is relatively more or less dispersed than the rest of the gallery. Gallery labels are available because the gallery represents enrolled people; no query identity is used in Equation (9). The profile mixes absolute and relative evidence. The first six components retain the distance scale, whereas CV and SoftCV are scale-invariant when ϵ=0 and approximately invariant for the small stabilizer used in practice. SoftCV adds an interpretable identity-conditioned quantity unavailable from unlabeled global summaries, although Section 4.10 shows that its aggregate incremental effect is small. The implementation follows the nine-statistic specification introduced in our preprint; nine is not assumed to be a uniquely necessary dimension. Across candidate encoders, ϕm(q) acts as a deterministic adapter from a model-specific feature space to a common evidence space: no trainable projection, classifier head, or gradient update is attached to fm. Computing the distances costs O(Ndm) per query. The current implementation sorts them in O(NlogN) time; selection-based order statistics could reduce profile extraction to O(N). 3.4. Source-Episodic Evaluator Learning For target dataset t, the remaining datasets form the auxiliary source set S−t. Identity-level episodes reproduce the hypotheses in Equation (4): an episode contains a support gallery, known queries drawn from its identities, and pseudo-unknown queries drawn from disjoint identities. Profiles and binary labels from all episodes are pooled to fit one regularized logistic evaluator. The target dataset is excluded from fitting, and the resulting evaluator is applied without target-domain adaptation. Let E−t be the source episodes, ue,q∈{0,1} the known/unknown episode label, and ℓbal class-balanced logistic loss. The evaluator parameters are(10)ψm(−t)=argminψ1|E−t|∑e∈E−t1|Qe|∑q∈Qeℓbalue,q,ψ(ϕm(q))+λ∥ψ∥22.The target dataset is excluded from this objective. Thus, success cannot be explained by fitting its cameras, identities, or unknown-query prevalence. In the main experiment, each source dataset contributes 80 episodes. Every episode draws 40 support identities and 40 disjoint pseudo-unknown identities, with two support and two query images per selected identity. Known queries share identities with the support gallery; pseudo-unknown queries do not. Equation (7) converts both groups to fixed-dimensional vectors. A standardizer and class-balanced logistic regressor estimate(11)sm(q)=Pψm(−t)(unknown∣ϕm(q)).The auxiliary episodes train ψm(−t) but never update fm. Because support and pseudo-unknown identities are disjoint within every episode, the labels correspond directly to the two hypotheses in Equation (4) rather than to memorized person identities. Each repetition resamples identities and images without replacement within the episode; identities may reappear in later episodes, so repetitions are independently resampled rather than globally disjoint. A deterministic seed makes the sequence reproducible. The fitted component is limited to the standardized logistic evaluator. 3.5. Target-Known-Only Calibration Each enrolled gallery image is temporarily treated as a query against all remaining gallery images, producing leave-one-out scores {sm,iLOO}. The default rejection threshold is(12)τm=Percentile95{sm,iLOO}i=1N.A query is rejected when sm(q)>τm; otherwise, the original embedding-distance ranking is returned. Under approximate exchangeability between future known queries and the leave-one-out gallery scores, the 95th percentile targets an empirical known rejection rate near 5%. It is a calibration rule, not a distribution-free coverage guarantee, because repeated identities and camera effects create dependence. Target unknown labels are not used to fit the evaluator or set the threshold. AUROC and AUPR evaluate score ranking. OSCR combines unknown rejection with correct Rank-1 identification, where correctness uses the same cross-camera filtering as official ReID Rank-1 and mean average precision (mAP). The false-positive rate at 95% true-positive rate (FPR@95) is lower-is-better. 