Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proposes a weakly supervised visible-infrared person re-identification framework that recovers cross-modal identity correspondences from single-modal labels alone, matching fully supervised accuracy on SYSU-MM01 and LLCM.

desk verdict Coherent method and clear writing, but the central claim—cross-modal correspondences are unknown—is not actually tested because both datasets use a shared identity-ID space; a label-permutation experiment is needed before the headline result can be taken at face value. read the letter →

arxiv 2507.12942 v1 pith:UVMNUY2N submitted 2025-07-17 cs.CV

Yafei Zhang , Lingqi Kong , Huafeng Li , Jie Wen This is my paper
classification cs.CV
keywords weaklysupervisedlearningvisible-infraredpersonre-identificationcross-modalidentitycorrespondencesheterogeneousexpertscollaborativeconsistencyCountPrioritySelectionSYSU-MM01LLCM
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether a person can be matched between visible-light and infrared images when the training data contains no cross-modal identity labels, meaning no annotation says which visible image and which infrared image show the same person. It proposes a training framework in which each modality trains its own classification expert on its own labels, and these experts then predict the other modality's identities; fusing those predictions produces the missing cross-modal correspondences. On the SYSU-MM01 and LLCM benchmarks, the method reaches Rank-1 accuracy comparable to certain fully supervised systems, which would make cross-modal person matching far cheaper to annotate. The central claim is that single-modal identity labels suffice to learn modality-invariant identities.

What carries the argument

The central mechanism is the heterogeneous expert pair, two classifiers $W^v$ and $W^r$ each trained only within its own modality, together with the cross-modal relationship matrices $M_c$, $M_s$, and $M_w$ built from their mutual predictions. $M_c$ marks identity pairs on which both experts agree, $M_s$ marks one-sided unique matches, and $M_w$ marks conflicting matches. These matrices convert expert opinion into a supervisory signal: a shared classifier $W^c$ is trained with cross-entropy on the confident correspondences and a weakened loss on conflicts, while the two experts are co-trained so their predictions on matched prototypes converge. The machinery's job is to bootstrap cross-modal alignment from purely intra-modal supervision and then to keep refining that alignment as the experts improve.

What would settle it

Use the identical training pipeline on SYSU-MM01 or LLCM but permute the identity labels within one modality independently of the other, so that no true cross-modal correspondence exists; if the method still reports retrieval accuracy far above chance, the correspondences it claims to recover must be leaking from the label structure, and the weak-supervision claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a weak supervision signal, labels available separately within each modality, can be converted into cross-modal identity supervision by having heterogeneous experts vouch for one another. After training a visible expert and an infrared expert on their own modalities, each expert classifies samples from the other modality, and a Count Priority Selection step turns the two predicted classification matrices into binary correspondence matrices. These are partitioned into consistent, unique, and conflicting matches: the consistent and unique matches supply pseudo-labels for a cross-modal cross-entropy loss, while conflicting matches are handled by a relaxed loss that discourages the shared classifier from firing on disputed identities. A prototype-based homogeneity loss then pushes the two experts to make consistent predictions on matched identities, and the loop repeats with the experts and the encoder mutually refining each other. The result is that on SYSU-MM01 all-search, the method reaches Rank-1 of 70.4% and mAP of 66.6%, close to the fully supervised CAJ method at 69.9% and 66.9%, and above the compared semi-supervised and unsupervised methods.

Load-bearing premise

The framework assumes that, before any cross-modal training happens, each modality's expert predicts the other modality's identities accurately enough that the first round of Count Priority Selection produces more correct than wrong correspondences, and this initial alignment quality is never measured.

