Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Person re-identification fails when neighbors are noisy; a new method adds negative weights to distinguish true from false matches.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:30 UTC pith:7KSYSZEN

load-bearing objection Solid extension of neighbor-based Re-ID with a plausible new mechanism; the headline seven-setting robustness claim needs a fixed-k rerun before I'd fully buy it. the 3 major comments →

arxiv 2607.25407 v1 pith:7KSYSZEN submitted 2026-07-28 cs.CV

ANFI: Rethinking Neighbor Feature Interaction in Person Re-ID

Xulin Li , Yan Lu , Bin Liu , Jiaze Li , Qinhong Yang , Tao Gong , Qi Chu , Nenghai Yu This is my paper
classification cs.CV
keywords person re-identificationfeature interactionnoisy neighborsaffinity relationsdiscrepancy relationsadaptive weightingnoise simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that existing neighbor-based person re-identification methods rely only on affinity (similarity) relations, which works when neighbors are reliable but fails when noisy neighbors are present. The authors propose ANFI, which jointly models affinity and discrepancy relations and adaptively mixes them per sample, allowing the model to push away false positive neighbors. They also introduce Noisy Relation Supervision (NRS) to train robustness to noisy relations. The paper argues that ANFI is the first Re-ID method to explicitly capture negative influences of noisy neighbors, and reports consistent gains across standard, cross-modal, and cross-domain settings.

Core claim

The central claim is that neighbor feature interaction in person Re-ID should not only aggregate features of similar neighbors (affinity relations) but also explicitly model differences from unreliable neighbors (discrepancy relations). When a sample's neighborhood is noisy, affinity-only interaction mistakenly pulls the sample toward false positives. ANFI uses a negative-weight discrepancy term, derived from a neighborhood similarity that leverages shared neighbors (third-party samples), and learns sample-wise weights to balance affinity and discrepancy. With NRS, which simulates noisy neighborhoods during training and regularizes the relation outputs, the model becomes robust across varyin

What carries the argument

The key mechanism is the discrepancy relation matrix learned via neighborhood similarity (E'ij = sum over shared neighbors of S'ik * S'jk), combined with adaptive mixing weights (softmax over affinity and discrepancy features). This allows the model to assign negative weights to non-self neighbors when they are unreliable, effectively computing feature differences (phi'(x_i) - phi'(x_j)) to push away false positives.

Load-bearing premise

The analysis assumes that when neighbors are unreliable, the correct update is to assign negative weights to all non-self neighbors and to use the feature difference (phi'(x_i) - phi'(x_j)) as the discrepancy signal; this is stated as an analysis conclusion rather than derived from an optimality criterion.

What would settle it

A direct test would be to construct a synthetic dataset where the optimal robustification is known (e.g., a low-dimensional manifold with controlled label noise) and compare ANFI's learned discrepancy weights against the oracle that minimizes a calibrated loss; if ANFI's negative weighting fails to outperform simple trimming of noisy neighbors, the claimed mechanism would be falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If ANFI is correct, neighbor-based Re-ID methods can be made robust to small gallery sizes, hard queries, and cross-domain shifts where noisy neighbors dominate.
  • The method suggests that combining affinity and discrepancy relations is a general principle that could be applied to other retrieval tasks beyond person Re-ID, such as face recognition or vehicle re-identification.
  • The adaptive mixing weights provide a per-sample mechanism to decide when to trust neighbors, potentially replacing hand-tuned neighborhood aggregation strategies.
  • NRS demonstrates that injecting controlled noise during training can bridge the gap between clean training neighborhoods and noisy test neighborhoods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The discrepancy relation could be interpreted as a form of repulsive force in feature space; a testable extension would be to see if the learned weights track an oracle noise ratio in other datasets without supervision.
  • The neighborhood similarity E' implicitly encodes local geometry; one could explore whether higher-order neighborhood overlaps (beyond pairwise) further improve discrepancy modeling.
  • Since NRS uses stop-gradient on the source feature, the noise injection might be seen as a type of self-distillation or data augmentation; it would be interesting to test whether simpler augmentation (e.g., feature mixup) achieves similar robustness.
  • The paper's claim that ANFI is 'the only method with positive ΔmAP in all seven settings' is a strong claim that depends on the chosen baselines and hyperparameter tuning; replicating the benchmark with more recent re-ranking methods would clarify the practical advantage.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper revisits neighbor-based person Re-ID and argues that affinity-only feature interaction is fragile when neighbor sets contain false positives. It proposes ANFI, which augments affinity relations with discrepancy relations built from a new neighborhood similarity (Eq. 10), adaptively mixes affinity and discrepancy features per sample (Eqs. 13–14), and adds Noisy Relation Supervision (NRS) with feature-noise simulation and relationship regularization (Eqs. 15–18). Experiments cover standard, cross-modal, and cross-domain settings, including same-backbone comparisons, ablations, efficiency, and hyperparameter analysis. The central empirical claim is that ANFI is the only method with positive ΔmAP across all seven neighbor-distribution settings in Table 3 (+5.4/+0.9/+2.3/+0.6/+1.1/+3.5/+0.2).

