Confidence-Guided Pseudo-Label Refinement for Unsupervised Cross-Domain Person Re-Identification

Sundas Iqbal; Danish Ali; Qing Tian; Yuhui Zheng; Jianping Gou; Weihua Ou

cs.CVOct 19, 2026

Confidence-Guided Pseudo-Label Refinement for Unsupervised Cross-Domain Person Re-Identification

Authors: Sundas Iqbal, Danish Ali, Qing Tian, Yuhui Zheng, Jianping Gou, Weihua Ou

Organizations: School of Software, Nanjing University of Information Science and Technology, Nanjing 210044, China · School of Computer Science, Wuhan University, Wuhan 430072, China · School of Computer Science, Nanjing University of Information Science and Technology, Nanjing 210044, China · Wuxi Institute of Technology, Nanjing University of Information Science and Technology, Wuxi 214000, China ·

Abstract

Unsupervised cross-domain person re-identification remains challenging because domain shifts generate noisy pseudo labels that are reinforced during iterative adaptation. Existing clustering-based methods often rely on limited reliability cues or global confidence criteria and may impose hard identity supervision on ambiguous target samples. To address these limitations, we propose CGPLR, a confidence-guided pseudo-label refinement framework for unsupervised domain-adaptive person re-identification. CGPLR estimates sample-level reliability by jointly modeling centroid compatibility, local neighborhood consistency, separation from the strongest competing cluster and temporal assignment stability. The resulting confidence distribution is used to derive cluster-adaptive upper and lower thresholds for a three-way selective decision. Reliable assignments are preserved, recoverable samples are reassigned only when an alternative cluster receives sufficient geometric and neighborhood support and highly ambiguous samples are deferred from hard identity supervision. The learning objective adapts to these decisions through confidence-weighted supervision for preserved samples, cautious corrected-label weighting for reassigned samples and label-free representation consistency for deferred samples. CGPLR achieves 76.4% Rank-1 accuracy and 47.6% mean average precision (mAP) on Market-1501 to MSMT17, and 95.1% Rank-1 accuracy and 87.7% mAP on MSMT17 to Market-1501, demonstrating effective reliability-aware selective adaptation. The code is available at https://github.com/Sundas058/CGPLR.

Figures & tables

Mechanism	CGC [ 33 ]	3C [ 13 ]	PPLR [ 26 ]	CMLR [ 17 ]	CGPLR (Ours)
Centroid-based reliability	$\checkmark$	$\checkmark$	–	–	$\checkmark$
Local/neighborhood evidence	–	–	$\checkmark$	$\checkmark$	$\checkmark$
Explicit centroid margin	–	–	–	–	$\checkmark$
Temporal stability	–	–	–	–	$\checkmark$
Cluster-adaptive threshold	–	$\checkmark$	–	–	$\checkmark$
Evidence-gated hard reassignment	–	–	–	–	$\checkmark$

TABLE I: Comparison of representative confidence- and pseudo-label-refinement methods for unsupervised person re-identification.

Fig. 1: Overview of the proposed CGPLR framework for unsupervised cross-domain person re-identification. Target pseudo labels are evaluated using structural, competing-cluster and temporal evidence, followed by cluster-adaptive preserve, reassign or defer decisions and reliability-conditioned target-domain learning.

Parameter	Value	Description
Batch size $B$	64	Mini-batch size
Learning rate $\eta$	$3.5\times 10^{-4}$	Initial Adam learning rate
Adaptation rounds $T_{t}$	50	Target adaptation rounds
Triplet margin $m$	0.3	Metric-learning margin
Neighborhood size $K$	5	Number of nearest neighbors
Structural balance $\alpha$	0.5	Balance between $c_{j}$ and $s_{j}$

TABLE II: Optimization and method-specific settings used by CGPLR.

