Pith. sign in

REVIEW 4 major objections 4 minor 78 references

Self-Reinforcing Prototype Evolution with Dual-Knowledge Cooperation for Semi-Supervised Lifelong Person Re-Identification

T0 review · 4 major / 4 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 A self-reinforcing cycle between prototype-generated pseudo-labels and new-old knowledge agreement allows person re-identification models to keep learning from mostly unlabeled lifelong data streams, beating prior methods at 10% label…

desk verdict A practical new Semi-LReID setting with a strong method and large low-label gains, but the self-reinforcing loop's stability and benchmark-tuned hyperparameters need more evidence before the exact numbers are bankable. read the letter →

arxiv 2507.01884 v2 pith:O4RIWBAC submitted 2025-07-02 cs.CV

Kunlun Xu , Fan Zhuo , Jiangmeng Li , Xu Zou , Jiahuan Zhou This is my paper
classification cs.CV
keywords semi-supervisedlifelongpersonre-identificationlearnableidentityprototypespseudo-labelpurificationdual-knowledgecooperationcatastrophicforgettingexemplar-freecontinuallearningneighborprototypelabeling
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 introduces Semi-LReID, a lifelong person re-identification setting in which each incoming dataset carries only a small labeled subset alongside a large unlabeled stream. It argues that prior lifelong methods and semi-supervised add-ons degrade sharply in this regime because pseudo-labels generated from scarce labels are noisy, and that noise accumulates as the model moves across domains. To fix this, SPRED creates a self-reinforcing cycle: learnable identity prototypes propose pseudo-labels through a top-2 neighbor scoring rule, a dual-knowledge purifier filters them using clusters from both the current and the historical model, and the purified labels retrain the prototypes. The paper reports state-of-the-art Seen-Avg results at 50%, 20%, and 10% label rates, with the largest margins at the lowest label rates, without storing any historical images.

What carries the argument

The load-bearing components are a set of learnable identity prototypes $P_t$, one per identity, trained with a prototype-orient identity loss plus triplet loss; the Neighbor Prototype Labeling rule, which converts the top-2 prototype similarities into a two-class softmax score $s_A$ and thresholds it; the set-based confidence $LC_s(x,S,C)$, the overlap ratio between a pseudo-label group and a cluster; two complementary clustering views, new-knowledge clustering by the current model and old-knowledge clustering by the previous model on style-aligned data; the union rule $Y^{pse}_t = Y^{pse}_{t-1}\cup Y^{pse}_{e-1}$ that fuses the purified sets; and a prototype-structure KL divergence that keeps the new model's affinity profile over old prototypes close to the old model's.

What would settle it

Evaluate SPRED at the 10% label rate and, each epoch, compute the precision of the purified pseudo-label set against a held-out subset of the unlabeled data whose identities are known. If precision does not increase monotonically, or if the union of the new- and old-model purified sets ever has lower precision than their intersection on the same epoch, the self-reinforcing and complementarity claims are contradicted.

Watch

Extended reading notes

Core claim

The central claim is that a self-reinforcing loop between prototype-generated pseudo-labels and new-old knowledge agreement can tame the noisy-label problem that blocks semi-supervised lifelong person re-identification. The paper shows that neighbor prototype labeling, which scores each unlabeled image by the softmax ratio of its top-2 nearest prototypes, produces better candidate labels than standard score-thresholding. Those candidates are then purified by a set-based label confidence, $LC_s(x,S,C)=|S\cap C|/|S\cup C|$, computed under two independent clusterings: one from the current model and one from the previous model applied after a style-alignment network. The union of the two purified sets outperforms either alone or their intersection, which the paper reads as evidence that new and old knowledge carry complementary errors. A prototype-structure KL loss against the old model preserves prior-domain structure while the loop trains the new domain.

Load-bearing premise

The loop's stability is the load-bearing premise: each round of filtering must remove more pseudo-label noise than it lets through, so that errors do not accumulate across epochs or across datasets.

Editorial extensions

If this is right

  • At a 10% label rate, SPRED reports 43.2% Seen-Avg mAP and 54.4% R@1, improvements of 15.5 and 17.3 points over the strongest LReID baseline; at 50% the margins are 5.7 and 4.2 points.
  • Pseudo-label prediction accuracy continues to rise with training epochs in SPRED while baseline methods plateau, so the paper's mechanism implies that unlabeled-data knowledge acquisition can be sustained over long streams rather than saturating.
  • Because the method is exemplar-free, it offers a route to lifelong ReID in privacy-sensitive deployments where storing historical identity images for replay is not acceptable.
  • The union-of-purified-sets result implies that, when two models disagree about which unlabeled samples belong together, treating their agreements as complementary rather than requiring consensus yields more usable training signal.