Editorial extensions

If this is right

  • Weakly supervised visible-infrared person re-identification is feasible: single-modality identity labels suffice to reach accuracy comparable to some fully supervised methods, reducing the need for expensive cross-modal annotation.
  • Splitting expert predictions into consistent, unique, and conflicting matches extracts more usable supervision than relying only on unanimous expert agreement, which would otherwise miss many valid correspondences.
  • Collaborative consistency learning between experts improves both the quality of the recovered correspondences and the modality-invariance of the learned features, as shown by the ablation study where the expert-collaboration component adds 2.4% Rank-1 on SYSU-MM01 all-search.
  • The framework outperforms the compared semi-supervised and unsupervised visible-infrared person re-identification methods on SYSU-MM01 and LLCM, indicating that intra-modal labels provide a cleaner training signal than pseudo-labels produced from unlabeled data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The method is never tested with independently permuted per-modality identity IDs, the setting in which cross-modal correspondence is truly unknown; a natural stress test is to randomly relabel the intra-modal IDs and check whether retrieval stays above chance.
  • If the bootstrap works, the expert-consensus idea could transfer to other multi-modal matching problems, such as sketch-to-photo or text-to-image person retrieval, where paired labels are expensive but within-modality labels are cheap.
  • The quality of the initial expert predictions is never quantified; measuring the precision of $M_c$, $M_s$, and $M_w$ immediately after the first round would reveal how much headroom the iterative refinement actually has.
  • Because the framework explicitly models conflicting expert opinions instead of discarding them, the same mechanism could be adapted to improve robustness against label noise within a single modality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a weakly supervised visible-infrared person re-identification (VIReID) framework consisting of three components: heterogeneous expert learning (HEL), which trains separate visible and infrared classification experts using only intra-modal identity labels with cross-entropy and weighted regularization triplet losses (Eqs. 1-2); cross-modal relationship establishment (CRE), which builds consistent, unique, and conflicting cross-modal identity correspondence matrices Mc, Ms, and Mw via Count Priority Selection applied to the experts' cross-predictions (Eqs. 4-6); and collaborative consistency learning (CCL), which uses these correspondences to supervise a shared classifier and to enforce consistency between the experts through prototype-based homogeneity losses (Eqs. 7-13). The paper claims that this is the first weakly supervised VIReID method that requires only single-modal identity labels and eliminates the need for cross-modal correspondence annotations. On SYSU-MM01 and LLCM it reports Rank-1/mAP of 70.4/66.6 (all-search) and 55.3/58.7 (VIS-to-IR), respectively, surpassing unsupervised and semi-supervised methods and approaching some fully supervised methods (Tables 1 and 2), with ablations in Tables 3 and 4 attributing monotonic gains to each component.

Significance. If the central claim is established, this is a valuable contribution: it opens a genuinely cheaper supervision paradigm for VIReID, the framework is clearly specified with explicit losses and matrix definitions, the code is publicly released, and the reported results are strong, beating state-of-the-art unsupervised (GUR, DLM) and semi-supervised (DPIS, OTLA) methods and rivaling the fully supervised CAJ. The ablations are sensible and show monotonic gains. That said, the significance is fully conditional on resolving the mismatch between the claimed unknown-correspondence setting and the aligned label space of the two evaluation datasets; as written, the experiments cannot establish that the framework discovers cross-modal correspondences rather than exploiting the implicit correspondence encoded in the shared identity-ID vocabulary.

