Source-linked AI summary
Identity-Conditioned Score Fusion for Open-Set Person Re-Identification
Manyi Yao, Jurijs Nazarovs, Eunji Chong, Abhishek Sharma, Rohan Sarkar, Yue Guo, Christian R. Shelton, Amit K. Roy-Chowdhury, Debashish Pal
TL;DR
Person ReID fusion must account for both variation in query quality and differences in model discriminability across enrolled identities. This paper estimates identity-specific model weights from gallery comparisons and combines them with query-conditioned fusion. Across benchmarks, it reports consistent gains over competing fusion approaches, with a stated limitation at ultra-low false-acceptance regimes.
Problem
Model usefulness varies across identities, while shared weights cannot express identity-specific discriminability; open-set ReID also requires separating true matches from impostors.
Source: §1 · p. 1, §3.2 · p. 5Method
Identity-conditioned score fusion estimates each gallery identity’s model-weight profile from gallery comparisons and composes with query-conditioned fusion through a parameter-free rule.
Source: §1 · pp. 1–2Results
Across three benchmarks, the method consistently outperforms fusion baselines; on CCVID, identity-conditioned-only fusion raises Rank-1 accuracy from 94.0% (AdaFace) to 95.3%.
Source: §1 · p. 2, §4.1 · p. 7Takeaways & Limitations
Across multiple model pools and clothes-changing benchmarks, identity-specific weighting outperforms uniform and quality-based baselines and integrates with query-dependent fusion.
Source: §5 · p. 9Open-set verification at ultra-low false-acceptance regimes remains constrained by limited non-mate query scale in existing clothes-changing benchmarks.
Source: §5 · p. 10
Abstract
from arXiv · showhide
Robust person re-identification often combines complementary cues such as face, gait, and body shape. While adaptive fusion typically targets query quality, model strength also varies across identities. We introduce identity-conditioned score fusion, a framework that tailors weights to each gallery identity without training. By contrasting intra-identity consistency against cross-identity impostors, it extracts identity-specific profiles that couple with query-conditioned adaptation via a parameter-free rule. This widens the separation between true and false matches while preserving score calibration. Evaluations on three clothes-changing person re-identification benchmarks show that our method consistently outperforms statistical, rank-based, and learned baselines, achieving up to an 8.8% absolute reduction in the false non-identification rate and demonstrating the value of identity-conditioned fusion in open-set person re-identification.
1 INTRODUCTION
Person ReID combines complementary cues, but their usefulness varies with both query conditions and the identity being matched. Identity-conditioned score fusion estimates gallery-specific model weights and can compose with query-conditioned fusion; evaluations report gains over competing fusion baselines.
1 INTRODUCTION
- A model’s discriminative value can differ by gallery identity, which shared model weights cannot express.The gallery provides each identity’s exemplars and comparisons with other enrolled identities for estimating these differences.Source: §1 · p. 1
- Identity-conditioned score fusion estimates a model-weight profile for each enrolled identity from within-identity consistency and cross-identity impostor comparisons.At query time, scores are combined using the profile of the gallery identity being considered, without updating recognition models or training an additional fusion network.Source: §1 · p. 1
- Experiments report gains both from identity-conditioned fusion alone and from its composition with query-dependent methods, and examine exemplar-count effects.The analysis also finds that different identities benefit from different models and that gallery comparisons help identify those preferences.Source: §1 · p. 2
- The method composes with query-conditioned fusion through a parameter-free rule, combining identity-specific weights with adaptation to the current observation.Query-conditioned methods address variation in observed cue quality, such as a poorly visible face or degraded gait sequence.Source: §1 · pp. 1–2
- Across three benchmarks, the method outperforms statistical, rank-based, and learned fusion baselines, widening the genuine–impostor separation margin.Experiments also report gains when identity-conditioned fusion is composed with query-dependent methods.Source: §1 · p. 2
2 RELATED WORK
Related work addresses open-set rejection, query-dependent fusion, and identity-dependent biometric performance. This framework combines gallery-based identity profiles with fusion for query conditions, while regularizing identity-specific weights to mitigate transfer errors.