Reading between the lines

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

  • The agreement principle is not tied to person images: the same prototype-plus-dual-clustering purification could be applied to streaming face, vehicle, or product identification with partial labels, at the cost of re-learning the domain-alignment network.
  • The paper's union-versus-intersection finding suggests a general design rule for noisy pseudo-labels: heterogeneous feature spaces (fresh fine-tuned features versus older, more general features) should be combined by collecting their separately purified endorsements rather than by intersecting them.
  • A natural next step would be to convert the empirical self-reinforcing loop into a convergence statement, e.g., bounding pseudo-label precision as a function of prototype purity and label rate; without such an analysis, the loop's stability rests on the observed accuracy curves.
  • One could stress-test the loop by measuring per-epoch precision on a held-out labeled sample of the unlabeled set; if precision ever falls below the previous epoch, the current filter is admitting noise faster than it removes it.
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 / 4 minor

Summary. The paper introduces Semi-LReID, a setting in which a lifelong person re-identification model must adapt to a stream of new domains where only a fraction of the data is labeled. It proposes SPRED, which combines learnable identity prototypes for training and pseudo-label generation (DPL and NPL) with a dual-knowledge purification step (DKCP) that clusters the unlabeled data using both the new model and the old model and keeps pseudo-labels accepted by either cluster agreement score (Eqs. 8 and 9). A prototype-structure KL distillation loss is added to mitigate forgetting. Experiments on a 5-domain training / 7-domain testing benchmark at 10%, 20%, and 50% label rates compare SPRED with LReID and SSL+LReID baselines; the supplement adds a second training order, a 100% label-rate comparison, and additional ablations.

Significance. If the reported gains are reproducible, this is a useful contribution: it identifies a realistic and under-studied setting, provides a benchmark and promises code, and shows consistent gains over two training orders, with the component ablation moving monotonically in the intended direction. The authors also honestly report the 0.4% Seen-Avg R@1 deficit against HDC at 50% label rate. The main risk is that the load-bearing mechanism, a self-reinforcing loop between pseudo-label generation and purification, is supported by single-draw accuracy curves and thresholds chosen on the evaluation benchmark, without an analysis of error accumulation or a multi-seed evaluation. This is a correctness-risk concern that can be addressed within the manuscript's scope rather than a reason to reject.

major comments (4)
  1. [Sections 3.3, 3.4 and Eq. (8); Figs. 6 and 8] The central claim that the self-reinforcing loop improves pseudo-labels monotonically and prevents error accumulation is supported only by the label-prediction accuracy curves in Fig. 8 and supplement Fig. 10, which appear to come from a single random label draw. Early epochs are exactly the regime the paper itself criticizes in Section 2.2, where prototypes can overfit scarce labeled data; a biased prototype set can then bias the clusters used to purify pseudo-labels in Eq. (8). Since the thresholds Tp, Tc, To and the weight alpha are ablated on the evaluation benchmark (Fig. 6, Section 4.2), the reported margins could partly reflect favorable threshold choices rather than intrinsic loop stability. Please add multi-seed experiments with mean and standard deviation for the main tables, and per-epoch pseudo-label precision and recall against the ground-truth labels of the unlabeled subset (which are available in these datasets), reporting the worst seed separately at the 10% label rate.
  2. [Eq. (9) and Section 3.4] The final pseudo-label set is defined as the union of the new-knowledge and old-knowledge purified sets, so an instance is accepted if either filter passes. This rule is the basis of the claimed complementarity, but the paper does not report the per-source error rates. A noisy pseudo-label that passes only the old model or only the new model is retained, and if such labels persist across datasets the union could accumulate errors even though each individual filter looks reasonable. Please report, per epoch and per dataset, the precision of the new-only, old-only, and intersection subsets, and confirm that the union's advantage over intersection is not driven by a single domain or by early epochs when both models are nearly identical.
  3. [Algorithm 1 and Section 3.4] Algorithm 1 initializes the new model at each training step as M_t^0 = M_{t-1}. Consequently, in the first epoch the features used for new-knowledge clustering and old-knowledge clustering are extracted by the same model, differing only in the input style alignment via DANet; the two knowledge sources are therefore not independent at the first epoch. The paper does not discuss how the complementarity asserted in Eq. (9) emerges in this first epoch or whether the union is meaningful before the new model has diverged from the old one. Please clarify the initial-state behavior and report the NKC and OKAC pseudo-label precision separately for the first few epochs.
  4. [Supplement Table 5 and Table 2] The ablations show that OKAC and NKC each add only about 1 to 3 mAP over the DPL+NPL baseline, while the full DKCP adds about 5 mAP, so the bulk of the double-knowledge gain comes from the combination. Given the union rule of Eq. (9), this combination effect is the key evidence for the dual-knowledge claim, but the main paper only reports the full ablation (Table 2). Please move the per-component DKCP decomposition into the main paper and state the threshold choices and clustering parameters used for the matched comparison.
minor comments (4)
  1. [Section 4.4] The sentence reporting the 50% comparison with HDC should read "a 2.6% increase in Seen-Avg mAP but a 0.4% decrease in Seen-Avg R@1" for consistency with the metric names.
  2. [Eq. (6) and Section 3.3] The pseudo-label value N/A is used for samples that do not pass the neighbor-prototype threshold, but the paper does not explicitly state how these samples are excluded from Lp and Ltri in Eq. (2); please clarify the treatment of discarded samples in the training loss.
  3. [Figures 4, 5, and 6] In the version of the manuscript provided for review, the axis labels of Figures 4, 5, and 6 are not legible; if this is not an artifact of the review copy, the figures should be regenerated with readable labels.
  4. [Supplement Section 11] The 100% label-rate comparison in supplement Table 7 shows that SPRED without the SSL augmentation strategy is not consistently better than DKP, and the paper's explanation is qualitative; please add a short quantitative discussion of the label-rate dependence of the prototype distribution modeling.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: target metrics are not inputs; self-citations are concentrated but not load-bearing.