major comments (4)
  1. [§3.1; §4.1; Tables 1-2] The claimed setting of unknown cross-modal correspondences is not enforced by the evaluation datasets. Section 3.1 states that cross-modal label correspondences remain unknown, but SYSU-MM01 and LLCM share a single global identity-ID vocabulary across the visible and infrared training sets, so a visible sample with identity i and an infrared sample with identity i are the same person by construction; the per-modality labels themselves therefore encode the cross-modal correspondence. Because Eqs. (4)-(6) build Mc, Ms, and Mw on these already aligned index spaces, the numbers in Tables 1 and 2 do not distinguish the hypothesis that the framework discovers correspondences from the hypothesis that it inherits them from the shared IDs. I request a decisive control experiment: apply a fixed random permutation to the infrared training identity labels while keeping the visible labels unchanged, train the full pipeline without any knowledge of the permutation, and evaluate Rank-1/mAP against the true mapping. High performance after permutation would support the claimed mechanism, while a collapse toward the corresponding baseline would refute it. This experiment is load-bearing for the paper's central claim and should be reported.
  2. [§4.4; Table 3] The strong baseline result in Table 3 is itself evidence of the aligned-label confound. The baseline B, trained only with the intra-modal losses of Eqs. (1)-(2), achieves 47.8% Rank-1 in all-search cross-modal retrieval; in a genuinely unaligned label space, an intra-modal classification objective contains no cross-modal pairing signal and should yield near-chance cross-modal Rank-1. The observed 47.8% is fully consistent with the shared-ID alignment being exploited implicitly, and the statement in Section 4.4 that a model trained solely within modalities 'already possesses some cross-modal retrieval capability' does not rule out this confound. The permuted-label experiment is therefore necessary to interpret not only the headline numbers but also the ablation ladder in Table 3.
  3. [§3.3; §3.4; Table 4] The quality of the pseudo-correspondences produced by CRE is never quantified, although these correspondences provide the supervision for all consistency losses in Eqs. (7)-(13). Since the claimed contribution is that the framework establishes robust cross-modal identity correspondences, the paper should report the accuracy (e.g., precision/recall or match rate) of Mc, Ms, and Mw against the ground-truth correspondences on the training identities, both at initialization and after collaborative learning, and ideally also under the permuted-label condition. Without this measurement, the gains shown in Tables 3 and 4 could reflect consistency losses that are robust to an almost-trivial correspondence structure in the aligned label space rather than accurate correspondence discovery.
  4. [§3.3; Eq. (5)] The definition of the single identity relationship matrix in Eq. (5) does not match its prose description. The indicator 1_{0}(·) is defined to return 1 when the input equals 0, so r_i = 1 when identity i has zero predicted correspondences in M^{t̄→t}, and c_j = 1 when identity j has zero predicted correspondences in the reverse direction; the term A^T ⊙ M^{t→t̄} then selects forward predictions only for pairs in which neither identity has any correspondence in the reverse matrix. As written, this selects identities with no matches, whereas the text describes Ms as capturing unique matches (identities with exactly one correspondence). Please correct the indicator (likely 1_{1}(·)) and/or the orientation of r and c relative to the rows and columns of M^{t→t̄}, and verify that the corrected definition matches the implementation.
minor comments (6)
  1. [§4.3; Table 1] The heading 'Comparasion' in Section 4.3 and in the Table 1 caption is a typo and should read 'Comparison'.
  2. [§3.3] The Count Priority Selection procedure is borrowed from [39] but never described; since Eqs. (4)-(6) and the whole CRE module depend on it, please provide the algorithm in the main text or in the supplementary material so that the method is self-contained and reproducible.
  3. [§3.4; Eq. (7)] The shared classifier W^c is introduced without a defined output dimension or a statement of how it relates to the per-modality label spaces C^v and C^r; please clarify whether C^v and C^r coincide in the experiments and how W^c is initialized and trained.
  4. [§4.4; Tables 3-4] No variance statistics are reported anywhere; because several ablation increments are small (e.g., Table 4 shows 68.0 vs. 67.7 Rank-1 and 66.6 vs. 64.6 mAP), please report results over at least three random seeds or provide error bars.
  5. [§3.4; Eqs. (12)-(13)] The entropy notation H(·) and the exact quantities p^{v→r} and p^{r→v} that enter the weights in Eq. (13) are not defined; please specify which prediction vectors are averaged to compute these weights.
  6. [§1; contribution bullet 1] The claim of exploring the weakly supervised VIReID setting 'for the first time' should be tempered or supported by a targeted comparison with prior label-efficient settings, since the related work already discusses methods that reduce cross-modal annotation such as OTLA [33] and DPIS [30].

Circularity Check

1 steps flagged · score 4.0 of 10

Claimed unknown cross-modal correspondences are pre-aligned in the benchmark label space, so the central weak-supervision setting is not actually tested.

  1. self definitional [Section 3.1 (problem definition) vs Section 4.1 (datasets and evaluation protocols)]
    "The corresponding label sets are given by Y v = {yv i}Nv i=1 and Y r = {yr i}Nr i=1, where yv i ∈ R1×Cv and yr i ∈ R1×Cr represent the labels associated with samples xv i and xr i, respectively. Under the weakly supervised setting considered in this paper, label information is available within each modality, whereas cross-modal label correspondences remain unknown. [...] The training set consists of 22,258 visible images and 11,909 infrared images from 395 identities, while the test set covers 96 identities."

    The paper defines the weakly supervised setting by saying cross-modal label correspondences are unknown, but both evaluation datasets use one global identity vocabulary for the two modalities: a visible label i and an infrared label i denote the same person by construction. Wv and Wr are therefore trained on the same class-index set, and the Count Priority Selection matrices in Eqs. 4-6 align class positions that are already aligned by the shared ID semantics. The inferred correspondences Mc, Ms, and Mw can be correct simply by inheriting this pre-alignment, rather than by solving the permutation problem the setting claims to address.

