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

cs.CVOct 5, 2026

Identity-Conditioned Score Fusion for Open-Set Person Re-Identification

Authors: Manyi Yao, Jurijs Nazarovs, Eunji Chong, Abhishek Sharma, Rohan Sarkar, Yue Guo, Christian R. Shelton, Amit K. Roy-Chowdhury, +1 more

Organizations: University of California, Riverside · Amazon

Abstract

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.

Figures & tables

Figure 1: Query-conditioned vs. Identity-conditioned score fusion. Existing multi-modal gait recognition often applies a global, shared weight to all identities (left). In contrast, our proposed Identity-conditioned approach adaptively tailors weights for each identity profile (right), allowing the system to leverage the most reliable modality on an identity-specific basis. Sliders conceptually illustrate the weight assignment.

Figure 2: Overview of our framework. Top: Our method leverages gallery intra- and inter-identity statistics to derive tailored model reliability weights

w^{i}_{k}

for each identity

i

without auxiliary training. Bottom: When an incoming query is compared against the gallery, its similarity scores with candidate identity

i

across modalities are combined using that candidate’s specific profile, prioritizing reliable modalities for each individual for final open-set verification.

Dataset	Type	# Identities	# Queries	# Gallery exemplars
Dataset	Type	Train / Test	# Queries	Total	Mean / identity	Range / identity
CCVID	Video	75 / 151	834	1,074	7.1	3–12
MEVID	Video	104 / 54	316	1,438	26.6	1–101
LTCC	Image	77 / 75	493	7,050	94.0	4–589

Table 1: Statistics of person re-identification benchmarks. Query and gallery counts refer to video sequences for CCVID ( Gu et al., 2022b ) and MEVID ( Davila et al., 2023 ) , and images for LTCC ( Qian et al., 2020 ) . Identity counts are reported as train/test. The mean and range of gallery exemplar counts are computed across gallery identities in the test split.

Table 2: Performance on CCVID and MEVID. Comb. specifies the model pool shared by each block (

\blacklozenge

AdaFace,

\spadesuit

CAL,

\blacktriangle

BigGait on CCVID and AGRL on MEVID,

\clubsuit

InsightFace/ArcFace, and

\circledast

LoRA-adapted CLIP-ReID); the top block reports individual-model results and the two below it the fusion methods. Best and second-best fusion results are marked within each pool; highlighted rows are ours. Values are in %; FNIR is mean

\pm

standard deviation over ten enrollment trials. R-1: Rank-1 accuracy, TAR: TAR@1%FAR, FNIR: FNIR@1%FPIR.

Figure 3: Identity profile permutation on CCVID. Expert pool: AdaFace, CAL, BigGait, and InsightFace. Gray histograms show performance distributions over

1{,}000

random identity-to-profile permutations while keeping models, gallery, and weight vectors identical. The genuine assignment (solid orange line) strictly outperforms all permutations across all three metrics, as well as equal weighting (dashed line), confirming that gains stem from identity-specific weights rather than arbitrary non-uniform weighting.

Panel of Figure 5 — Figure 5: Robustness to gallery enrichment. Queries are added into the gallery under their predicted identity $\hat{i}$ when the calibrated match score satisfies $S_{q,\hat{i}}>\eta$ . Solid curves and shaded regions denote the mean and spread across 9 arrival sequences, while dashed lines mark the static-gallery baselines. We report: (a) TAR@ $1\%$ FAR and (b) number of admitted queries across enrollment thresholds $\eta$ . Ours (vanilla) consistently outperforms both equal weighting and QME across all thresholds, remaining robust regardless of enrichment aggressiveness.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Separation of match and non-match score distributions. Dataset: CCVID; model pool: AdaFace, BigGait, and CAL. Dashed lines indicate distribution means and the

1\%

FAR threshold. Our identity-conditioned fusion markedly enlarges the margin between genuine matches and impostors by driving match scores rightward, effectively suppressing distribution overlap under strict verification thresholds.

Combination Rule	Rank-1	mAP	TAR@1%FAR	FNIR@1%FPIR ( $\downarrow$ )
Query only	93.77	91.96	82.40	10.38
Identity only	93.53	93.10	85.36	11.23
Product (Default)	94.60	93.99	85.60	9.96
Weighted sum ( $\lambda=0.25$ )	94.00	92.94	84.45	10.27
Weighted sum ( $\lambda=0.50$ )	93.89	93.38	85.94	10.51
Weighted sum ( $\lambda=0.75$ )	93.64	93.31	86.08	10.72

Appendix

Table A1: Ablation on weight combination rules for joint fusion. Dataset: CCVID; Models pool: AdaFace, BigGait, CAL, and zero-shot CLIP-ReID. We compare different strategies for composing query-conditioned weights

w^{q}_{k}

and identity-specific weights

w^{i}_{k}

, keeping each branch’s weights fixed.