full rationale

SPRED's reported quantities (Seen-Avg and UnSeen-Avg mAP/R@1) are never used as training inputs or as fitting targets: the pipeline in Eqs. (1)-(10) takes only labeled and unlabeled images, prototype-derived pseudo-labels, and cluster memberships, and the comparisons are against external and prior methods. The self-reinforcing loop (Eqs. 6-10) is an iterative bootstrap, not a derivation in which the conclusion is an input by construction; its stability is asserted empirically via Fig. 8 and the ablations, and the absence of multi-seed analysis is a robustness concern rather than circularity. Hyperparameters (alpha, Tc, To, Tp) are tuned on the benchmark in Fig. 6, but they are not fitted to the exact predicted mAP/R@1 values and sensitivity is reported, which is a standard model-selection caveat rather than a definitional reduction. Refs [53]-[55] are self-citations and supply the base model and problem framing, but the novel DKCP/SPE equations are specified in the paper and the method is evaluated against those same works as baselines; no load-bearing claim is justified solely by a self-citation and no uniqueness theorem or prior ansatz is imported to force the design. Therefore no circular step is identifiable.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new conceptual entities are introduced: the learnable prototypes are the same class of parameter set as the authors' prior DKP work, and the DANet is a color-style reconstruction network from reference [18]. The load-bearing inputs the reader does not see specified are the tuned thresholds and the cluster-identity alignment assumption.

free parameters (5)
  • Alpha (weight of structure-preserving loss Ls) = 4.0
    Balances new-knowledge learning against anti-forgetting in Eq. 10; selected by grid search (Fig. 6a) on the evaluation benchmark.
  • Tc (new-knowledge cluster agreement threshold) = 0.1
    Filters pseudo-labels by agreement with new-model clustering (Eq. 8); grid-searched on the benchmark (Fig. 6b).
  • To (old-knowledge cluster agreement threshold) = 0.6
    Filters pseudo-labels by agreement with old-model clustering; grid-searched on the benchmark (Fig. 6c).
  • Tp (neighbor prototype labeling threshold) = 0.7
    Confidence gate on the top-2 prototype softmax (Eqs. 5 and 6); grid-searched on the benchmark (Fig. 6d).
  • DBSCAN parameters (eps, min_samples) = not reported in main text
    Both clustering modules (OKAC, NKC) invoke DBSCAN via reference [18] without stating the parameters, so the cluster granularity is an unreported tuned input.
assumptions (5)
  • domain assumption Identity label sets are disjoint across training datasets (Y_l^t ∩ Y_l^s = ∅)
    Stated in Section 3.1; keeps prototypes and pseudo-labels from colliding across steps and is inherited from the standard LReID protocol.
  • domain assumption Color-style alignment makes old-model features usable on new domains
    OKAC (Section 3.4) clusters DANet-transferred new data with the old model; if the domain gap is semantic rather than stylistic, the clusters are unreliable. Supplement Table 3 shows only 0.6 to 1.0 percent contribution.
  • domain assumption DBSCAN cluster membership is a valid proxy for identity agreement
    The LCs score (Eq. 7) treats cluster-set agreement with pseudo-label sets as label correctness; this requires cluster granularity to match identities, which is not guaranteed for arbitrary eps and min_samples.
  • ad hoc to paper Union of new and old purified label sets beats either alone or their intersection
    Eq. 9 is justified only by an experimental finding (Section 3.4); no error-correlation analysis supports union over intersection under correlated clustering errors.
  • domain assumption The top-2 prototype softmax score is a calibrated confidence estimate
    NPL (Eq. 5) uses the two-class softmax over the nearest two prototypes as a certainty gate; this is a heuristic from SSL practice, not a calibrated probability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Reinforcing Prototype Evolution with Dual-Knowledge Cooperation for Semi-Supervised Lifelong Person Re-Identification." pith.science (2026). https://pith.science/paper/O4RIWBAC

@misc{pith2026250701884,
  author       = {Pith},
  title        = {Pith review of: Self-Reinforcing Prototype Evolution with Dual-Knowledge Cooperation for Semi-Supervised Lifelong Person Re-Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O4RIWBAC}},
  note         = {Machine review of arXiv:2507.01884}
}
read the original abstract