Method	R1	R5	R10	mAP	$\Delta$ R1	$\Delta$ mAP
Source-only baseline	37.2	54.9	62.7	23.1	$-23.7$	$-12.5$
Clustering baseline	60.9	77.8	83.6	35.6	$0.0$	$0.0$
Full CGPLR	67.3	83.4	88.8	40.7	$+6.4$	$+5.1$

TABLE IV: Controlled additional UDA evaluation on Market-1501

\rightarrow

CUHK03-NP-D. Rank-

k

accuracy and mAP are reported (%).

\Delta

R1 and

\Delta

mAP denote performance differences relative to the clustering baseline.

Market-1501 $\rightarrow$ MSMT17
Setting	PL	CC	NC	SR	RSL	R1	mAP
Source-only baseline	–	–	–	–	–	38.6	15.4
Clustering baseline	✓	–	–	–	–	64.8	35.2
+ Centroid compatibility	✓	✓	–	–	–	68.7	38.4
+ Neighborhood consistency	✓	✓	✓	–	–	71.3	41.2
+ Selective refinement	✓	✓	✓	✓	–	74.1	44.6

TABLE V: Stage-wise ablation of CGPLR. PL, CC, NC, SR, and RSL denote pseudo labeling, centroid compatibility, neighborhood consistency, selective refinement, and reliability-aware selective learning, respectively. R1 and mAP are reported (%).

Market-1501 $\rightarrow$ MSMT17
Variant	R1	mAP	$\Delta$ mAP vs. full
Without competing-centroid margin	74.9	45.7	$-1.9$
Without temporal stability	75.3	46.2	$-1.4$
Global thresholds instead of cluster-adaptive thresholds	74.6	45.2	$-2.4$
Forced reassignment without deferral	73.8	44.1	$-3.5$
Discard deferred samples	75.0	45.8	$-1.8$

TABLE VI: Direct ablation of the individual mechanisms introduced in CGPLR. Rank-1 accuracy (R1) and mAP are reported (%).

\Delta

mAP denotes the difference relative to the full CGPLR model under the corresponding transfer setting. Therefore, the full model has

\Delta

mAP

=0.0

Metric	Clustering baseline	Multi-evidence reliability	+ Cluster-adaptive refinement	Full CGPLR
Preserved precision	89.4	96.2	97.0	97.8
Reassignment precision	–	–	84.7	87.3
Correction rate	–	–	67.5	72.6
Preserved rate	100.0	81.4	78.2	79.1
Reassignment rate	0.0	0.0	9.8	10.5
Hard-label coverage	100.0	81.4	88.0	89.6

TABLE VII: Decision-level pseudo-label analysis on MSMT17

\rightarrow

Market-1501. Target labels are used only for post-hoc diagnosis; statistics are averaged over the final five adaptation rounds (%).

Parameter	Tested value	R1	mAP	$\Delta$ mAP
Temporal window $H$	1	94.4	86.8	$-0.9$
	2	94.8	87.3	$-0.4$
	3	95.1	87.7	0.0
	4	94.9	87.4	$-0.3$
Threshold factors $(\rho_{H},\rho_{L})$	$(0.8,0.8)$	94.6	87.0	$-0.7$
	$(0.9,0.9)$	94.9	87.4	$-0.3$

TABLE VIII: Sensitivity of the revised CGPLR parameters on MSMT17

\rightarrow

Market-1501. One parameter is varied at a time, while all remaining parameters are fixed at their default values.

\Delta

mAP denotes the difference from the selected configuration.

Fig. 2: Qualitative retrieval comparison between the clustering baseline and CGPLR under challenging target-domain appearance variations. Green and red boxes denote correct and incorrect matches, respectively.

Fig. 3: Distributions of the competing-centroid distance margin under different cross-domain adaptation settings.