Significance. If the claims hold, ANFI would be a useful step beyond affinity-only neighbor interaction, and the introduction of discrepancy relations plus a noise-supervision training scheme is a plausible way to improve robustness in small-gallery, hard-query, and cross-domain Re-ID. The paper is also strong in scope: it includes five datasets, same-backbone controls, ablation of each component, efficiency measurements, and sensitivity analysis. However, the headline 'positive in all seven settings' rests on per-setting best-k selection and on a comparison protocol whose exact application to ANFI is ambiguous. Those issues need to be resolved before the robustness claim can be accepted.

major comments (3)
  1. [Sec. 5.3 / Sec. 5.7, Table 3] The protocol states that neighbor-related hyper-parameters are adjusted for all methods in each configuration and that 'we report each method with its best k for each dataset and testing mode.' Since Fig. 5(c–e) shows that the optimal k varies with dataset and testing protocol, the reported ΔmAP values are maxima over k per test distribution. At deployment the test distribution is unknown, so the claim that ANFI is the only method with positive ΔmAP in all seven settings does not yet establish robustness to neighbor distributions. Please report fixed-k results (e.g., k = 10 for all methods) and/or a held-out model-selection protocol, and show whether the positive-ΔmAP pattern persists.
  2. [Sec. 5.3, item 2 / Sec. 5.2, Table 2] The text says 'we use the same baseline model for all approaches to extract single-image representations,' but ANFI is trained end-to-end with the identity loss and NRS (Eq. 19), which updates the backbone. It is therefore unclear whether the ANFI rows in Tables 2 and 3 come from a fixed baseline backbone or from ANFI's own trained backbone. If the latter, the comparison is not apples-to-apples with training-free re-ranking and plug-and-play neighbor-based methods. Please clarify the exact protocol, and if ANFI is trained on the same backbone features, state how the ANFI module is trained (frozen backbone vs. end-to-end) and whether the 'Poor Model' early-stopping protocol in Table 3 also applies to ANFI.
  3. [Tables 1 and 3] All results are single-run with no error bars. Several decisive numbers are small (e.g., +0.2 mAP for M→S and +0.6 for Hard Query in Table 3). In this regime, run-to-run variation could change the sign of ΔmAP and invert the 'only method positive' conclusion. Please provide multiple seeds with standard deviations or a statistical test for the headline comparisons, at least for Table 3.
minor comments (5)
  1. [Eq. (7)–(8), Sec. 3.2] The 'bad case' analysis assumes equal negative weights for all non-self neighbors and uses φ'(x_i) − φ'(x_j) as the discrepancy signal. This is presented as an analysis conclusion rather than as a consequence of an optimality criterion. A brief discussion of why this form is preferred over, e.g., magnitude-dependent gating would help.
  2. [Table 5] The row labeled 'AIM [2]' appears to refer to SAAI [2] (Table 1); please correct the label for consistency.
  3. [Table 4] The table header formatting is hard to follow: the NRS columns 'Noise' and 'L_rr' are not visually separated from the ANFI columns, and the row labels are ambiguous. Please reformat for readability.
  4. [Abstract / Sec. 1] The phrase 'the first Re-ID method to capture negative influences of noisy neighbors' is a strong novelty claim. It would be safer to say 'to our knowledge' and to briefly note the closest prior work on negative relation modeling, if any.
  5. [Fig. 2] The panel labels 'Our ANFI', 'Neighbor-based Methods', and 'Ranking Optimization Methods' are placed in a way that may confuse readers; consider adding clear captions or separating the three pipelines spatially.