2 RELATED WORK
- Open-set person ReID must match enrolled identities and reject queries from identities absent from the gallery; reliable rejection requires separating true matches from high-scoring non-matches.Prior work has explored clothing-robust representations, body shape and gait cues, and training losses that account for impostor scores as well as correct matches.Source: §2 · p. 2
- Quality-dependent fusion adjusts modality contributions according to observed input reliability, while query-conditioned methods can select or weight models for each query.Examples include QME’s query-quality-based learned fusion, FusionAgent’s query-level model selection, and HAMoBE’s pair-conditioned feature weights.Source: §2 · p. 2
- The framework estimates a mixture profile for each gallery identity from gallery comparisons, then uses profiles for score fusion and open-set decisions.It also describes composition of these profiles with query-dependent fusion.Source: Figure 2 · p. 3, §2 · p. 3
- Earlier biometric fusion used client-specific separation statistics, but the paper notes that camera and viewpoint changes can undermine enrollment-to-test transfer in person ReID.Its regularized weighting strategy is designed to mitigate these transfer errors and integrate model consensus into identity scoring.Source: §2 · p. 3
3 PROPOSED METHOD
The method assigns each gallery identity a profile of model weights, then uses those weights to fuse query scores and support open-set decisions. It can also combine identity-specific separability with query-dependent fusion.
3.1 GALLERY-DERIVED MIXTURE PROFILES
- For each identity and model, d′ measures separation between genuine and impostor similarities by scaling their mean difference by pooled standard deviation.Genuine scores compare distinct exemplars of the identity; impostor scores compare its exemplars with the rest of the gallery.Source: §3.1 · p. 4
- The identity-specific profile interpolates softmax weights over separation scores with uniform weights to hedge against noisy estimates from limited gallery observations.The profile is computed in closed form from gallery comparisons, without training a fusion network.Source: §3.1 · p. 4
3.2 IDENTITY-CONDITIONED SCORE FUSION
- Each candidate’s profile weights its models’ standardized query scores, which are fused per exemplar and max-pooled across that identity’s exemplars.The model-level normalization parameters are identity-agnostic and computed once per model.Source: §3.2 · pp. 4–5
- Because candidate identities can elicit different non-match score baselines, the method calibrates fused scores against candidate-specific impostor statistics before applying a shared open-set threshold.It estimates impostor tail statistics from scores above the 90th percentile and regularizes them toward pooled gallery statistics.Source: §3.2 · p. 5
3.3 COMPOSITION WITH QUERY-CONDITIONED FUSION
- Composition multiplies identity-specific weights by query-dependent modality weights, giving greater contribution to models that are both distinctive for the candidate and reliable for the query.The composite score replaces the identity-conditioned score in the calibration pipeline.Source: §3.3 · pp. 5–6
4 EXPERIMENTS
Across three clothes-changing benchmarks, the experiments find that identity-conditioned fusion improves recognition, while calibration, identity-specific profiles, exemplar availability, and gallery updates are examined as factors in its performance.
4 EXPERIMENTS
- Experiments use three clothes-changing benchmarks, fixed heterogeneous face, body, and gait backbones, and standard ranking, verification, and open-set identification metrics.Baselines include training-free fusion rules and methods optimized on the training split.Source: §4 · p. 6
4.1 RESULTS
- Across three benchmarks, Ours (vanilla) beats strong single-model and supervised fusion baselines without fusion training, while query-conditioned composition further improves performance.On CCVID, Rank-1 rises from 94.0% (AdaFace) to 95.3%; on MEVID, Ours (FusionAgent) raises mAP from 28.7% to 31.2% and TAR from 34.9% to 37.3%; on LTCC, FNIR@1%FPIR falls from 60.7% to 29.1%.Source: Table 2 · p. 7, §4.1 · p. 7, Table 3 · p. 8
4.2 ABLATION STUDIES
- Candidate calibration lowers FNIR@1%FPIR across configurations by up to 8.1% and reduces CCVID FNIR from 19.1% to 11.0%.Calibration stabilizes global thresholding across gallery identities, while identity weighting provides the primary discriminative power.Source: §4.2 · pp. 7–8, Table 4 · p. 8
4.3 ANALYSIS
- On CCVID, genuine identity-to-profile assignments beat equal weighting and all 1,000 random permutations across three metrics, indicating that weights encode identity-specific priors.The permutation test keeps the models and gallery data identical while shuffling weight-to-identity assignments.Source: §4.3 · pp. 8–9, Figure 3 · p. 8
- On MEVID, Ours (vanilla) leads equal weighting and QME from 2 to 20 exemplars per identity, retaining its advantage with sparse enrollment.Performance scales with more exemplars for all methods, but Ours (vanilla) maintains its advantage across exemplar counts.Source: Figure 4 · p. 8, §4.3 · p. 9
- On CCVID, Ours (vanilla) stays ahead of equal weighting and QME across enrollment thresholds and nine arrival orders, including low thresholds with potential misidentifications.All methods benefit at low-to-moderate thresholds (η ≤ 4) before converging to static-gallery baselines as thresholds rise and fewer queries are admitted.Source: §4.3 · p. 9, Figure 5 · p. 9
5 CONCLUSION
The conclusion reports that gallery-derived identity-specific reliability priors improve multi-model open-set identification across the evaluated settings. It also identifies unmodeled cross-model dependencies and limited non-mate query scale at ultra-low false-acceptance regimes as limitations.