Method	Adaptation time (h)	Time ratio ( $\times$ )	Peak memory (GB)	FPS
Clustering baseline	4.82	1.00	6.3	302
+ Multi-evidence reliability	5.35	1.11	6.7	302
+ Selective refinement	5.61	1.16	6.9	301
Full CGPLR	6.07	1.26	7.1	300

TABLE IX: Computational comparison of the clustering baseline and CGPLR on MSMT17

\rightarrow

Market-1501. Adaptation time and peak memory are reported for 50 adaptation rounds using one NVIDIA RTX 4090.

Person Re-Identification

Person re-identification matches the same individual across camera views, times, clothing, and sensing modalities despite changes in appearance.

See all 92 papers

AIDA-ReID: Adaptive Intermediate Domain Adaptation for Generalizable and Source-Free Person Re-Identification
Apr 30, 2026Sundas Iqbal, Qing Tian, Danish Ali +2Multi-Source Domain Adaptation Source-Free Domain Adaptation
School of Software, Nanjing University of Information Science and Technology, Nanjing 210044, China · Wuxi Institute of Technology, Nanjing University of Information Science and Technology, Wuxi 214000, China · School of Computer Science, Wuhan University, Wuhan 430070, China +2
Multimodal LLM-Empowered Re-Ranking for Generalizable Person Re-Identification
Jun 15, 2026Jiachen Li, Xiaojin GongMultimodal Reranking Multimodal Large Language Models
College of Information Science and Electronic Engineering Zhejiang University
From Calibration to Refinement: Seeking Certainty via Probabilistic Evidence Propagation for Noisy-Label Person Re-Identification
Feb 26, 2026Xin Yuan, Zhiyong Zhang, Xin Xu +2Person Re-Identification Learning with Noisy Labels
School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan, Hubei 430065, China · Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System, Wuhan University of Science and Technology, Wuhan 430065, China · School of Computer Science, Wuhan University, Wuhan, Hubei 430072, China +1

Apr 30, 2026J/K move · Enter open · S save

AIDA-ReID: Adaptive Intermediate Domain Adaptation for Generalizable and Source-Free Person Re-Identification

Sundas Iqbal, Qing Tian, Danish Ali, Jianping Gou+1

School of Software, Nanjing University of Information Science and Technology, Nanjing 210044, China · Wuxi Institute of Technology, Nanjing University of Information Science and Technology, Wuxi 214000, China · School of Computer Science, Wuhan University, Wuhan 430070, China+2

Multi-Source Domain Adaptation Source-Free Domain Adaptation Domain Adaptation Cross-Domain Generalization Person Re-Identification

Open paper AskPDF arXiv

Person re-identification (Re-ID) aims to match images of the same individual across non-overlapping camera views and remains challenging due to domain shifts caused by variations in illumination, background, camera characteristics, and population distributions. Although supervised models perform well under matched training and testing conditions, their performance degrades significantly when deployed in unseen environments. Existing intermediate domain approaches such as IDM and IDM++ alleviate this gap by constructing bridge feature distributions between domains; however, they rely on fixed mixing strategies and joint source-target access, limiting their applicability to multi-source and source-free settings. To address these limitations, this paper proposes Adaptive Intermediate Domain Adaptation (AIDA), also referred to as Source-Free Multi-Source Intermediate Domain Adaptation (SF-MIDA). The proposed framework treats intermediate-domain learning as a dynamically regulated process, where feature mixing and regularization strength are adaptively controlled using feedback signals derived from model uncertainty and training stability. A multi-source intermediate domain generator synthesizes diverse intermediate representations, while a pseudo-mirror regularization strategy preserves identity consistency under domain perturbations. Extensive experiments across domain generalization and source-free settings demonstrate the effectiveness of the proposed framework.

Abstract

Person Re-Identification

Explore similar work

AIDA-ReID: Adaptive Intermediate Domain Adaptation for Generalizable and Source-Free Person Re-Identification

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

From Calibration to Refinement: Seeking Certainty via Probabilistic Evidence Propagation for Noisy-Label Person Re-Identification