\lambda

denotes the interpolation weight for the identity branch. Bold marks the best result among all combination strategies.

	CCVID				MEVID
Approach	R-1 $\uparrow$	mAP $\uparrow$	TAR $\uparrow$	FNIR $\downarrow$	R-1 $\uparrow$	mAP $\uparrow$	TAR $\uparrow$	FNIR $\downarrow$
Equal weights (no identity score)	$92.03$	$90.62$	$80.72$	$15.96$	$62.13$	$36.62$	$40.37$	$58.92$
$d^{\prime}$ (Default)	$\mathbf{94.21}$	$\mathbf{92.62}$	$\mathbf{84.36}$	$10.60$	$\mathbf{64.03}$	$38.10$	$41.75$	$57.57$
Mean genuine similarity	$93.65$	$91.76$	$83.09$	$\mathbf{9.67}$	$61.50$	$36.10$	$40.12$	$59.97$
Genuine $-$ impostor mean	$93.94$	$91.95$	$83.34$	$9.74$	$61.71$	$36.44$	$40.23$	$59.82$
AUC	$92.18$	$90.76$	$82.66$	$16.46$	$63.40$	$37.89$	$41.57$	$57.84$

Appendix

Table A2: Ablation on per-identity separability metrics. Comparison of different gallery-derived separation scores and learned per-identity weighting baselines ( ‡ ), evaluated across four CCVID and three MEVID expert pools. FNIR and TAR are measured at

1\%

FPIR and

1\%

FAR, respectively. Bold indicates the best result per column.

Panel of Figure 11 — Figure A2: Robustness to gallery enrichment with a margin-based admission rule. Queries are enrolled when the gap between the top two candidate scores satisfies $S_{q,\hat{i}}-\max_{i\neq\hat{i}}S_{q,i}>\eta$ . Across all subplots, curves and shaded areas indicate the mean and spread over 9 arrival orders, and dashed lines show static-gallery baselines: (a) TAR@ $1\%$ FAR, (b) pseudo-label precision of admitted queries, and (c) total admitted queries ( $\eta=0$ admits all queries). Ours (vanilla) consistently maintains its performance advantage over equal weighting and QME at every threshold.

Combination	Method	Rank-1	mAP	TAR@1%FAR	FNIR@1%FPIR ( $\downarrow$ )
$\blacklozenge$	AdaFace ( Kim et al., 2022 )	94.0	87.9	75.7	13.0 $\,{\scriptstyle\pm}\,$ 3.5
$\spadesuit$	CAL ( Gu et al., 2022a )	81.4	74.7	66.3	52.8 $\,{\scriptstyle\pm}\,$ 13.3
$\blacktriangle$	BigGait ( Ye et al., 2024 )	76.7	61.0	49.7	71.1 $\,{\scriptstyle\pm}\,$ 6.1
$\clubsuit$	InsightFace/ArcFace ( Deng et al., 2019 )	94.1	88.2	78.9	18.1 $\,{\scriptstyle\pm}\,$ 3.1
⚫	CLIP-ReID ( Li et al., 2023a )	64.1	63.0	67.9	63.1 $\,{\scriptstyle\pm}\,$ 2.5
$\circledast$	LoRA-adapted CLIP-ReID ( Hu et al., 2022 )	85.4	84.3	80.3	42.8 $\,{\scriptstyle\pm}\,$ 7.5

Appendix

Table A4: Extended multi-model performance on CCVID. Combination specifies the model pool shared by each block (

\blacklozenge

AdaFace,

\spadesuit

CAL,

\blacktriangle

BigGait,

\clubsuit

InsightFace/ArcFace ( Deng et al., 2019 ) , and ⚫ zero-shot and

\circledast

LoRA-adapted CLIP-ReID ( Li et al., 2023a ; Hu et al., 2022 ) ). Best and second-best results are marked within each model pool.

Combination	Method	Rank-1	mAP	TAR@1%FAR	FNIR@1%FPIR ( $\downarrow$ )
$\blacklozenge$	AdaFace ( Kim et al., 2022 )	25.0	8.1	5.4	98.8 $\,{\scriptstyle\pm}\,$ 1.2
$\spadesuit$	CAL ( Gu et al., 2022a )	52.5	27.1	34.7	67.8 $\,{\scriptstyle\pm}\,$ 7.3
$\blacktriangle$	AGRL ( Wu et al., 2020 )	51.9	25.5	30.7	69.4 $\,{\scriptstyle\pm}\,$ 8.9
⚫	CLIP-ReID ( Li et al., 2023a )	63.0	33.0	33.6	68.3 $\,{\scriptstyle\pm}\,$ 3.3
$\circledast$	LoRA-adapted CLIP-ReID ( Hu et al., 2022 )	69.9	45.1	47.7	49.7 $\,{\scriptstyle\pm}\,$ 4.3
	Min-Fusion ( Jain et al., 2005 )	46.8	21.2	28.0	70.4 $\,{\scriptstyle\pm}\,$ 8.0

Appendix

Table A5: Extended multi-model evaluation on MEVID. Combination specifies the model pool shared by each block (

\blacklozenge

AdaFace,

\spadesuit

CAL,

\blacktriangle

AGRL, and ⚫ zero-shot and

\circledast

LoRA-adapted CLIP-ReID ( Li et al., 2023a ; Hu et al., 2022 ) ). Best and second-best results are marked within each block.