Circularity Check

1 steps flagged

Central mAP claim is externally benchmarked; main circularity is confined to the NRS/adaptive-weight analysis, where the supervision target is built from the model's own relations and the learned weight is trained to mimic NNR.

specific steps
  1. fitted input called prediction [Sec. 4.2 Eqs. (17)-(18); Sec. 5.5 Fig. 4]
    "Finally, we replace the adaptive weights wi with nnr(xi) for each sample, yielding the final ground truth feature g_mix_i: ... By matching f_mix_i to g_mix_i through L_rr, this supervision propagates the nnr(xi) trend to wi, encouraging larger wi under noisier neighbors."

    Eq. (17) constructs the supervision target from label-derived nnr and from the model's own relation matrices (Eqs. (3), (10)-(12)); L_rr then explicitly trains wi to follow nnr. Therefore Sec. 5.5's observation that the learned wi tracks nnr is a consequence of the objective, not an independent finding, and the 'ground truth' relation feature is a masked, renormalized version of the model's own outputs rather than an external target. This is a contained self-referential loop: final mAP is measured on held-out identities, and oracle-nnr replacement hurts mAP, so the central benchmark claim does not reduce to this fit.

full rationale

The paper's main derivation is self-contained and its principal empirical claims are tested on held-out identities. Eq. (8) is an algebraic rewrite of the negative-weight construction in Eq. (7); the neighborhood similarity in Eq. (10) is a defined structural feature; no uniqueness theorem or load-bearing self-citation is invoked, and the authors' own prior work appears only as a baseline/loss source (CIFT). The per-setting best-k protocol of Table 3 is a validity/robustness concern (test-set hyperparameter selection), but it is not a circularity in the sense of a derivation reducing to its inputs. The one genuinely circular element is the NRS adaptive-weight analysis: because L_rr deliberately trains wi to mimic nnr and builds gmix from the same relation matrices being learned, reporting wi-vs-nnr agreement in Sec. 5.5 and calling gmix 'ground truth' are partially self-referential. Since the central mAP claims are external and the oracle-nnr replacement test still hurts, this is minor and local rather than a collapse of the paper's contribution.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

ANFI is an empirical pipeline that introduces no new physical entities. Its load-bearing assumptions are modeling choices about how to represent and supervise noisy-neighbor relations, plus hyper-parameters tuned on benchmark test sets. The 'discrepancy relation' and 'neighborhood similarity' are mathematical constructs, not independently evidenced entities.

free parameters (4)
  • temperature τ = 0.4
    Used in pairwise and neighborhood similarity (Eqs. 1 and 10); tuned on Market1501 (Fig. 5b). Directly controls the sharpness of relation weights.
  • noise upper limit αmax = 0.4
    Upper bound for the uniform noise-injection distribution in Eq. (16); selected via hyperparameter search on Market1501 (Fig. 5a). Controls the amount of training-time noise.
  • neighborhood size k = dataset/test-mode dependent; exact values not reported
    Determines the neighbor matrix via Top-k (Eq. 2), and thus the affinity and discrepancy graphs. The paper states k was tuned per dataset and test mode for all methods, but does not list the chosen values.
  • 'Poor Model' early-stopping epoch = 30 of 120 epochs
    Defines the 'Poor Model' benchmark in Sec. 5.3; an arbitrary choice that shapes the reported robustness comparison.
axioms (4)
  • domain assumption In the 'bad case', all non-self neighbors should receive negative weights, and the discrepancy update (φ'(x_i) - φ'(x_j)) in Eq. (7)-(8) is the correct robustification.
    Introduced in Sec. 3.2 as an analysis conclusion rather than derived from an optimality criterion. The entire discrepancy-relation design rests on this premise.
  • domain assumption The neighborhood similarity E' in Eq. (10), built from shared neighbors, captures discrepancy information better than pairwise similarity.
    Stated in Step 4-2 without theoretical or independent empirical justification; only the final Re-ID results support it.
  • ad hoc to paper Uniform feature corruption with intensity αmax(1−nnr(X)) in Eqs. (15)-(16) simulates the distribution of noisy neighbors encountered at test time.
    A heuristic training augmentation designed for this paper; no evidence connects this corruption model to real test-time neighbor noise.
  • domain assumption The sample-wise mixing weight w_i can be predicted by two linear layers and learned through KL matching to a label-derived oracle target in Eq. (17)-(18).
    Assumes image features contain enough signal to predict neighbor noise; the paper's Fig. 4 provides partial evidence on Market1501 only.

pith-pipeline@v1.3.0-alltime-deepseek · 14372 in / 11235 out tokens · 119588 ms · 2026-08-01T02:30:16.039879+00:00 · methodology