3.6. Source-Selected Unknown-Candidate Gate UCG is asymmetric: it re-examines queries accepted by SDSE and may reject those with high auxiliary risk. It cannot recover a known query that SDSE has already rejected. Its first cue is the gallery-distance energy:(13)Em(q)=−log∑i=1Nexp[−di(m)(q)].This retrieval-space analogue of energy scoring requires no classifier logits. The second cue is the diagonal Mahalanobis deviation of the query embedding from the enrolled gallery:(14)Mm(q)=∑j=1dm(zj−μj)2vj+ϵ1/2, where μj and vj are the gallery mean and variance in embedding dimension j. The diagonal form remains defined when the embedding dimension exceeds the number of enrolled samples and avoids an unstable covariance inverse. Let CE,m and CM,m denote the corresponding known-gallery reference scores; self-distance is excluded when computing gallery energy. Each cue is placed on the scale of the gallery used in that split: a source gallery during weight search and the enrolled target gallery during deployment. The normalization is(15)r˜m(q)=rm(q)−minCr,mmaxCr,m−minCr,m+ϵ,r∈{E,M}.For α∈A={0,0.1,…,1}, the auxiliary risk is(16)hm,α(q)=αE˜m(q)+(1−α)M˜m(q).The two labeled auxiliary datasets are each partitioned into enrolled and pseudo-unknown identities under the same disjoint query/gallery protocol used for evaluation. For every source proxy split, r, and candidate weight, α, the source-side threshold is computed only from its enrolled reference scores,(17)ηm,r,αsrc=Percentile95{hm,α(gi):gi∈Grk}, and the corresponding known acceptance and pseudo-unknown rejection are evaluated using this threshold. Their means across the two source domains define A¯ksrc and R¯usrc. The energy weight is then selected without target queries:(18)αm*=argmaxα∈AR¯usrc(α)s.t.A¯ksrc(α)≥0.95,Ties favor higher known acceptance and then higher candidate precision. If no weight satisfies the 0.95 constraint, the same lexicographic criterion is applied to all weights. After αm* is frozen, the target threshold is(19)ηm=Percentile95{hm,αm*(gi):gi∈Gk}.With bm(q)=I[sm(q)>τm], the final decision is(20)bmUCG(q)=bm(q)∨I[hm,αm*(q)>ηm].UCG changes only the binary operating decision. It is not applied when computing AUROC, AUPR, FPR@95, or OSCR, and it does not change the ReID ranking. Auxiliary labels are used to select the weight, while target calibration uses only the enrolled gallery. The Mahalanobis branch serves as an ablation of the two-cue hypothesis. When source-side selection returns α=1, that branch is inactive and can be omitted at inference. 3.7. Exploratory Frozen-Encoder Selection from the Enrolled Gallery Assume a bank, M={m1,…,mK}, of already trained encoders. The nine-dimensional profile provides the same gallery-evidence interface for every candidate, while candidate-specific standardization handles distance-scale differences. This interface defines a possible selection protocol; its empirical validity must be established for the composition of the candidate bank and is not implied by dimensional compatibility alone. Selection compares proxy AUROC or OSCR, not raw distances, and must not inspect Qk or Qu. For every candidate and target split, SDSE is fitted from the same auxiliary datasets using 40 fixed-budget source episodes with 30 support and 30 pseudo-unknown identities. The enrolled target gallery then generates 20 proxy episodes:Sample at most 40 pseudo-known identities with cross-camera observations and at most 40 disjoint withheld identities.Use two images per pseudo-known identity as support and one different-camera image as a known query.Use two images per withheld identity as pseudo-unknown queries.Average proxy AUROC and proxy OSCR across episodes. Sample at most 40 pseudo-known identities with cross-camera observations and at most 40 disjoint withheld identities. Sample at most 40 pseudo-known identities with cross-camera observations and at most 40 disjoint withheld identities. Use two images per pseudo-known identity as support and one different-camera image as a known query. Use two images per pseudo-known identity as support and one different-camera image as a known query. Use two images per withheld identity as pseudo-unknown queries. Use two images per withheld identity as pseudo-unknown queries. Average proxy AUROC and proxy OSCR across episodes. Average proxy AUROC and proxy OSCR across episodes. The unknown-detection selector is(21)mAUC*=argmaxm∈MAUROC^G(m), and the open-set-identification selector replaces the criterion with OSCR^G(m). All candidates receive the same identity and episode budgets. Only after selection are official target queries evaluated. Selection regret for metric J is(22)RJ=maxm∈MJ(m)−J(mJ*).Target-gallery identities are resampled independently across proxy repetitions, with pseudo-known and withheld pseudo-unknown sets disjoint within each episode but not forced to be disjoint across episodes. The seed–episode pair determines each split. The same identity and image indices are reused for every candidate in a dataset–seed trial, so model differences cannot be attributed to different proxy samples. The procedure selects among existing frozen models; training the candidate bank is outside its scope. Task-aligned resampling supplies a testable proxy: if an encoder preserves cross-camera