Combination	Method	Rank-1	mAP	TAR@1%FAR	FNIR@1%FPIR ( $\downarrow$ )
$\blacklozenge$	AdaFace ( Kim et al., 2022 )	18.5	5.9	2.4	99.8 $\,{\scriptstyle\pm}\,$ 0.2
$\spadesuit$	CAL ( Gu et al., 2022a )	74.4	40.6	36.7	59.7 $\,{\scriptstyle\pm}\,$ 7.3
$\blacksquare$	AIM ( Yang et al., 2023 )	74.8	40.9	37.0	66.2 $\,{\scriptstyle\pm}\,$ 7.5
⚫	CLIP-ReID ( Li et al., 2023a )	70.7	33.7	26.5	47.8 $\,{\scriptstyle\pm}\,$ 13.0
$\circledast$	LoRA-adapted CLIP-ReID ( Hu et al., 2022 )	76.7	40.7	31.9	42.0 $\,{\scriptstyle\pm}\,$ 10.3
	Min-Fusion ( Jain et al., 2005 )	37.1	12.9	12.4	81.9 $\,{\scriptstyle\pm}\,$ 6.0

Appendix

Table A6: Extended multi-model performance on LTCC. Combination specifies the model pool shared by each block (

\blacklozenge

AdaFace,

\spadesuit

CAL,

\blacksquare

AIM, and ⚫ zero-shot and

\circledast

LoRA-adapted CLIP-ReID ( Li et al., 2023a ; Hu et al., 2022 ) ). Best and second-best results are marked within each model pool.

Explore similar work

Thinking Before Matching: A Reinforcement Reasoning Paradigm Towards General Person Re-Identification
Apr 21, 2026Quan Zhang, Jingze Wu, Jialong Wang +3Re-Identification Scene Reasoning
Sun Yat-sen University · Alibaba Cloud Computing
Multimodal LLM-Empowered Re-Ranking for Generalizable Person Re-Identification
Jun 15, 2026Jiachen Li, Xiaojin GongRe-Identification Large Language Model Reranking
College of Information Science and Electronic Engineering Zhejiang University
From Global to Local: Rethinking CLIP Feature Aggregation for Person Re-Identification
Apr 24, 2026Aotian Zheng, Winston Sun, Bahaa Alattar +2Re-Identification Contrastive Language-Image Pre-Training Model
Department of Electrical and Computer Engineering, University of Washington, Seattle, WA 98195 · Applied Physics Laboratory, University of Washington, Seattle, WA 98195

Apr 21, 2026, cs.CVJ/K move · Enter open · S save

Thinking Before Matching: A Reinforcement Reasoning Paradigm Towards General Person Re-Identification

Quan Zhang, Jingze Wu, Jialong Wang, Xiaohua Xie+2

Sun Yat-sen University · Alibaba Cloud Computing

Re-Identification Scene Reasoning Fine-Grained Perception Discriminative

Open paper AskPDF arXiv

Learning identity-discriminative representations with multi-scene generality has become a critical objective in person re-identification (ReID). However, mainstream perception-driven paradigms tend to identify fitting from massive annotated data rather than identity-causal cues understanding, which presents a fragile representation against multiple disruptions. In this work, ReID-R is proposed as a novel reasoning-driven paradigm that achieves explicit identity understanding and reasoning by incorporating chain-of-thought into the ReID pipeline. Specifically, ReID-R consists of a two-stage contribution: (i) Discriminative reasoning warm-up, where a model is trained in a CoT label-free manner to acquire identity-aware feature understanding; and (ii) Efficient reinforcement learning, which proposes a non-trivial sampling to construct scene-generalizable data. On this basis, ReID-R leverages high-quality reward signals to guide the model toward focusing on ID-related cues, achieving accurate reasoning and correct responses. Extensive experiments on multiple ReID benchmarks demonstrate that ReID-R achieves competitive identity discrimination as superior methods using only 14.3K non-trivial data (20.9% of the existing data scale). Furthermore, benefit from inherent reasoning, ReID-R can provide high-quality interpretation for results.

Abstract

Explore similar work

Thinking Before Matching: A Reinforcement Reasoning Paradigm Towards General Person Re-Identification

Multimodal LLM-Empowered Re-Ranking for Generalizable Person Re-Identification

From Global to Local: Rethinking CLIP Feature Aggregation for Person Re-Identification