0 comments
read the original abstract

In person re-identification, neighbor-based methods have achieved significant success by interacting with neighbor samples to obtain more robust representations. However, existing methods rely only on affinity relations, causing their success to depend heavily on the reliability of selected neighbors. We find that affinity-only interaction often fails in challenging scenarios due to the inevitable presence of noisy neighbors. To enable effective interactions under noisy neighborhoods, we revisit neighbor-based methods under distinct reliability conditions and propose a novel Adaptive Neighbor Feature Interaction (ANFI) method. The core idea of ANFI is to account for negative effects from noisy neighbors, allowing samples to remain distinguishable from false positive neighbors. Unlike existing methods, ANFI models not only affinity relations but also discrepancy relations, and employs sample-wise adaptive weighting for these two types of relations. Given that capturing negative effects from noisy neighbors differs significantly from traditional relation learning, we derive discrepancy relations from a new neighborhood similarity, which provides more information than pairwise similarity. In addition, we propose Noisy Relation Supervision (NRS) to train ANFI, gradually injecting robustness to noisy relations into the model. Extensive experiments conducted under standard, cross-modal, and cross-domain settings, including comparisons with neighbor-based methods and re-ranking methods, demonstrate the superiority of our method across various neighbor distributions.

Figures

Figures reproduced from arXiv: 2607.25407 by Bin Liu, Jiaze Li, Nenghai Yu, Qi Chu, Qinhong Yang, Tao Gong, Xulin Li, Yan Lu.

Figure 1
Figure 1. Figure 1: (a) Existing neighbor-based approaches always aggregate neighbor represen￾tations by affinity-only relations. (b) Our approach dynamically incorporates both positive (affinity) and negative (discrepancy) relations, allowing samples to remain distinguishable from false positive neighbors. viewpoint, and resolution, it is hard to extract sufficiently discriminative features from each image alone. To overcome… view at source ↗
Figure 2
Figure 2. Figure 2: Pipelines of ANFI and related methods. where Aˆ ij ≥ 0, N + i = {xj |xj ∈ Ni , yi = yj} is the positive neighbor set, N − i = {xj |xj ∈ Ni , yi ̸= yj} is the negative neighbor set. • Fully Reliable Neighbors (Good Case). When all neighbors are positive samples (i.e., |N − i | = 0), the affinity rela￾tions assign positive weights to positive samples, effectively reducing the distance between positive sample… view at source ↗
Figure 3
Figure 3. Figure 3: The framework of the proposed Adaptive Neighbor Feature Interaction (ANFI) model achieves robust representations through interactions among multiple images. reliance on affinity interaction, whereas a higher nnr(xi) calls for stronger dis￾crepancy interaction. To quantify this trade-off, we define the Noise Neighbor Ratio (NNR) as a sample-level noise indicator: \begin {aligned} \mathrm {nnr}(x_i)=\frac {|… view at source ↗
Figure 4
Figure 4. Figure 4: Distributions of oracle nnr(xi) and learned adaptive weights wi on Market1501 under Standard and Hard Query settings. with using only affinity relations, mAP increases by 1.0% and 2.2% under the ‘1-shot’ and ‘Hard Query’ test modes. Effectiveness of NRS. As shown in the 4th and 5th rows, we analyze the effects of relationship regularization Lrr and noise simulation in NRS. Experi￾mental results indicate th… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of hyperparameters αmax and τ on Market1501, and k on Market1501 and SYSU-MM01. 5.7 Hyperparameter Analysis We analyzed the influences of key hyperparameters, including αmax, τ , and k. As shown in [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE visualization of the distributions of image features. Different colors rep￾resent different identities. similar tuning parameter. The optimal neighborhood size varies with the dataset and testing protocol. In practice, we first determine a coarse candidate range of k according to dataset protocol statistics, including the average number of gallery images per identity, and then tune k for all compared… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of retrieval ranking lists on the MSMT17 dataset. Green/red boxes represent correct/incorrect retrieval results. 6 Conclusion In this paper, we revisit neighbor-based Re-ID under varying neighbor relia￾bility and show that affinity-only interaction is insufficient in the presence of noisy neighbors. Based on this analysis, we propose Adaptive Neighbor Fea￾ture Interaction (ANFI), which jointl… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 1 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, Y., Fan, Z., Chen, Z., Zhu, Y.: Ca-jaccard: camera-aware jaccard distance for person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 17532–17541 (2024)

    2024

  2. [2]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fang, X., Yang, Y., Fu, Y.: Visible-infrared person re-identification via semantic alignment and affinity inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11270–11279 (2023)