full rationale

Most of the optimization chain (Eqs. 1-15) is a normal alternating pseudo-labeling procedure: modality-specific experts make predictions, Count Priority Selection converts those predictions into relation matrices, and the cross-entropy, triplet, and homogeneity losses push encoders and experts toward agreement. This is self-referential in the style of self-training, but the final Rank-1/mAP numbers are measured on held-out test identities with ground-truth labels, so the reported empirical result is not a tautology derived from the model's own fitted values. The one self-citation, reference [57] used for style augmentation, is an implementation detail and is not load-bearing. The real circularity is at the level of the problem setup: Section 3.1 declares cross-modal correspondences unknown, while Section 4.1 uses datasets whose visible and infrared training labels share a common identity-index vocabulary (395 identities on SYSU-MM01 and 713 on LLCM). In that vocabulary, the single-modal labels already encode which visible and infrared identities correspond, and Eqs. 4-6 operate on this pre-aligned index space. The central claim therefore needs a label-permutation experiment before it can be accepted as evidence that the framework discovers cross-modal correspondences; as designed, the benchmark may be reading off an alignment that is already implicit in the labels. Because the final evaluation is external and the learning loop can still fail, the circularity is partial rather than complete, giving a score of 4.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method rests on common ML machinery (pretrained ResNet-50, cross-entropy and triplet losses), plus hand-set hyperparameters and an unquantified assumption that single-modality classifiers produce useful cross-modal predictions at initialization. The most consequential premise is that the label IDs across modalities are meaningful in the same identity space; the experiments use datasets with aligned IDs, so the claimed unknown correspondences are only hidden by choice, not by construction.

free parameters (4)
  • lambda1 (loss balance weight for weighted regularization triplet loss) = 0.25
    Set by hand in Section 4.2; balances intra-modality and cross-modal triplet losses in Eqs. (14) and (15). No sensitivity analysis is reported.
  • lambda2 (weight for relaxed identity discrimination loss) = 0.25
    Set by hand in Section 4.2; controls the contribution of the conflicting-correspondence loss in Eq. (8).
  • lambda (prototype momentum coefficient) = 0.8
    Set by hand in Section 4.2; controls prototype update in Eq. (10).
  • Count Priority Selection thresholds and parameters from ADCA [39] = not specified in the paper
    The decision matrices M in Section 3.3 rely on this borrowed procedure, but the paper does not state its hyperparameters, so the exact mapping from expert prediction scores to binary correspondences is underspecified.
assumptions (4)
  • domain assumption ResNet-50 pretrained on ImageNet provides a feature space in which visible and infrared images of the same person are initially closer than random.
    Section 4.2 uses ResNet-50 without explaining how cross-modal alignment is initialized; the whole HEL and CRE pipeline depends on this warm start.
  • ad hoc to paper The label IDs in the visible and infrared training sets refer to the same global identity space.
    Section 3.1 defines Y^v and Y^r as labels from shared datasets but claims cross-modal correspondences are unknown; the experiments on SYSU-MM01 and LLCM use aligned label IDs, making the problem easier than a setting with independently permuted label sets.
  • domain assumption The shared encoder layers and the two expert classifiers can be aligned through within-modality classification losses alone, without explicit cross-modal pairs.
    Section 3.2 trains W^v and W^r only on single-modality losses; later cross-modal consistency depends on this implicit alignment.
  • domain assumption Reliable cross-modal correspondences can be derived from agreement between expert predictions.
    Section 3.3 defines Mc, Ms, and Mw and treats them as supervision; no precision or recall of these pseudo-correspondences is measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning." pith.science (2026). https://pith.science/paper/UVMNUY2N

@misc{pith2026250712942,
  author       = {Pith},
  title        = {Pith review of: Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UVMNUY2N}},
  note         = {Machine review of arXiv:2507.12942}
}
read the original abstract

To reduce the reliance of visible-infrared person re-identification (ReID) models on labeled cross-modal samples, this paper explores a weakly supervised cross-modal person ReID method that uses only single-modal sample identity labels, addressing scenarios where cross-modal identity labels are unavailable. To mitigate the impact of missing cross-modal labels on model performance, we propose a heterogeneous expert collaborative consistency learning framework, designed to establish robust cross-modal identity correspondences in a weakly supervised manner. This framework leverages labeled data from each modality to independently train dedicated classification experts. To associate cross-modal samples, these classification experts act as heterogeneous predictors, predicting the identities of samples from the other modality. To improve prediction accuracy, we design a cross-modal relationship fusion mechanism that effectively integrates predictions from different experts. Under the implicit supervision provided by cross-modal identity correspondences, collaborative and consistent learning among the experts is encouraged, significantly enhancing the model's ability to extract modality-invariant features and improve cross-modal identity recognition. Experimental results on two challenging datasets validate the effectiveness of the proposed method.

Figures

Figures reproduced from arXiv: 2507.12942 by the authors.

