MINER: Modeling Inter- and Intra-Person Variance for Reidentification

Oron Nir; Eliyahu Strugo; Ariel Shamir

cs.CVOct 20, 2026

MINER: Modeling Inter- and Intra-Person Variance for Reidentification

Authors: Oron Nir, Eliyahu Strugo, Ariel Shamir

Organizations: Reichman University & Microsoft · Microsoft · Reichman University

Abstract

Clothes-Changing Person Re-Identification (CC-ReID) remains challenging due to drastic appearance shifts across outfits. Most prior works devise biometric signal-extraction heuristics to enhance inter-person variance, e.g., using segmentation masks, while leaving intra-person variance uncontrolled. We present MINER, a simple model-agnostic framework that suppresses garment bias by contrasting hard examples and adaptively balancing dataset-dependent intra- and inter-person variance. Extensive experiments show that MINER improves retrieval across CC-ReID benchmarks, achieving state-of-the-art performance on four datasets and ranking among the leaders in all three tracks of the open TVRID RGB-Depth ReID challenge.

Figures & tables

Figure 1: MINER Contrastive Loss over a bipartite sampled batch. Hard-Negative person IDs and Hard-Positive examples per person distill both inter-person (white) and intra-person (blue) variance simultaneously using a block-diagonal label matrix.

Figure 4: The same identity exhibits greater intra-person variance in Celeb-L (top) than in PRCC (bottom), spanning clothing, pose, and scene.

CC-ReID	Backbone	Vanilla	PRCC		LTCC		Celeb-L		DeepChange		VC-Clothes		LaST
Method		Weights	R@1	mAP	R@1	mAP	R@1	mAP	R@1	mAP	R@1	mAP	R@1	mAP
CAL [ 13 ] ‘22	ResNet50	$\checkmark$	55.2	55.8	40.1	18.0	–	–	54.0	19.0	81.4	81.7	73.7	37.2
LaST [ 52 ] ‘22	ResNet50	$\checkmark$	57.7	54.7	–	–	54.4	11.8	–	–	–	–	71.0	28.0
MBUNet [ 71 ] ‘23	ResNet50		68.7	65.2	40.3	15.0	35.5	21.5	–	–	82.7	70.3	–	–
ACID [ 65 ] ‘23	ResNet50	$\checkmark$	55.4	66.1	29.1	14.5	27.9	15.8	–	–	84.3	74.2	–	–
SCNet [ 14 ] ‘23	ResNet50	$\checkmark$	58.7	59.9	47.5	25.5	–	–	53.5	18.7	–	–	–	–

Table 1: Comparison of various methods on CC-ReID: MINER Rank@1 and mAP performance across CC-ReID datasets vs. SOTA. We initialize to both EVA-CLIP and DIFFER separately. All methods use pre-trained weights, Vanilla Weights were trained on generic datasets like ImageNet while others used additional ReID finetuning. VC-Clothes is synthetic. Best result is in bold , the second is underlined .

CC-ReID	Data	Base-reproduced		+MINER
Method	set	R@1	mAP	R@1	mAP
CLIP-ReID	PRCC	17.5	18.0	29.0	30.3
MADE	PRCC	61.7	56.2	71.2	64.7
DIFFER	PRCC	68.5	64.7	73.9	69.7
CLIP-ReID	LTCC	17.3	8.3	32.7	11.2
MADE	LTCC	44.0	19.4	47.4	23.3

Table 2: Gains from using MINER on various reproduced backbone scores in Rank@1 and mAP across three CC-ReID datasets. MINER improves both metrics on all backbones and datasets.

$\overline{R@1}$ over 3 paired seeds	Random-Negative	Hard-Negative
Random-Positive	$67.1/74.3$	$72.0/75.3$
Hard-Positive	$67.5/76.0$	$\mathbf{73.6}/\mathbf{76.4}$

Table 3: Sampler ablation reporting

\overline{R@1}

on PRCC/Celeb-L. MINER uses both hard-positive and hard-negative mining.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Hyper-parameter	Value
Input image size	$224\times 224$
Normalization mean ( $\mu$ )	$(0.485,0.456,0.406)$
Normalization std. ( $\sigma$ )	$(0.229,0.224,0.225)$
Optimizer	SGD
Learning rate	$2\times 10^{-7}$
LR scheduler	Linear

Appendix

Table 4: Default training configuration (DIFFER-initialized runs).

B_{\max}

is the hard memory ceiling on the per-batch image budget

2mn

. EVA-CLIP-initialized runs use

30

epochs at learning rate

1\times 10^{-6}

; the cross-modal TVRID setting uses

\tau=0.2

( Sec. B.4 ).

SC-ReID	PRCC		LTCC
Method	R@1	mAP	R@1	mAP
TransReID ‘21	97.3	95.9	73.8	39.4
CAL ‘22	100.0	99.8	74.2	40.8
MADE ‘24	100.0	98.6	84.2	48.2
RFLNet ‘25	100.0	99.6	–	–
FA-ReID ‘25	100.0	99.0	75.7	41.3

Appendix

Table 5: Same-clothes (PRCC-SC) and General (LTCC) ReID vs. prior work, in R@1/mAP. Best is bold , second underlined . PRCC-SC R@1 is saturated at

100.0

across methods; MINER sets a new state of the art on LTCC-General.