    2023

  3. [3]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

    2016

  4. [4]

    In: Proceedings of the IEEE/CVF international con- ference on computer vision

    He, S., Luo, H., Wang, P., Wang, F., Li, H., Jiang, W.: Transreid: Transformer- based object re-identification. In: Proceedings of the IEEE/CVF international con- ference on computer vision. pp. 15013–15022 (2021)

    2021

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    He, W., Deng, Y., Tang, S., Chen, Q., Xie, Q., Wang, Y., Bai, L., Zhu, F., Zhao, R., Ouyang, W., et al.: Instruct-reid: A multi-purpose person re-identification task with instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17521–17531 (2024)

    2024

  6. [6]

    In: European Conference on Computer Vision

    Jiang, Y., Cheng, X., Yu, H., Liu, X., Chen, H., Zhao, G.: Domain shifting: A generalized solution for heterogeneous cross-modality person re-identification. In: European Conference on Computer Vision. pp. 289–306. Springer (2024)

    2024

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Kim, M., Kim, S., Park, J., Park, S., Sohn, K.: Partmix: Regularization strategy to learn part discovery for visible-infrared person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18621– 18632 (2023)

    2023

  8. [8]

    In: European Conference on Computer Vision

    Li,C.,Chen,S.,Ye,M.:Adaptivehigh-frequencytransformerfordiversewildlifere- identification. In: European Conference on Computer Vision. pp. 296–313. Springer (2024)

    2024

  9. [9]

    In: Proceedings of the AAAI conference on artificial intelligence

    Li, S., Sun, L., Li, Q.: Clip-reid: exploiting vision-language model for image re- identification without concrete text labels. In: Proceedings of the AAAI conference on artificial intelligence. vol. 37, pp. 1405–1413 (2023)

    2023

  10. [10]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 152–159 (2014)

    2014

  11. [11]

    IEEE Transactions on Circuits and Systems for Video Technology (2026)

    Li, X., Lu, Y., Liu, B., Li, J., Liu, Y., Chu, Q., Ye, M., Ouyang, W., Yu, N.: Causal clothes-invariant feature learning for cloth-changing person re-id. IEEE Transactions on Circuits and Systems for Video Technology (2026)

    2026

  12. [12]

    In: Proceed- ings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25

    Li, X., Lu, Y., Liu, B., Li, J., Yang, Q., Gong, T., Chu, Q., Ye, M., Yu, N.: Towards anytime retrieval: A benchmark for anytime person re-identification. In: Proceed- ings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25. pp. 1467–1475. International Joint Conferences on Artificial Intelligence Organization (2025)

    2025

  13. [13]

    In: European conference on computer vision

    Li, X., Lu, Y., Liu, B., Liu, Y., Yin, G., Chu, Q., Huang, J., Zhu, F., Zhao, R., Yu, N.: Counterfactual intervention feature transfer for visible-infrared person re- identification. In: European conference on computer vision. pp. 381–398. Springer (2022)

    2022

  14. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, X., Lu, Y., Liu, B., Yang, Q., Chu, Q., Gong, T., Yu, N.: Mfen: Multi-frequency expert network for visible-infrared person re-id. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18471–18480 (2026) 20 X. Li et al

    2026

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Lu, Y., Wu, Y., Liu, B., Zhang, T., Li, B., Chu, Q., Yu, N.: Cross-modality per- son re-identification with shared-specific feature transfer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13379– 13389 (2020)

    2020

  16. [16]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Luo, C., Chen, Y., Wang, N., Zhang, Z.: Spectral feature transformation for person re-identification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4976–4985 (2019)

    2019

  17. [17]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops

    Luo, H., Gu, Y., Liao, X., Lai, S., Jiang, W.: Bag of tricks and a strong baseline for deep person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0–0 (2019)

    2019

  18. [18]

    Sensors17(3), 605 (2017)

    Nguyen, D.T., Hong, H.G., Kim, K.W., Park, K.R.: Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors17(3), 605 (2017)

    2017

  19. [19]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Rao, Y., Chen, G., Lu, J., Zhou, J.: Counterfactual attention learning for fine-grained visual categorization and re-identification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1025–1034 (2021)

    2021

  20. [20]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Sarfraz, M.S., Schumann, A., Eberle, A., Stiefelhagen, R.: A pose-sensitive embed- ding for person re-identification with expanded cross neighborhood re-ranking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 420–429 (2018)