Figure 1
Figure 1. VIReID faces two key challenges: asynchronous visible [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method. HEL first performs intra-modality training to obtain experts for each modality, generating [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization comparison of feature distributions be [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 57 canonical work pages

  1. [1]

    Structure-aware positional trans- former for visible-infrared person re-identification

    Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Jianguo Jiang, and Chia-Wen Lin. Structure-aware positional trans- former for visible-infrared person re-identification. IEEE Transactions on Image Processing, 31:2352–2364, 2022. 2

    work page 2022

  2. [2]

    Joint feature learning network for visible-infrared person re-identification

    Kunfeng Chen, Zhisong Pan, Jiabao Wang, Shanshan Jiao, Zhicheng Zeng, and Zhuang Miao. Joint feature learning network for visible-infrared person re-identification. In Chi- nese Conference on Pattern Recognition and Computer Vi- sion (PRCV), pages 652–663. Springer, 2020. 2

    work page 2020

  3. [3]

    Unveiling the power of clip in unsupervised visible-infrared person re-identification

    Zhong Chen, Zhizhong Zhang, Xin Tan, Yanyun Qu, and Yuan Xie. Unveiling the power of clip in unsupervised visible-infrared person re-identification. In Proceedings of the 31st ACM International Conference on Multimedia (ACMMM), pages 3667–3675, 2023. 3, 7

    work page 2023

  4. [4]

    Efficient bilateral cross- modality cluster matching for unsupervised visible-infrared person reid

    De Cheng, Lingfeng He, Nannan Wang, Shizhou Zhang, Zhen Wang, and Xinbo Gao. Efficient bilateral cross- modality cluster matching for unsupervised visible-infrared person reid. In Proceedings of the 31st ACM International Conference on Multimedia (ACMMM) , pages 1325–1333,

    work page

  5. [5]

    Hi-cmd: Hierarchical cross-modality disentanglement for visible-infrared person re-identification

    Seokeon Choi, Sumin Lee, Youngeun Kim, Taekyung Kim, and Changick Kim. Hi-cmd: Hierarchical cross-modality disentanglement for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 10257– 10266, 2020. 2

    work page 2020

  6. [6]

    Cross-modality person re-identification with generative adversarial training

    Pingyang Dai, Rongrong Ji, Haibin Wang, Qiong Wu, and Yuyu Huang. Cross-modality person re-identification with generative adversarial training. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI), pages 677–683, 2018. 2

    work page 2018

  7. [7]

    Shape-erased feature learning for visible-infrared person re-identification

    Jiawei Feng, Ancong Wu, and Wei-Shi Zheng. Shape-erased feature learning for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 22752– 22761, 2023. 2

    work page 2023

  8. [8]

    Learning modality-specific representations for visible-infrared person re-identification

    Zhanxiang Feng, Jianhuang Lai, and Xiaohua Xie. Learning modality-specific representations for visible-infrared person re-identification. IEEE Transactions on Image Processing , 29:579–590, 2020. 2

    work page 2020

Show all 57 references
  1. [9]

    Cm-nas: Cross-modality neural architecture search for visible-infrared person re-identification

    Chaoyou Fu, Yibo Hu, Xiang Wu, Hailin Shi, Tao Mei, and Ran He. Cm-nas: Cross-modality neural architecture search for visible-infrared person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 11823–11832, 2021. 2

    2021

  2. [10]

    Hsme: Hy- persphere manifold embedding for visible thermal person re- identification

    Yi Hao, Nannan Wang, Jie Li, and Xinbo Gao. Hsme: Hy- persphere manifold embedding for visible thermal person re- identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 8385–8392, 2019. 1, 2

    2019

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 4

    2016

  4. [12]

    Exploring homogeneous and heterogeneous consistent label associations for unsupervised visible-infrared person reid

    Lingfeng He, De Cheng, Nannan Wang, and Xinbo Gao. Exploring homogeneous and heterogeneous consistent label associations for unsupervised visible-infrared person reid. International Journal of Computer Vision , 133:3129–3148,

  5. [13]

    Adversarial decoupling and modality- invariant representation learning for visible-infrared person re-identification

    Weipeng Hu, Bohong Liu, Haitang Zeng, Yanke Hou, and Haifeng Hu. Adversarial decoupling and modality- invariant representation learning for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5095–5109, 2022. 2

    2022

  6. [14]

    Deep learning for visible-infrared cross- modality person re-identification: A comprehensive review

    Nianchang Huang, Jianan Liu, Yunqi Miao, Qiang Zhang, and Jungong Han. Deep learning for visible-infrared cross- modality person re-identification: A comprehensive review. Information Fusion, 91:396–411, 2023. 2

    2023

  7. [15]