Method	PRCC		LTCC		Celeb-L
Method	R@1	mAP	R@1	mAP	R@1	mAP
EVA-CLIP (No finetune)	39.3	23.9	15.0	5.1	57.7	37.1
Baseline (ID Loss)	65.1	63.0	54.6	30.4	72.7	53.3
DIFFER	68.5	64.7	58.2	31.6	75.6	54.3
MINER (EVA-CLIP)	74.5	70.0	60.2	34.4	77.2	59.1
MINER (DIFFER)	73.9	69.7	60.9	33.3	76.9	56.3

Appendix

Table 6: MINER versus three initialization baselines on PRCC, LTCC, and Celeb-L under the clothes-changing protocol, in R@1/mAP. Upper block: EVA-CLIP off the shelf, EVA-CLIP finetuned with a plain ID-loss, and the ReID-pretrained DIFFER encoder. Lower block: MINER applied to each of the two initializations. Best per column in bold , second underlined .

Task	Rank	Method	R@1	mAP
RGB	1st	Duy et al.	100.0	100.0
	2nd	Carlo Metta	100.0	99.9
	3rd	MINER (Ours)	99.8	99.9
	Avg.	Task average	85.7	89.6
Depth	1st	Jiang et al.	98.8	99.4
	2nd	MINER (Ours)	97.2	98.3

Appendix

Table 7: TVRID top-3 entries and per-track average over

732

submissions. MINER is the only method on the podium of all three tracks. The large podium-to-average gap, especially on Depth and RGB-Depth, reflects an unsaturated, face-free setting.

Figure 5: MINER’s contrastive loss over a cross-modal bipartite batch. Hard-negative identities and hard-positive cross-modal pairs distill inter-person (white) and intra-person (blue) variance simultaneously via a block-diagonal label matrix, here over the RGB-Depth data of the TVRID 2026 Competition.

Dataset	Setting/Factor	Sampler configuration	$\overline{\text{R@1}}$ (%)	$\Delta$ R@1, $p_{value}$
PRCC	(i)	Random positive + random negative	$67.13\pm 0.25$	–
	(ii)	Random positive + hard negative	$72.00\pm 0.26$	$+4.87$ , $p_{value}=0.0002$
	(iii)	Hard positive + random negative	$67.47\pm 0.25$	$+0.33$ , $p_{value}=0.29$
	(iv)	MINER: hard positive + hard negative	$\mathbf{73.63\pm 0.30}$	$\mathbf{+6.50}$ , $p_{value}=0.0007$
Celeb-L	(i)	Random positive + random negative	$74.27\pm 0.60$	–
	(ii)	Random positive + hard negative	$75.30\pm 0.53$	$+1.03$ , $p_{value}=0.0072$

Appendix

Table 8: Sampler-mining ablation on PRCC and Celeb-ReID-Light over three paired seeds. Results report mean

\pm

standard deviation over 3 seeds. Paired

t

-tests compare each configuration against the random-positive/random-negative baseline (i), with

\mathrm{dof}=2

Dataset	Deterministic Hard Mining	MINER
	$\overline{\text{R@1}}$ (%)	$\overline{\text{R@1}}$ (%)
PRCC	$69.93\pm 0.15$	$\mathbf{73.63\pm 0.30}$
Celeb-L	$75.87\pm 0.09$	$\mathbf{76.43\pm 0.40}$

Appendix

Table 9: Deterministic mining selects the hardest negative identities and hardest positive clothing pairs, while MINER samples stochastically from the corresponding hardness distributions. Results report mean

\pm

sample standard deviation over 3 seeds. MINER consistently improves R@1 over deterministic hard mining, with a larger gain on PRCC than on Celeb-L.

$\langle m,n\rangle$	PRCC	LTCC	Celeb-L	PRCC	LTCC	Celeb-L	Mean	Worst	Rank
	mAP			Regret			Agg. regret
$\langle 29,3\rangle$	67.6	33.9	55.7	2.1	0.6	0.8	1.17	2.1	8.83
$\langle 21,4\rangle$	67.6	34.0	56.5	2.1	0.5	0.0	0.87	2.1	6.50
$\langle 17,5\rangle$	68.2	34.4	56.3	1.5	0.1	0.2	0.60	1.5	5.50
$\langle 12,7\rangle$	68.6	34.2	56.4	1.1	0.3	0.1	0.50	1.1	5.17
$\langle 9,~~9\rangle$	68.9	34.4	56.3	0.8	0.1	0.2	0.37	0.8	4.50

Appendix

Table 10: Per-allocation mAP and regret (gap to each dataset’s grid optimum) on PRCC/LTCC/Celeb-L, with mean and worst-case regret and Demšar average rank across the three. The Variance Controller attains the lowest mean and worst-case regret of all configurations—Pareto-dominating every fixed split—and the best average rank. Lower is better throughout; per-dataset optima are underlined.