identity structure in the enrolled gallery, repeated withheld-identity episodes should predict its behavior on unseen queries. Correlation, top-1 agreement, and regret measure that hypothesis within the tested candidate bank. 3.8. End-to-End Procedure The complete algorithm follows five stages, summarized visually in Figure 2 and Figure 3. Prepare a gallery interface. Freeze each candidate encoder, compute query-to-gallery distances, and map every distance vector to the profile in Equation (7). Prepare a gallery interface. Freeze each candidate encoder, compute query-to-gallery distances, and map every distance vector to the profile in Equation (7). Learn across auxiliary identity tasks. For target dataset t, fit ψm(−t) with Equation (10); independently select the optional UCG weight with Equation (18). Learn across auxiliary identity tasks. For target dataset t, fit ψm(−t) with Equation (10); independently select the optional UCG weight with Equation (18). Select a candidate when needed. Generate equal-budget pseudo-known and pseudo-unknown episodes from the enrolled target gallery and minimize the proxy regret in Equation (22). With a single encoder, this stage is skipped. Select a candidate when needed. Generate equal-budget pseudo-known and pseudo-unknown episodes from the enrolled target gallery and minimize the proxy regret in Equation (22). With a single encoder, this stage is skipped. Calibrate from enrolled target data. Use leave-one-out gallery scores to set τm in Equation (12) and, if UCG is enabled, set ηm in Equation (19). Calibrate from enrolled target data. Use leave-one-out gallery scores to set τm in Equation (12) and, if UCG is enabled, set ηm in Equation (19). Infer without adaptation. Compute sm(q) for a new query, apply Equation (20) at the selected operating point, and return the original cross-camera distance ranking only when the query is accepted. Infer without adaptation. Compute sm(q) for a new query, apply Equation (20) at the selected operating point, and return the original cross-camera distance ranking only when the query is accepted. This factorization is central to the design: source episodes learn a transferable presence rule, target-known data calibrate its operating point, and official target queries provide the final independent evaluation. No stage requires target unknown labels or a gradient update to the ReID encoder. 4. Experiments 4.1. Datasets, Embeddings, and Protocol We evaluate Market-1501, DukeMTMC-reID, and CUHK03 with their official disjoint query/gallery partitions [23,24,25]. The partitions encode six, eight, and two cameras, respectively. For each dataset, 50% of test identities are sampled as enrolled; they retain their gallery images and contribute known queries. The remaining identities contribute unknown queries only. The primary seven-method comparison uses seeds 1–5, whereas the confirmatory SDSE–nearest comparison uses seeds 1–20. Within every paired comparison, all scorers share the encoder, gallery, queries, and identity split. Table 6 reports the resulting gallery and query composition. CUHK03 uses its labeled new-protocol query/gallery split. Exported feature archives retain the image path, identity, camera, and split flag for each row; the public scripts verify this alignment before producing qualitative retrieval results. The main comparison uses a 512-dimensional OSNet-x1.0 initialized from ImageNet and trained for 60 epochs on Market-1501. Training images are resized to 256×128 and randomly flipped horizontally. The batch contains 64 images, eight per identity. Adam uses a learning rate of 3×10−5 and weight decay 5×10−4. Label-smoothed identity classification and triplet loss receive equal weight, with triplet margin 0.3. The encoder uses training seed 1; the open-set identity splits use seeds 1–5 for the primary comparison and 1–20 for confirmation. Euclidean scorers operate on exported pre-classifier vectors without additional L2 normalization. Market-1501 is, therefore, the in-domain representation reference, while DukeMTMC-reID and CUHK03 test transfer to different camera systems. Rank-1 and mAP follow the official cross-camera filtering rules, and OSCR uses exactly the same filtered Rank-1 correctness. Known queries without a valid cross-camera positive are excluded from the retrieval and OSCR denominators. Unknown scores are computed for every query because a deployed system cannot use query identity to decide whether a camera filter should apply. This common protocol keeps unknown detection, known retrieval, and combined open-set identification distinct but compatible. The baseline set comprises energy, unit-normalized cosine distance (1− maximum cosine similarity), nearest Euclidean distance, mean 5-nearest-neighbor distance, novelty-mode local outlier factor (LOF, k=20), and shrinkage Mahalanobis distance with Ledoit–Wolf covariance. Whenever fitting is required, it uses enrolled gallery embeddings only. Hyperparameters k=5 and k=20 remain fixed across datasets. The SDSE standardizer is fitted on auxiliary profiles; its class-balanced logistic regressor uses an L2 penalty, C=1, L-BFGS optimization, and at most 1000 iterations. For each target and seed, UCG compares the 11 weights in A on the two auxiliary source-domain proxy splits, never on official target queries. Primary results are means and sample standard deviations over five paired seeds. Confirmatory inference first macro-averages the three datasets within each of 20 seeds and then reports the mean paired difference, a 95% percentile bootstrap interval from 20,000 resamples, and a two-sided Wilcoxon signed-rank test. FPR@95 differences are sign-reversed so that every positive effect favors SDSE. This paired design controls identity-partition variability and avoids treating dataset-level measurements from the same split as independent observations. The main SDSE experiment uses 80 source episodes per auxiliary dataset with 40 enrolled and 40 pseudo-unknown identities because it provides the final query-level score. The encoder-selection sweep repeats fitting for six candidates, three targets, and five seeds; it, therefore, uses a fixed 40-episode source budget with 30 enrolled and 30 pseudo-unknown identities, followed by 20 target-gallery proxy episodes. This reduced budget controls the repeated computational cost and is applied identically to every candidate. The selection experiment is a ranking study within the candidate bank, not a replacement estimate of the main 80-episode SDSE result. Source and target-proxy identities are resampled independently across repetitions, with disjoint known/pseudo-unknown sets only within an episode. For a given target proxy episode, the sampled identities and image indices are identical across all candidate encoders. All candidate features are exported from the same test images after the same deterministic preprocessing: resize to 256×128, tensor conversion, and channel normalization by ImageNet mean (0.485,0.456,0.406) and standard deviation (0.229,0.224,0.225). Thus, preprocessing does not vary with the selected encoder. Gallery sensitivity is evaluated separately over seeds 1–5. Starting from the same 50% known split, we test nested 25%, 50%, and 100% subsets of the enrolled identities; one image, at most two images, and all available images per enrolled identity; a balanced one-image gallery and an induced long tail in which alternating identities retain one or all images; and unknown-query proportions of 25%, 50%, and 75%. The evaluator is not refitted for any target-gallery condition. To equalize computation and class support, sensitivity trials use at most 400 known and 400 unknown official queries per dataset–seed; the main and 20-split comparisons continue to use every official query. The experiments test four linked hypotheses. First, a distributional profile should improve unknown-risk ranking when the nearest extreme is poorly calibrated. Second, any gain should persist under controlled gallery perturbations but cannot replace missing same-identity evidence. Third, open-set identification should remain bounded by frozen-encoder retrieval. Fourth, withheld-identity gallery episodes should provide an informative proxy for ranking frozen encoders. The result sections follow this order, keeping score, representation, threshold, and selection effects identifiable. 4.2. Open-Set Scoring Results Table 7 reports the macro average across datasets. We first average the three datasets within each seed and then compute the mean and standard deviation over the five seeds. Relative to nearest distance, SDSE gains 0.0092 AUROC, 0.0087 AUPR, and 0.0038 OSCR, and lowers FPR@95 by 0.0192. The baseline suite isolates three competing explanations. Mean 5-NN distance tests whether averaging the local tail is sufficient; LOF tests local density; shrinkage Mahalanobis scoring tests global embedding deviation. SDSE achieves the highest macro AUROC and AUPR among the evaluated frozen-embedding scorers, indicating that its advantage is not explained by any one of these alternatives. OpenMax and PROSER are methodologically adjacent but require a fixed parametric classifier interface or classifier retraining [13,14]. They are, therefore, compared conceptually in Section 2.2 rather than converted into different algorithms for the replaceable-gallery setting. Table 8 and Figure 5 reveal where the macro gain arises. On Market-1501, the training-domain encoder already gives nearest distance an AUROC of 0.993; SDSE is effectively tied and nearest distance retains a small AUPR advantage. On DukeMTMC-reID and CUHK03, the nearest extreme is less stable under transfer, and the additional distributional evidence raises AUROC, AUPR, and OSCR in every split. CUHK03 FPR@95 decreases from 0.395 to 0.339, while DukeMTMC-reID shows no corresponding FPR@95 improvement. The scientific conclusion is conditional: the profile contributes when domain transfer weakens nearest-distance calibration, not when that statistic is already near the ceiling. 