    Alleviating modality bias training for infrared-visible person re-identification

    Yan Huang, Qiang Wu, Jingsong Xu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Alleviating modality bias training for infrared-visible person re-identification. IEEE Transactions on Multimedia, 24:1570–1582, 2022. 2

    2022

  8. [16]

    Kajal Kansal, A. V . Subramanyam, Zheng Wang, and Shin’ichi Satoh. Sdl: Spectrum-disentangled representation learning for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 30(10):3422–3432, 2020. 2

    2020

  9. [17]

    Infrared-visible cross-modal person re-identification with an x modality

    Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In Proceedings of the AAAI Conference on Arti- ficial Intelligence (AAAI), pages 4610–4617, 2020. 2

    2020

  10. [18]

    Shape-centered representation learning for visible-infrared person re-identification

    Shuang Li, Jiaxu Leng, Ji Gan, Mengjingcheng Mo, and Xinbo Gao. Shape-centered representation learning for visible-infrared person re-identification. Pattern Recogni- tion, 167:111756, 2025. 2

    2025

  11. [19]

    Inter-intra modality knowledge learning and clustering noise alleviation for unsupervised visible-infrared person re- identification

    Zhiyong Li, Haojie Liu, Xiantao Peng, and Wei Jiang. Inter-intra modality knowledge learning and clustering noise alleviation for unsupervised visible-infrared person re- identification. IEEE Transactions on Knowledge and Data Engineering, 36(8):3934–3947, 2024. 3

    2024

  12. [20]

    Homogeneous-to-heterogeneous: Unsupervised learning for rgb-infrared person re-identification

    Wenqi Liang, Guangcong Wang, Jianhuang Lai, and Xiao- hua Xie. Homogeneous-to-heterogeneous: Unsupervised learning for rgb-infrared person re-identification. IEEE Transactions on Image Processing, 30:6392–6407, 2021. 3

    2021

  13. [21]

    Parameter sharing exploration and hetero-center triplet loss for visible- thermal person re-identification

    Haijun Liu, Xiaoheng Tan, and Xichuan Zhou. Parameter sharing exploration and hetero-center triplet loss for visible- thermal person re-identification. IEEE Transactions on Mul- timedia, 23:4414–4425, 2021. 2

    2021

  14. [22]

    Learning memory-augmented unidirectional metrics for cross-modality person re-identification

    Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19366–19375, 2022. 7

    2022

  15. [23]

    Pose-guided modality- invariant feature alignment for visible-infrared object re- identification

    Min Liu, Yeqing Sun, Xueping Wang, Yuan Bian, Zhu Zhang, and Yaonan Wang. Pose-guided modality- invariant feature alignment for visible-infrared object re- identification. IEEE Transactions on Instrumentation and Measurement, 73:1–10, 2024. 2

    2024

  16. [24]

    Disentan- gling modality and posture factors: Memory-attention and orthogonal decomposition for visible-infrared person re- identification

    Zefeng Lu, Ronghao Lin, and Haifeng Hu. Disentan- gling modality and posture factors: Memory-attention and orthogonal decomposition for visible-infrared person re- identification. IEEE Transactions on Neural Networks and Learning Systems, 36(3):5494–5508, 2025. 2

    2025

  17. [25]

    Mimr: Modality-invariance modeling and refinement for unsupervised visible-infrared person re-identification

    Zhiqi Pang, Chunyu Wang, Honghu Pan, Lingling Zhao, Junjie Wang, and Maozu Guo. Mimr: Modality-invariance modeling and refinement for unsupervised visible-infrared person re-identification. Knowledge-Based Systems , 285: 111350, 2024. 3

    2024

  18. [26]

    Cross-modality hierarchical clustering and refinement for unsupervised visible-infrared person re- identification

    Zhiqi Pang, Chunyu Wang, Lingling Zhao, Yang Liu, and Gaurav Sharma. Cross-modality hierarchical clustering and refinement for unsupervised visible-infrared person re- identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2706–2718, 2024. 3

    2024

  19. [27]

    Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences

    Hyunjong Park, Sanghoon Lee, Junghyup Lee, and Bum- sub Ham. Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12046–12055, 2021. 1, 2, 7

    2021

  20. [28]

    Bakker, and Michael S

    Nan Pu, Wei Chen, Yu Liu, Erwin M. Bakker, and Michael S. Lew. Dual gaussian-based variational subspace disentangle- ment for visible-infrared person re-identification. In Pro- ceedings of the 28th ACM International Conference on Mul- timedia (ACMMM), pages 2149–2158, 2020. 1