Current lifelong person re-identification (LReID) methods predominantly rely on fully labeled data streams. However, in real-world scenarios where annotation resources are limited, a vast amount of unlabeled data coexists with scarce labeled samples, leading to the Semi-Supervised LReID (Semi-LReID) problem where LReID methods suffer severe performance degradation. Existing LReID methods, even when combined with semi-supervised strategies, suffer from limited long-term adaptation performance due to struggling with the noisy knowledge occurring during unlabeled data utilization. In this paper, we pioneer the investigation of Semi-LReID, introducing a novel Self-Reinforcing Prototype Evolution with Dual-Knowledge Cooperation framework (SPRED). Our key innovation lies in establishing a self-reinforcing cycle between dynamic prototype-guided pseudo-label generation and new-old knowledge collaborative purification to enhance the utilization of unlabeled data. Specifically, learnable identity prototypes are introduced to dynamically capture the identity distributions and generate high-quality pseudo-labels. Then, the dual-knowledge cooperation scheme integrates current model specialization and historical model generalization, refining noisy pseudo-labels. Through this cyclic design, reliable pseudo-labels are progressively mined to improve current-stage learning and ensure positive knowledge propagation over long-term learning. Experiments on the established Semi-LReID benchmarks show that our SPRED achieves state-of-the-art performance. Our source code is available at https://github.com/zhoujiahuan1991/ICCV2025-SPRED

Figures

Figures reproduced from arXiv: 2507.01884 by the authors.

Figure 1
Figure 1. (a) In Semi-LReID, the training data is continuously [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Given the labeled and unlabeled data of the new do [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overview of our SPRED. Given a dataset [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Seen domain performance curves under 10% label rate. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Ablation studies on hyperparameters under 10% label rate. The default values are highlighted by the dashed lines. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Label prediction accuracy of different methods. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 7
Figure 7. Figure 7: T-SNE visualization of labeled and unlabeled features in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Training process of distribution alignment network [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Label prediction accuracy of different methods. The compared methods are trained on Training Order-1 with label rate [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Visualization of test feature distribution. The com [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 77 canonical work pages

  1. [1]

    Mixmatch: A holistic approach to semi-supervised learning

    David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. NeurIPS, 32,

    work page

  2. [2]

    Hy- pernetworks for continual semi-supervised learning

    Dhanajit Brahma, Vinay Kumar Verma, and Piyush Rai. Hy- pernetworks for continual semi-supervised learning. arXiv preprint arXiv:2110.01856, 2021. 3

    arXiv 2021

  3. [3]

    Imagery li- brary for intelligent detection systems (i-lids)

    Home Office Scientific Development Branch. Imagery li- brary for intelligent detection systems (i-lids). In 2006 IET conference on crime and security, pages 445–448. IET, 2006. 5, 13, 14

    work page 2006

  4. [4]

    Exponential moving average normalization for self-supervised and semi- supervised learning

    Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Char- less Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi- supervised learning. In CVPR, pages 194–203. IEEE, 2021. 5

    work page 2021

  5. [5]

    Learn from yesterday: A semi-supervised continual learning method for supervision- limited text-to-sql task streams

    Yongrui Chen, Xinnan Guo, Tongtong Wu, Guilin Qi, Yang Li, and Yang Dong. Learn from yesterday: A semi-supervised continual learning method for supervision- limited text-to-sql task streams. In AAAI, pages 12682– 12690, 2023. 4

    work page 2023

  6. [6]

    Boosting semi- supervised learning by exploiting all unlabeled data

    Yuhao Chen, Xin Tan, Borui Zhao, Zhaowei Chen, Ren- jie Song, Jiajun Liang, and Xuequan Lu. Boosting semi- supervised learning by exploiting all unlabeled data. In CVPR, pages 7548–7557, 2023. 3

    work page 2023

  7. [7]

    Dma: Dual modality-aware alignment for visible-infrared person re- identification

    Zhenyu Cui, Jiahuan Zhou, and Yuxin Peng. Dma: Dual modality-aware alignment for visible-infrared person re- identification. T-IFS, 19:2696–2708, 2024. 1

    work page 2024

  8. [8]

    Learning continual compatible representation for re-indexing free lifelong person re-identification

    Zhenyu Cui, Jiahuan Zhou, Xun Wang, Manyu Zhu, and Yuxin Peng. Learning continual compatible representation for re-indexing free lifelong person re-identification. In CVPR, pages 16614–16623, 2024. 1

    work page 2024

Show all 78 references
  1. [9]

    Dkc: Differ- entiated knowledge consolidation for cloth-hybrid lifelong person re-identification

    Zhenyu Cui, Jiahuan Zhou, and Yuxin Peng. Dkc: Differ- entiated knowledge consolidation for cloth-hybrid lifelong person re-identification. In CVPR, pages 3573–3582, 2025. 1

    2025

  2. [10]