4.3. Qualitative Query–Gallery Analysis Figure 6 connects the score-level results to the retrieval decisions seen by a camera-network operator. The four rows are selected deterministically from seed 1 after all query scores and target-known thresholds have been fixed: the first two maximize the normalized decision-margin change among cases corrected by SDSE, and the last two show the corresponding highest-margin SDSE errors. Ground-truth unknown labels are used only at this final visualization stage to assign a scored query to a case category; they do not enter evaluator learning, threshold calibration, or quantitative model selection. The corrected known case retains valid cross-camera matches at ranks 1–2 even though its absolute nearest distance crosses the baseline threshold; SDSE accepts it because the complete gallery profile remains compatible with an enrolled identity. In the corrected unknown case, a deceptively close impostor passes the nearest-distance threshold, whereas the broader profile triggers rejection. The failure rows delimit the same mechanism: a shifted known profile can still be rejected despite correct ranking, and an unknown person with several gallery look-alikes can resemble a known profile. These examples explain the decision layer but do not replace the paired population-level evidence in Table 7, Table 8 and Table 9. 4.4. Paired Effects and Split Stability Figure 7 presents the primary five-split effect sizes and orients every difference so that positive favors SDSE. Against nearest distance, DukeMTMC-reID gains are +0.0133 AUROC, +0.0149 AUPR, and +0.0086 OSCR. CUHK03 gains are +0.0145 AUROC, +0.0138 AUPR, and +0.0026 OSCR, with FPR@95 lower by 0.0560. Market-1501 differences are negligible in AUROC and OSCR, while its AUPR favors nearest distance by 0.0025. Inferential claims are based on the larger paired analysis in Table 9, not on five observations. Figure 8 shows the AUROC ordering for every primary seed. SDSE exceeds nearest distance and maximum similarity on both transferred datasets in all five splits, so the displayed gain is not driven by a single favorable partition. All four intervals exclude zero, and every paired difference favors SDSE. The magnitude remains important: AUROC and AUPR improve by about one percentage point, whereas OSCR improves by only 0.00365 because a known query must be both accepted and retrieved correctly. Table 10 connects this central effect to the evidence for each additional design component. 4.5. Gallery-Composition Sensitivity Table 11 tests whether the distributional advantage survives changes to the amount and balance of gallery evidence. SDSE retains a positive AUROC and OSCR difference from the nearest distance in every controlled condition. Across small, medium, and large identity galleries, the AUROC gain remains between 0.0098 and 0.0120. Absolute performance, however, depends strongly on gallery evidence: reducing the gallery from all available images to one image per identity lowers mean SDSE AUROC from 0.9182 to 0.7078 and Rank-1 from 0.449 to 0.331. The induced long tail is also harder than the full gallery despite a positive relative gain. These results agree with the mixture rationale: the evaluator can interpret a same-identity component when evidence exists, but it cannot reconstruct missing reference observations. Varying unknown prevalence leaves AUROC and OSCR nearly stable while changing AUPR substantially, as expected from the prevalence dependence of precision–recall. 