Figure 6: Risk (orange) versus mAP regret (blue) across Tab. 10 ’s nine allocations, each normalized within its dataset; lower is better. Dashed lines mark the Controller’s choice — the risk minimum by construction (Eq. 15) — and shaded cells the best measured. The test is whether the curves track. They do on PRCC (

r_{s}=-0.97

) and Celeb-L (

r_{s}=-0.85

), where lowest risk is also best measured. LTCC spans only

0.6

mAP, too flat to order (

r_{s}=+0.14

p_{value}=0.73

); the Controller’s

\langle 9,9\rangle

lands

0.1

mAP off.

Figure 7: Effect of the risk weighting

\lambda

on Celeb-L. Top: retrieval performance (mAP, R@1) is flat across

\lambda

. Bottom: the controller allocation

\langle m^{\star},n^{\star}\rangle

shifts markedly with

\lambda

. The allocation moves while performance does not, confirming the equal-weighting choice is immaterial. Mean over three seeds; error bars are standard deviation.

	Allocation $\langle m,n\rangle$ (iii)		(ii)
Dataset	Proj. $\nabla_{\mathbf{z}}\ell$	Full $\nabla_{\theta}\ell$	$r_{\mathrm{cross}}$
PRCC	$\langle 5,17\rangle$	$\langle 5,17\rangle$ (exact)	$-0.23$
LTCC	$\langle 9,~~9\rangle$	$\langle 10,8\rangle$ (within $1$ )	$-0.25$

Appendix

Table 11: Validation of Assumptions (ii) and (iii). The last column reports the trace-correlation

r_{\mathrm{cross}}

between the invariance and discrimination gradient signals (Assumption ii); it is small and negative, so the additive risk is a conservative upper bound. The first two columns report the allocation recovered by the projection-head versus the full-parameter estimator (Assumption iii); they agree, exactly on PRCC and within one unit per coordinate on LTCC.

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

CMCC-ReID: Cross-Modality Clothing-Change Person Re-Identification
Apr 3, 2026Haoxuan Xu, Hanzi Wang, Guanglin NiuVisible-Infrared Person Re-Identification Person Re-Identification
School of Artificial Intelligence, Beihang University, Beijing, China · School of Informatics, Xiamen University, Xiamen, Fujian, China
Learning Instance-Adaptive Low-Rank Orthogonal Subspaces for Clothes-Changing Person Re-Identification
Jun 10, 2026Dong-Woo Kim, Tae-Kyun KimInvariant Representation Learning Person Re-Identification
School of Computing, KAIST, Daejeon, South Korea.
Identity-Conditioned Score Fusion for Open-Set Person Re-Identification
Oct 5, 2026Manyi Yao, Jurijs Nazarovs, Eunji Chong +6Multimodal Fusion Person Re-Identification
University of California, Riverside · Amazon

Apr 3, 2026J/K move · Enter open · S save

CMCC-ReID: Cross-Modality Clothing-Change Person Re-Identification

Haoxuan Xu, Hanzi Wang, Guanglin Niu

School of Artificial Intelligence, Beihang University, Beijing, China · School of Informatics, Xiamen University, Xiamen, Fujian, China

Visible-Infrared Person Re-Identification Person Re-Identification Multimodal Disentangled Representation Learning

Open paper AskPDF arXiv

Person Re-Identification (ReID) faces severe challenges from modality discrepancy and clothing variation in long-term surveillance scenario. While existing studies have made significant progress in either Visible-Infrared ReID (VI-ReID) or Clothing-Change ReID (CC-ReID), real-world surveillance system often face both challenges simultaneously. To address this overlooked yet realistic problem, we define a new task, termed Cross-Modality Clothing-Change Re-Identification (CMCC-ReID), which targets pedestrian matching across variations in both modality and clothing. To advance research in this direction, we construct a new benchmark SYSU-CMCC, where each identity is captured in both visible and infrared domains with distinct outfits, reflecting the dual heterogeneity of long-term surveillance. To tackle CMCC-ReID, we propose a Progressive Identity Alignment Network (PIA) that progressively mitigates the issues of clothing variation and modality discrepancy. Specifically, a Dual-Branch Disentangling Learning (DBDL) module separates identity-related cues from clothing-related factors to achieve clothing-agnostic representation, and a Bi-Directional Prototype Learning (BPL) module performs intra-modality and inter-modality contrast in the embedding space to bridge the modality gap while further suppressing clothing interference. Extensive experiments on the SYSU-CMCC dataset demonstrate that PIA establishes a strong baseline for this new task and significantly outperforms existing methods.

Abstract

Person Re-Identification

Explore similar work

CMCC-ReID: Cross-Modality Clothing-Change Person Re-Identification

Learning Instance-Adaptive Low-Rank Orthogonal Subspaces for Clothes-Changing Person Re-Identification

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