    Dynamic sub-graph distillation for robust semi-supervised continual learning

    Yan Fan, Yu Wang, Pengfei Zhu, and Qinghua Hu. Dynamic sub-graph distillation for robust semi-supervised continual learning. In AAAI, pages 11927–11935, 2024. 2, 3

    2024

  3. [11]

    Semi-supervised learning made simple with self-supervised clustering

    Enrico Fini, Pietro Astolfi, Karteek Alahari, Xavier Alameda-Pineda, Julien Mairal, Moin Nabi, and Elisa Ricci. Semi-supervised learning made simple with self-supervised clustering. In CVPR, pages 3187–3197, 2023. 3, 4

    2023

  4. [12]

    Domain adaptive learning with multi-granularity features for unsupervised person re- identification

    Lihua Fu, Yubin Du, Yu Ding, Dan Wang, Hanxu Jiang, and Haitao Zhang. Domain adaptive learning with multi-granularity features for unsupervised person re- identification. Chinese Journal of Electronics , 31(1):116– 128, 2022. 1

    2022

  5. [13]

    Lifelong person re-identification by pseudo task knowledge preservation

    Wenhang Ge, Junlong Du, Ancong Wu, Yuqiao Xian, Ke Yan, Feiyue Huang, and Wei-Shi Zheng. Lifelong person re-identification by pseudo task knowledge preservation. In AAAI, pages 688–696, 2022. 1

    2022

  6. [14]

    Exemplar-free continual representation learning via learn- able drift compensation

    Alex Gomez-Villa, Dipam Goswami, Kai Wang, Andrew D Bagdanov, Bartlomiej Twardowski, and Joost van de Weijer. Exemplar-free continual representation learning via learn- able drift compensation. In ECCV, pages 473–490. Springer,

  7. [15]

    Cross-modality perturbation synergy attack for person re-identification

    Yunpeng Gong, Zhun Zhong, Yansong Qu, Zhiming Luo, Rongrong Ji, and Min Jiang. Cross-modality perturbation synergy attack for person re-identification. 1

  8. [16]

    Person re- identification method based on color attack and joint de- fence

    Yunpeng Gong, Liqing Huang, and Lifei Chen. Person re- identification method based on color attack and joint de- fence. In CVPR, pages 4313–4322, 2022. 1

    2022

  9. [17]

    Viewpoint invariant pedestrian recognition with an ensemble of localized features

    Douglas Gray and Hai Tao. Viewpoint invariant pedestrian recognition with an ensemble of localized features. InECCV, pages 262–275. Springer, 2008. 5, 13, 14

    2008

  10. [18]

    Color prompting for data-free continual un- supervised domain adaptive person re-identification

    Jianyang Gu, Hao Luo, Kai Wang, Wei Jiang, Yang You, and Jian Zhao. Color prompting for data-free continual un- supervised domain adaptive person re-identification. arXiv preprint arXiv:2308.10716, 2023. 5, 12

    2023 arXiv

  11. [19]

    Transreid: Transformer-based object re- identification

    Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re- identification. In ICCV, pages 14993–15002. IEEE, 2021. 1

    2021

  12. [20]

    Approximating the kull- back leibler divergence between gaussian mixture models

    John R Hershey and Peder A Olsen. Approximating the kull- back leibler divergence between gaussian mixture models. In ICASSP, pages IV–317. IEEE, 2007. 4

    2007

  13. [21]

    Person re-identification by descriptive and dis- criminative classification

    Martin Hirzer, Csaba Beleznai, Peter M Roth, and Horst Bischof. Person re-identification by descriptive and dis- criminative classification. In Image Analysis, pages 91–102. Springer, 2011. 5, 13, 14

    2011

  14. [22]

    Searching for mo- bilenetv3

    Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mo- bilenetv3. In ICCV, pages 1314–1324. IEEE, 2019. 12

    2019

  15. [23]

    A soft nearest-neighbor framework for contin- ual semi-supervised learning

    Zhiqi Kang, Enrico Fini, Moin Nabi, Elisa Ricci, and Kar- teek Alahari. A soft nearest-neighbor framework for contin- ual semi-supervised learning. In ICCV, pages 11868–11877,

  16. [24]

    Cdmad: Class- distribution-mismatch-aware debiasing for class-imbalanced semi-supervised learning

    Hyuck Lee and Heeyoung Kim. Cdmad: Class- distribution-mismatch-aware debiasing for class-imbalanced semi-supervised learning. In CVPR, pages 23891–23900. IEEE, 2024. 6, 15

    2024

  17. [25]

    Comatch: Semi-supervised learning with contrastive graph regulariza- tion

    Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regulariza- tion. In ICCV, pages 9475–9484, 2021. 3

    2021

  18. [26]

    Exemplar-free lifelong person re-identification via prompt- guided adaptive knowledge consolidation.IJCV, pages 1–16,

    Qiwei Li, Kunlun Xu, Yuxin Peng, and Jiahuan Zhou. Exemplar-free lifelong person re-identification via prompt- guided adaptive knowledge consolidation.IJCV, pages 1–16,

  19. [27]