4.6. Exploratory Target-Unknown-Free Frozen-Encoder Selection The candidate bank contains six frozen representations: an ImageNet-initialized ResNet50-AE feature extractor, an ImageNet-pretrained OSNet-x1.0, and four Market-1501 OSNet-x1.0 checkpoints saved after 10, 30, 60, and 80 epochs. The bank varies initialization and training stage but contains only two architecture families and is dominated by one ReID source domain. It, therefore, tests within-bank feasibility rather than transfer across unrelated architectures, self-supervised objectives, or source domains. No candidate is trained or fine-tuned during selection. Table 12 evaluates the selector defined in Section 3.7. Spearman correlation is computed across the six candidates in each dataset-seed trial and then averaged over seeds. Top-1 hit records exact agreement with the oracle chosen from official target-query performance, while regret measures the cost of a different choice. For AUROC, mean Spearman correlation ranges from 0.943 to 0.977 and mean regret from 0.00027 to 0.00264. For OSCR, the corresponding correlations range from 0.886 to 0.977. Market-1501 has the largest mean OSCR regret because its candidates have nearly identical unknown-detection scores but small differences in retrieval-coupled OSCR. Across the 15 dataset–seed trials, mean regret is 0.00150 for AUROC selection and 0.00549 for OSCR selection. The selector exactly matches the oracle in 53.3% and 40.0% of trials, respectively; low regret despite a top-1 mismatch indicates that several candidates can be practically equivalent. Figure 9 visualizes the proxy–oracle relationship and shows that the gallery episodes preserve the candidate ordering more reliably than exact winner identity. The experiment, therefore, supports task-aligned ranking within this bank. Its four checkpoints from one OSNet training run, however, do not test architecture-general selection. 4.7. Representation Boundary and Evaluator Capacity Unknown scoring cannot compensate for a representation that fails to retrieve enrolled identities. Table 13 reports known-query performance for the 60-epoch OSNet embeddings. Because SDSE changes only the acceptance score, OSCR remains bounded by the Rank-1 accuracy of the underlying representation. At every OSCR threshold, the correctly accepted fraction cannot exceed the Rank-1 fraction of evaluable known queries; integrating over thresholds preserves this upper bound. On CUHK03, the transferred encoder obtains Rank-1 0.066±0.008, and SDSE obtains OSCR 0.064±0.007. The low OSCR is, therefore, a representation ceiling, not a protocol inconsistency: both quantities use the same official cross-camera filtering. CUHK03, therefore, separates the three layers of evidence. AUROC 0.900 shows that the profile orders unknown risk effectively; Rank-1 0.066 shows that the frozen representation rarely retrieves the correct enrolled identity; OSCR 0.064 combines both requirements and approaches the retrieval ceiling. The dataset supports the unknown-person detector but not strong end-to-end identification with this transferred encoder. This distinction also explains why improving the post hoc score cannot repair a weak representation. On seed 1, replacing logistic regression with Extra Trees or gradient boosting yields three-dataset macro AUROC values of 0.899 and 0.909, compared with 0.921 for logistic regression. Their macro OSCR values are 0.427 and 0.431, compared with 0.436. The higher-capacity evaluators provide no advantage in this experiment, supporting the use of a simple, well-regularized decision model. 4.8. Adaptive False-Accept Correction Table 14 evaluates the decision rule from Section 3.6. Adaptive UCG raises macro unknown rejection from 0.380 to 0.472 while retaining 0.946 known acceptance. Queries flagged by the gate have precision 0.865, and the decision changes for only 5.82% of all queries. The benefit is accompanied by the expected cost: some difficult known queries are also rejected. The selected weights identify the operative mechanism. Source-constrained search chooses α=1 in 14 of the 15 dataset–seed trials and α=0.8 once. Adaptive UCG and the fixed energy gate consequently have the same mean unknown rejection and known acceptance, while adaptive UCG has slightly lower gate recall. Source-side search contributes defensively by suppressing an ineffective cue without inspecting target unknowns; it does not improve on fixed energy in these experiments. Because the continuous SDSE score and all model parameters remain unchanged, Table 14 measures a threshold-level intervention rather than a new representation or scoring gain. For the evaluated setting, the parsimonious deployment rule is, therefore, the energy-only gate. The Mahalanobis row functions as a negative control for the proposed two-cue hypothesis, and the branch can be omitted whenever α=1. This result is scientifically useful because it locates the improvement in energy-based threshold control rather than in unsupported cue synergy. 