    2020

  21. [29]

    High-order structure based middle- feature learning for visible-infrared person re-identification

    Liuxiang Qiu, Si Chen, Yan Yan, Jing-Hao Xue, Da-Han Wang, and Shunzhi Zhu. High-order structure based middle- feature learning for visible-infrared person re-identification. In Proceedings of the AAAI Conference on Artificial Intelli- gence (AAAI), pages 4596–4604, 2024. 2

    2024

  22. [30]

    Dual pseudo-labels interactive self-training for semi- supervised visible-infrared person re-identification

    Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi- supervised visible-infrared person re-identification. In Pro- ceedings of the IEEE/CVF International Conferenc...

    2023

  23. [31]

    Multi- memory matching for unsupervised visible-infrared person re-identification

    Jiangming Shi, Xiangbo Yin, Yeyun Chen, Yachao Zhang, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Multi- memory matching for unsupervised visible-infrared person re-identification. In European Conference on Computer Vi- sion (ECCV), pages 456–474. Springer, 2024. 3, 7

    2024

  24. [32]

    Enhanc- ing unsupervised visible-infrared person re-identification with bidirectional-consistency gradual matching

    Xiao Teng, Xingyu Shen, Kele Xu, and Long Lan. Enhanc- ing unsupervised visible-infrared person re-identification with bidirectional-consistency gradual matching. In Pro- ceedings of the 32nd ACM International Conference on Mul- timedia (ACMMM), pages 9856–9865, 2024. 3

    2024

  25. [33]

    Optimal transport for label-efficient visible-infrared person re-identification

    Jiangming Wang, Zhizhong Zhang, Mingang Chen, Yi Zhang, Cong Wang, Bin Sheng, Yanyun Qu, and Yuan Xie. Optimal transport for label-efficient visible-infrared person re-identification. In European Conference on Computer Vi- sion (ECCV), pages 93–109. Springer, 2022. 1, 3, 7

    2022

  26. [34]

    Learning to reduce dual-level discrepancy for infrared-visible person re-identification

    Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin’ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 618–626...

    2019

  27. [35]

    Rgb-infrared cross-modality per- son re-identification

    Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality per- son re-identification. In Proceedings of the IEEE interna- tional Conference on Computer Vision (ICCV), pages 5380– 5389, 2017. 2, 6

    2017

  28. [36]

    Asym- metric mutual learning for unsupervised transferable visible- infrared re-identification

    Ancong Wu, Chengzhi Lin, and Wei-Shi Zheng. Asym- metric mutual learning for unsupervised transferable visible- infrared re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(11):10524–10537, 2024. 1, 3

    2024

  29. [37]

    Unsupervised visible-infrared per- son re-identification via progressive graph matching and al- ternate learning

    Zesen Wu and Mang Ye. Unsupervised visible-infrared per- son re-identification via progressive graph matching and al- ternate learning. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pages 9548–9558, 2023. 3, 7

    2023

  30. [38]

    Fmcnet +: Feature-level modality com- pensation for visible-infrared person re-identification

    Ruida Xi, Nianchang Huang, Changzhou Lai, Qiang Zhang, and Jungong Han. Fmcnet +: Feature-level modality com- pensation for visible-infrared person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 36 (7):13247–13261, 2025. 2

    2025

  31. [39]

    Aug- mented dual-contrastive aggregation learning for unsuper- vised visible-infrared person re-identification

    Bin Yang, Mang Ye, Jun Chen, and Zesen Wu. Aug- mented dual-contrastive aggregation learning for unsuper- vised visible-infrared person re-identification. In Proceed- ings of the 30th ACM International Conference on Multime- dia (ACMMM), pages 2843–2851, 2022. 3, 4, 7

    2022

  32. [40]

    Towards grand unified representation learning for unsupervised visible-infrared per- son re-identification

    Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared per- son re-identification. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV) , pages 11069–11079, 2023. 3, 7

    2023

  33. [41]

    Dual consistency-constrained learning for unsupervised visible- infrared person re-identification

    Bin Yang, Jun Chen, Cuiqun Chen, and Mang Ye. Dual consistency-constrained learning for unsupervised visible- infrared person re-identification. IEEE Transactions on In- formation Forensics and Security, 19:1767–1779, 2024. 3

    2024

  34. [42]

    Shallow-deep collab- orative learning for unsupervised visible-infrared person re- identification

    Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collab- orative learning for unsupervised visible-infrared person re- identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16870–16879, 2024. 3

    2024

  35. [43]