    Locally aligned feature trans- forms across views

    Wei Li and Xiaogang Wang. Locally aligned feature trans- forms across views. In CVPR, pages 3594–3601. IEEE,

  20. [28]

    Human reidentifica- tion with transferred metric learning

    Wei Li, Rui Zhao, and Xiaogang Wang. Human reidentifica- tion with transferred metric learning. InACCV, pages 31–44. Springer, 2012. 5, 13, 14

    2012

  21. [29]

    Deep- reid: Deep filter pairing neural network for person re- identification

    Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deep- reid: Deep filter pairing neural network for person re- identification. In CVPR, pages 152–159. IEEE, 2014. 5, 13, 14

    2014

  22. [30]

    A bottom-up clustering approach to unsupervised per- 9 son re-identification

    Yutian Lin, Xuanyi Dong, Liang Zheng, Yan Yan, and Yi Yang. A bottom-up clustering approach to unsupervised per- 9 son re-identification. In AAAI, pages 8738–8745, 2019. 2, 3

    2019

  23. [31]

    Certainty driven consistency loss on multi-teacher networks for semi-supervised learning.Pat- tern Recognition, 120:108140, 2021

    Lu Liu and Robby T Tan. Certainty driven consistency loss on multi-teacher networks for semi-supervised learning.Pat- tern Recognition, 120:108140, 2021. 3

    2021

  24. [32]

    Time- delayed correlation analysis for multi-camera activity under- standing

    Chen Change Loy, Tao Xiang, and Shaogang Gong. Time- delayed correlation analysis for multi-camera activity under- standing. IJCV, 90(1):106–129, 2010. 5, 13, 14

    2010

  25. [33]

    Learning to predict gradients for semi-supervised con- tinual learning

    Yan Luo, Yongkang Wong, Mohan Kankanhalli, and Qi Zhao. Learning to predict gradients for semi-supervised con- tinual learning. TNNLS, 2024. 3

    2024

  26. [34]

    Lifelong person re-identification via adaptive knowl- edge accumulation

    Nan Pu, Wei Chen, Yu Liu, Erwin M Bakker, and Michael S Lew. Lifelong person re-identification via adaptive knowl- edge accumulation. In CVPR, pages 7897–7906. IEEE,

  27. [35]

    Meta reconciliation normalization for lifelong person re-identification

    Nan Pu, Yu Liu, Wei Chen, Erwin M Bakker, and Michael S Lew. Meta reconciliation normalization for lifelong person re-identification. In ACM MM, pages 541–549, 2022. 3

    2022

  28. [36]

    A memorizing and generalizing framework for lifelong person re-identification

    Nan Pu, Zhun Zhong, Nicu Sebe, and Michael S Lew. A memorizing and generalizing framework for lifelong person re-identification. IEEE TPAMI, 45(11):13567–13585, 2023. 3, 13

    2023

  29. [37]

    Regularization with stochastic transformations and perturba- tions for deep semi-supervised learning

    Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturba- tions for deep semi-supervised learning. NeurIPS, 29, 2016. 3

    2016

  30. [38]

    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. InICCV, pages 11218–11228. IEEE, 2023. 2, 3, 5, 6, 15

    2023

  31. [39]

    Learning commonality, divergence and variety for unsupervised visible-infrared person re- identification

    Jiangming Shi, Xiangbo Yin, Yachao Zhang, Yuan Xie, Yanyun Qu, et al. Learning commonality, divergence and variety for unsupervised visible-infrared person re- identification. NeurIPS, 37:99715–99734, 2024. 1

    2024

  32. [40]

    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 ECCV, pages 456–474. Springer, 2025. 2

    2025

  33. [41]

    Memory-efficient semi-supervised continual learning: The world is its own replay buffer

    James Smith, Jonathan Balloch, Yen-Chang Hsu, and Zsolt Kira. Memory-efficient semi-supervised continual learning: The world is its own replay buffer. In IJCNN, pages 1–8. IEEE, 2021. 3

    2021

  34. [42]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 33:596–608, 2020. 2, 3

    2020

  35. [43]

    Region-based quality estimation network for large-scale person re-identification

    Guanglu Song, Biao Leng, Yu Liu, Congrui Hetang, and Shaofan Cai. Region-based quality estimation network for large-scale person re-identification. In AAAI, 2018. 5, 13, 14

    2018

  36. [44]

    Marginmatch: Im- proving semi-supervised learning with pseudo-margins

    Tiberiu Sosea and Cornelia Caragea. Marginmatch: Im- proving semi-supervised learning with pseudo-margins. In CVPR, pages 15773–15782, 2023. 3

    2023

  37. [45]

    Patch-based knowledge dis- tillation for lifelong person re-identification

    Zhicheng Sun and Yadong Mu. Patch-based knowledge dis- tillation for lifelong person re-identification. In ACM MM, pages 696–707, 2022. 5, 6, 15

    2022

  38. [46]

    Ordisco: Effective and efficient usage of incremental unlabeled data for semi-supervised continual learning