4.9. Profile Diagnostics and Threshold Calibration Figure 10 compares the four open-set metrics, Figure 11 traces the calibrated operating-point trade-off, Figure 12 tests the contribution of each profile component, and Figure 13 compares scorers on fixed embeddings. All four analyses use the same official cross-camera correctness rule as Table 7 and Table 8. The four diagnostics answer different questions about the decision layer. Figure 10 confirms that the main advantage lies in unknown-score separation, whereas the smaller OSCR change reflects the additional requirement of correct known-person retrieval. Figure 11 shows the expected exchange between unknown rejection and known acceptance when only the threshold changes. The leave-one-feature-out results in Figure 12 indicate that the profile is distributed and partly redundant rather than being driven by a single statistic. Finally, Figure 13 keeps the embedding fixed and, therefore, separates score calibration from representation quality: all scorers inherit the same known-query mAP. 4.10. Nested Profile-Size Analysis Table 15 complements leave-one-feature-out deletion by adding evidence groups cumulatively. Local order statistics improve only slightly over nearest distance, and global summaries alone are insufficient. Combining the two raises macro AUROC from 0.9139 to 0.9208; the full profile reaches 0.9209. The result isolates the main principle: local match evidence becomes more informative when interpreted relative to the global gallery distribution. SoftCV remains an interpretable candidate-conditioned diagnostic, but the negligible eight-to-nine-statistic difference shows that it is not required for the aggregate gain. 4.11. Limitations The nine statistics are designed projections of an empirical distribution, not sufficient statistics derived from a complete generative model. Equations (5) and (6) motivate their roles, while the nested analysis shows that the first eight match the full profile in aggregate. Gallery perturbations establish robustness only over the tested range; one or two reference images cause substantial absolute degradation, and arbitrary gallery growth, severe camera imbalance, and temporal change remain open. The evaluator also requires labeled auxiliary identities, even though target unknown labels and encoder adaptation are unnecessary. The numerical baseline comparison is intentionally restricted to post hoc scorers that share the frozen, replaceable-gallery interface; classifier-dependent OpenMax and PROSER answer a different intervention question. Near-ceiling nearest-distance separation on Market-1501 leaves little room for SDSE. UCG reduces to energy in 14 of 15 trials, and the selector is evaluated on only six candidates from two architecture families. Low CUHK03 Rank-1 accuracy constrains OSCR independently of score quality. Finally, the experiments assume a static gallery and do not address camera drift, temporal association, discovery, or continual enrollment. 5. Conclusions This study asked whether a complete query-to-gallery distance distribution can provide stronger identity-presence evidence than its nearest value when the ReID encoder is fixed. SDSE answers that question by combining local order, global distribution, and candidate-conditioned summaries within a source-episodic evaluator. Identity-disjoint auxiliary episodes supply the evaluator’s training examples, while target-known leave-one-out calibration fixes the operating point without target unknown labels. Its meta-validation structure transfers the decision rule across identity tasks and camera domains, while target-known leave-one-out calibration fixes the operating point without target unknown labels. On shared embeddings and paired identity splits, SDSE provides the strongest macro unknown detection among seven post hoc scorers; the 20-split analysis confirms a 0.0095 AUROC gain over nearest distance. The gain concentrates on the transferred DukeMTMC-reID and CUHK03 domains, where the nearest extreme is less reliable, and disappears on the near-ceiling Market-1501 reference. Profile-size and gallery-sensitivity experiments explain both the mechanism and its boundary: local evidence benefits from global context, but no decision layer can replace sparse same-identity references or weak retrieval. UCG further identifies energy as the useful threshold-control cue, while the gallery-only selector demonstrates low-regret ranking within a limited candidate bank. The resulting contribution is a scientifically grounded, lightweight decision layer that extends frozen ReID systems to fixed-gallery open-set operation while preserving their representations and accepted-query rankings. Broader encoder banks, dynamic galleries, temporal drift, and continual enrollment remain the next validation frontier.