    Progressive cross-modal association learning for unsupervised visible- infrared person re-identification

    Yiming Yang, Weipeng Hu, and Haifeng Hu. Progressive cross-modal association learning for unsupervised visible- infrared person re-identification. IEEE Transactions on In- formation Forensics and Security, 20:1290–1304, 2025. 1, 3, 7

    2025

  36. [44]

    Hi- erarchical discriminative learning for visible thermal person re-identification

    Mang Ye, Xiangyuan Lan, Jiawei Li, and Pong Yuen. Hi- erarchical discriminative learning for visible thermal person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7501–7508, 2018. 2

    2018

  37. [45]

    Mang Ye, Zheng Wang, Xiangyuan Lan, and Pong C. Yuen. Visible thermal person re-identification via dual-constrained top-ranking. In Proceedings of the Twenty-Seventh Interna- tional Joint Conference on Artificial Intelligence, (IJCAI) , pages 1092–1099, 2018. 2, 6

    2018

  38. [46]

    Visible-infrared person re-identification via homogeneous augmented tri- modal learning

    Mang Ye, Jianbing Shen, and Ling Shao. Visible-infrared person re-identification via homogeneous augmented tri- modal learning. IEEE Transactions on Information Foren- sics and Security, 16:728–739, 2020. 2

    2020

  39. [47]

    Channel augmented joint learning for visible-infrared recog- nition

    Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recog- nition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13547–13556, 2021. 2, 7

    2021

  40. [48]

    Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re- identification: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(6):2872– 2893, 2022. 2, 4, 6, 7

    2022

  41. [49]

    Dual-level matching with outlier filtering for unsupervised visible-infrared person re- identification

    Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re- identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(5):3815–3829, 2025. 3, 7

    2025

  42. [50]

    Robust pseudo-label learning with neighbor relation for unsuper- vised visible-infrared person re-identification

    Xiangbo Yin, Jiangming Shi, Yachao Zhang, Yang Lu, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Robust pseudo-label learning with neighbor relation for unsuper- vised visible-infrared person re-identification. In Proceed- ings of the 32nd ACM International Conference on Multime- d...

    2024

  43. [51]

    Modality unifying network for visible-infrared per- son re-identification

    Hao Yu, Xu Cheng, Wei Peng, Weihao Liu, and Guoying Zhao. Modality unifying network for visible-infrared per- son re-identification. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV) , pages 11185–11195, 2023. 7

    2023

  44. [52]

    Clip-driven semantic discovery network for visible-infrared person re-identification

    Xiaoyan Yu, Neng Dong, Liehuang Zhu, Hao Peng, and Dapeng Tao. Clip-driven semantic discovery network for visible-infrared person re-identification. IEEE Transactions on Multimedia, 27:4137–4150, 2025. 7

    2025

  45. [53]

    Fmcnet: Feature-level modality compensation for visible-infrared person re-identification

    Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) , pages 7349–7358,

  46. [54]

    Diverse embedding expan- sion network and low-light cross-modality benchmark for visible-infrared person re-identification

    Yukang Zhang and Hanzi Wang. Diverse embedding expan- sion network and low-light cross-modality benchmark for visible-infrared person re-identification. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2153–2162, 2023. 1, 2, 6, 7

    2023

  47. [55]

    Joint color-irrelevant consistency learning and identity- aware modality adaptation for visible-infrared cross modal- ity person re-identification

    Zhiwei Zhao, Bin Liu, Qi Chu, Yan Lu, and Nenghai Yu. Joint color-irrelevant consistency learning and identity- aware modality adaptation for visible-infrared cross modal- ity person re-identification. In Proceedings of the AAAI con- ference on artificial intelligence (AAAI) ,...

  48. [56]

    Semi-supervised visible-infrared person re-identification via modality unification and confidence guidance

    Xiying Zheng, Yukang Zhang, Yang Lu, and Hanzi Wang. Semi-supervised visible-infrared person re-identification via modality unification and confidence guidance. In Proceed- ings of the 32nd ACM International Conference on Multime- dia (ACMMM), pages 5761–5770, 2024. 1, 3

    2024

  49. [57]

    Video-based visible-infrared person re- identification via style disturbance defense and dual interac- tion

    Chuhao Zhou, Jinxing Li, Huafeng Li, Guangming Lu, Yong Xu, and Min Zhang. Video-based visible-infrared person re- identification via style disturbance defense and dual interac- tion. In Proceedings of the 31st ACM International Confer- ence on Multimedia (ACMMM), page 46–55, 2023. 6

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.