    Liyuan Wang, Kuo Yang, Chongxuan Li, Lanqing Hong, Zhenguo Li, and Jun Zhu. Ordisco: Effective and efficient usage of incremental unlabeled data for semi-supervised continual learning. In CVPR, pages 5383–5392, 2021. 3

    2021

  39. [47]

    Persistence homology distillation for semi-supervised con- tinual learning

    Yu Wang, Pengfei Zhu, Dongyue Chen, and Qinghua Hu. Persistence homology distillation for semi-supervised con- tinual learning. NeurIPS, 37:76332–76355, 2024. 2

    2024

  40. [48]

    Person transfer gan to bridge domain gap for person re- identification

    Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re- identification. In CVPR, pages 79–88. IEEE, 2018. 5, 13, 14

    2018

  41. [49]

    Semi- supervised learning with heterogeneous distribution consis- tency for visible infrared person re-identification

    Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Semi- supervised learning with heterogeneous distribution consis- tency for visible infrared person re-identification. IEEE TIP,

  42. [50]

    Rewarded semi-supervised re-identification on identities rarely cross- ing camera views

    Ancong Wu, Wenhang Ge, and Wei-Shi Zheng. Rewarded semi-supervised re-identification on identities rarely cross- ing camera views. IEEE TPAMI, 2023. 2, 3

    2023

  43. [51]

    Generalising without for- getting for lifelong person re-identification

    Guile Wu and Shaogang Gong. Generalising without for- getting for lifelong person re-identification. In AAAI, pages 2889–2897, 2021. 1, 2

    2021

  44. [52]

    End-to-end deep learning for person search

    Tong Xiao, Shuang Li, Bochao Wang, Liang Lin, and Xi- aogang Wang. End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850, 2(2):4, 2016. 5, 13, 14

    2016 arXiv

  45. [53]

    Mitigate catastrophic remembering via con- tinual knowledge purification for noisy lifelong person re- identification

    Kunlun Xu, Haozhuo Zhang, Yu Li, Yuxin Peng, and Ji- ahuan Zhou. Mitigate catastrophic remembering via con- tinual knowledge purification for noisy lifelong person re- identification. In ACM MM, pages 1–9, 2024. 1, 2, 3, 5, 16

    2024

  46. [54]

    Distribution-aware knowledge prototyping for non-exemplar lifelong person re-identification

    Kunlun Xu, Xu Zou, Yuxin Peng, and Jiahuan Zhou. Distribution-aware knowledge prototyping for non-exemplar lifelong person re-identification. In CVPR, pages 16604– 16613. IEEE, 2024. 1, 2, 3, 5, 6, 14, 15

    2024

  47. [55]

    Lstkc: Long short-term knowledge consolidation for lifelong person re- identification

    Kunlun Xu, Xu Zou, and Jiahuan Zhou. Lstkc: Long short-term knowledge consolidation for lifelong person re- identification. In AAAI, pages 16202–16210, 2024. 1, 3, 4, 5, 6, 13, 15

    2024

  48. [56]

    Dask: Distribution rehearsing via adaptive style kernel learning for exemplar-free lifelong person re- identification

    Kunlun Xu, Chenghao Jiang, Peixi Xiong, Yuxin Peng, and Jiahuan Zhou. Dask: Distribution rehearsing via adaptive style kernel learning for exemplar-free lifelong person re- identification. In AAAI, pages 8915–8923, 2025. 2

    2025

  49. [57]

    Long short-term knowledge decomposition and con- solidation for lifelong person re-identification.IEEE TPAMI,

    Kunlun Xu, Zichen Liu, Xu Zou, Yuxin Peng, and Jiahuan Zhou. Long short-term knowledge decomposition and con- solidation for lifelong person re-identification.IEEE TPAMI,

  50. [58]

    Compo- nential prompt-knowledge alignment for domain incremen- tal learning

    Kunlun Xu, Xu Zou, Gang Hua, and Jiahuan Zhou. Compo- nential prompt-knowledge alignment for domain incremen- tal learning. In ICML, 2025. 1

    2025

  51. [59]

    Shrinking class space for enhanced cer- tainty in semi-supervised learning

    Lihe Yang, Zhen Zhao, Lei Qi, Yu Qiao, Yinghuan Shi, and Hengshuang Zhao. Shrinking class space for enhanced cer- tainty in semi-supervised learning. In ICCV, pages 16187– 16196, 2023. 6, 7, 15

    2023

  52. [60]

    Handling label uncertainty for camera incremen- tal person re-identification

    Zexian Yang, Dayan Wu, Wanqian Zhang, Bo Li, and Weip- inng Wang. Handling label uncertainty for camera incremen- tal person re-identification. In ACM MM, page 6253–6263,

  53. [61]

    Unsupervised cross-media hashing learning via knowledge graph

    Zhaoda Ye, Xiangteng He, and Yuxin Peng. Unsupervised cross-media hashing learning via knowledge graph. Chinese Journal of Electronics, 31(6):1081–1091, 2022. 1