5 CONCLUSION
- Across multiple model pools and clothes-changing benchmarks, identity-specific weighting outperforms uniform and quality-based baselines and remains robust under label noise, data scarcity, and dynamic gallery evolution.The method also integrates with query-dependent fusion.Source: §5 · p. 9
- The method estimates each model’s reliability profile individually and does not explicitly model cross-model feature correlations or statistical redundancies.Capturing joint multi-model dependencies is identified as a direction for further refinement.Source: §5 · p. 10
- Open-set verification at ultra-low false-acceptance regimes remains constrained by the limited non-mate query scale in existing clothes-changing benchmarks.Source: §5 · p. 10
A.1 RECOGNITION MODELS AND SCORE PROCESSING
The appendix describes benchmark-specific model configurations and how embeddings or released score matrices supply query-gallery and gallery-gallery similarities.
A.1 RECOGNITION MODELS AND SCORE PROCESSING
- Evaluated backbones span face, gait, and CLIP-ReID models, with configurations varying by benchmark and CLIP-ReID tested with or without LoRA adaptation.Source: §A.1 · p. 15
- Video tracklet frame embeddings are mean-pooled except BigGait and AGRL, which provide sequence embeddings; face embeddings use detected and aligned frames.Source: §A.1 · p. 15
- CCVID and MEVID use cosine similarities between unit-normalized embeddings, whereas LTCC uses released QME query-gallery scores for AdaFace, CAL, and AIM.LTCC CLIP-ReID and all gallery-gallery scores are computed from embeddings using cosine similarity.Source: §A.1 · p. 15
A.2 ADDITIONAL RESULTS
Additional results show that identity-conditioned fusion outperforms static fusion across model pools and that its combination with query-conditioned weights performs best with a multiplicative rule. Ablations and sensitivity analyses examine separability metrics, gallery admission, score distributions, and hyperparameters.
Score distribution analysis
- On CCVID, identity-conditioned fusion improves distributional separability over equal weighting, elevating genuine-match confidence and reducing rejected true matches at a strict 1% FAR threshold.The analysis is shown in Figure A1.Source: §A.2 · p. 15
Weight combination strategies
- Multiplicative composition achieves Rank-1 (94.60%), mAP (93.99%), and open-set FNIR (9.96%), outperforming additive and geometric-mean rules across mixing weights λ.RRF's rank-transformed scores disrupt per-identity score calibration and yield open-set FNIR (68.09%).Source: Table A1 · p. 16, §A.2 · p. 16
Alternative admission rule for gallery enrichment
- 86.1% mean TAR for Ours (vanilla) exceeds either baseline's 85.5% peak under margin admission; at η = 0, label accuracy is 95.8% versus ≤94.0% for baselines.Figure A2 reports trends across nine arrival orders; as η increases, Ours consistently admits more clean exemplars than the baselines under the same threshold.Source: Figure A2 · p. 17, §A.2 · p. 17
Sensitivity to hyperparameters
- Intermediate α values improve over equal weights, peaking around α = 0.5; recognition remains stable for 0.25 ≤ τ ≤ 2, with τ = 0.5 a solid trade-off.Discarding the uniform prior (α = 1) degrades verification; anchoring weights to that prior prevents overfitting to sparse gallery samples.Source: Table A3 · p. 17, §A.2 · pp. 17–18
Choice of per-identity separability metric
- Across both benchmarks, d′ yields the most reliable overall performance, making score spread a stronger basis for per-identity weight estimation than the evaluated alternatives.Alternatives include genuine similarity, genuine–impostor mean gap, AUC, 1−EER, LDA, and logistic regression.Source: §A.2 · pp. 16–17, Table A2 · p. 16
Extended evaluations across model pools
- Across CCVID, MEVID, and LTCC model pools, Ours (vanilla) outperforms single-model and classical fusion baselines, while query-adaptive variants achieve top-tier performance and reduce open-set FNIR@1%FPIR over static fusion.The extended evaluations are reported in Tables A4–A6.Source: §A.2 · p. 18, Table A4 · p. 19, Table A5 · p. 20