    2022

  54. [62]

    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 ACM MM, pages 2242–2251, 2024. 1

    2024

  55. [63]

    Lifelong person re-identification via knowledge re- freshing and consolidation

    Chunlin Yu, Ye Shi, Zimo Liu, Shenghua Gao, and Jingya Wang. Lifelong person re-identification via knowledge re- freshing and consolidation. In AAAI, pages 3295–3303,

  56. [64]

    S4l: Self-supervised semi-supervised learning

    Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lu- cas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, pages 1476–1485, 2019. 3

    2019

  57. [65]

    Scap: Transductive test-time adaptation via supportive clique-based attribute prompting

    Chenyu Zhang, Kunlun Xu, Zichen Liu, Yuxin Peng, and Jiahuan Zhou. Scap: Transductive test-time adaptation via supportive clique-based attribute prompting. InCVPR, pages 30032–30041, 2025. 1

    2025

  58. [66]

    Spindle net: Person re-identification with human body region guided feature decomposition and fusion

    Haiyu Zhao, Maoqing Tian, Shuyang Sun, Jing Shao, Junjie Yan, Shuai Yi, Xiaogang Wang, and Xiaoou Tang. Spindle net: Person re-identification with human body region guided feature decomposition and fusion. In CVPR, pages 907–915. IEEE, 2017. 5, 13, 14

    2017

  59. [67]

    Scalable person re-identification: A benchmark

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jing- dong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, pages 1116–1124. IEEE, 2015. 5, 13, 14

    2015

  60. [68]

    Simmatch: Semi-supervised learn- ing with similarity matching

    Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learn- ing with similarity matching. In CVPR, pages 14471–14481,

  61. [69]

    Simmatchv2: Semi- supervised learning with graph consistency

    Mingkai Zheng, Shan You, Lang Huang, Chen Luo, Fei Wang, Chen Qian, and Chang Xu. Simmatchv2: Semi- supervised learning with graph consistency. In CVPR, pages 16432–16442. IEEE, 2023. 6, 15

    2023

  62. [70]

    Camera style adaptation for person re- identification

    Zhun Zhong, Liang Zheng, Zhedong Zheng, Shaozi Li, and Yi Yang. Camera style adaptation for person re- identification. In CVPR, pages 5157–5166. IEEE, 2018. 1 11 Self-Reinforcing Prototype Evolution with Dual-Knowledge Cooperation for Semi-Supervised Lifelong Person Re-Identif...

    2018

  63. [71]

    As illustrated in Fig

    Implementation of Distribution Alignment Network Inspired by [18], which revealed that the primary distribu- tion difference across ReID datasets lies in color styles, we propose a distribution alignment network (DANet) based on the architecture of [18]. As illustrated in Fig....

  64. [72]

    Ablation Studies on DKCP Our Dual-Knowledge Cooperation-driven Pseudo-label Pu- rification (DKCP) module is designed to filter noisy pseudo- labels by leveraging both old and new model knowledge. To achieve this, two complementary strategies are introduced: Old Knowledge Align...

  65. [73]

    Algorithm The pseudo-code of our SPRED method is shown in Alg 1

  66. [74]

    The de- tailed dataset statistics are presented in Tab

    Datasets Details of Semi-LReID Benchmark We establish the Semi-LReID benchmark based on the ex- isting LReID configuration[34], incorporating 12 widely- used ReID datasets: Market1501 [67], LPW [43], CUHK-SYSU [52], MSMT17-V2 [48], CUHK03 [29], CUHK01 [28], CUHK02 [27], VIPeR ...

  67. [75]

    Experimental Results on Different Dataset Orders In our main paper, the default training dataset order follows Market-1501 →CUHK-SYSU→LPW→MSMT17 →CUHK03, denoted as Training Training Order-1 1. In this section, we extend the evaluation by compar- ing our method with existing a...

  68. [76]

    As shown in Tab

    Experimental Results Under Label Rate r=100% We also evaluate our method under the fully-labeled sce- nario, where the label rater=100%. As shown in Tab. 7, our SPRED achieves competitive results with the latest LReID methods across both training orders. Additionally, the en- ...

  69. [77]

    Besides, the exemplar-free SSLL method LDC [14] is also compared (with sequential fine-tune as the baseline)

    Quantitative comparisons with state-of- the-art NoisyLReID and SSLL methods We also compare our proposed method with the latest Noisy Lifelong Person Re-Identification (NoisyLReID) method CKP [53] by assigning all the unlabeled data with the same label, serving as noisy data. ...

  70. [78]

    Be- yond the prediction accuracy curves presented in the main paper, we provide additional visualizations for CUHK- SYSU, LPW, and CUHK03 datasets in Fig

    Additional Visualization on Prediction and Features More Identity Prediction Capacity Visualization . Be- yond the prediction accuracy curves presented in the main paper, we provide additional visualizations for CUHK- SYSU, LPW, and CUHK03 datasets in Fig. 10. The re- sults sh...

Pith tools

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