    2018

  21. [21]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Sun, Y., Cheng, C., Zhang, Y., Zhang, C., Zheng, L., Wang, Z., Wei, Y.: Circle loss: A unified perspective of pair similarity optimization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6398–6407 (2020)

    2020

  22. [22]

    In: Pro- ceedings of the European conference on computer vision (ECCV)

    Sun, Y., Zheng, L., Yang, Y., Tian, Q., Wang, S.: Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In: Pro- ceedings of the European conference on computer vision (ECCV). pp. 480–496 (2018)

    2018

  23. [23]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Wang, H., Shen, J., Liu, Y., Gao, Y., Gavves, E.: Nformer: Robust person re- identification with neighbor transformer. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 7297–7307 (2022)

    2022

  24. [24]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 79–88 (2018)

    2018

  25. [25]

    In: Proceedings of the IEEE international conference on computer vision

    Wu, A., Zheng, W.S., Yu, H.X., Gong, S., Lai, J.: Rgb-infrared cross-modality person re-identification. In: Proceedings of the IEEE international conference on computer vision. pp. 5380–5389 (2017)

    2017

  26. [26]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wu, Q., Dai, P., Chen, J., Lin, C.W., Wu, Y., Huang, F., Zhong, B., Ji, R.: Discover cross-modality nuances for visible-infrared person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4330–4339 (2021)

    2021

  27. [27]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yang, B., Chen, J., Ye, M.: Towards grand unified representation learning for unsu- pervised visible-infrared person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11069–11079 (2023)

    2023

  28. [28]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition

    Yang, J., Li, H., Du, B., Ye, M.: Cheb-gr: Rethinking k-nearest neighbor search in re-ranking for person re-identification. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition. pp. 19261–19270 (2025)

    2025

  29. [29]

    IEEE Transactions on Multimedia18(12), 2553–2566 (2016) ANFI: Rethinking Neighbor Feature Interaction in Person Re-ID 21

    Ye, M., Liang, C., Yu, Y., Wang, Z., Leng, Q., Xiao, C., Chen, J., Hu, R.: Per- son reidentification via ranking aggregation of similarity pulling and dissimilarity pushing. IEEE Transactions on Multimedia18(12), 2553–2566 (2016) ANFI: Rethinking Neighbor Feature Interaction in Person Re-ID 21

    2016

  30. [30]

    In: Proceedings of the IEEE/CVF international con- ference on computer vision

    Ye, M., Ruan, W., Du, B., Shou, M.Z.: Channel augmented joint learning for visible-infrared recognition. In: Proceedings of the IEEE/CVF international con- ference on computer vision. pp. 13567–13576 (2021)

    2021

  31. [31]

    Crandall, D., Shao, L., Luo, J.: Dynamic dual-attentive aggre- gationlearningforvisible-infraredpersonre-identification.In:Europeanconference on computer vision

    Ye, M., Shen, J., J. Crandall, D., Shao, L., Luo, J.: Dynamic dual-attentive aggre- gationlearningforvisible-infraredpersonre-identification.In:Europeanconference on computer vision. pp. 229–247. Springer (2020)

    2020

  32. [32]

    IEEE transactions on pattern analysis and machine intelligence44(6), 2872–2893 (2021)

    Ye, M., Shen, J., Lin, G., Xiang, T., Shao, L., Hoi, S.C.: Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence44(6), 2872–2893 (2021)

    2021

  33. [33]

    arXiv preprint arXiv:2012.07620 (2020)

    Zhang, X., Jiang, M., Zheng, Z., Tan, X., Ding, E., Yang, Y.: Understanding image retrieval re-ranking: A graph neural network perspective. arXiv preprint arXiv:2012.07620 (2020)

    Pith/arXiv arXiv 2012

  34. [34]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhang, Y., Wang, H.: Diverse embedding expansion network and low-light cross- modality benchmark for visible-infrared person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2153– 2162 (2023)

    2023

  35. [35]

    IEEE Transactions on Multimedia 26, 1089–1101 (2023)

    Zhang, Y., Qian, Q., Wang, H., Liu, C., Chen, W., Wang, F.: Graph convolution based efficient re-ranking for visual retrieval. IEEE Transactions on Multimedia 26, 1089–1101 (2023)

    2023

  36. [36]

    In: Proceedings of the IEEE international conference on computer vision

    Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re- identification: A benchmark. In: Proceedings of the IEEE international conference on computer vision. pp. 1116–1124 (2015)

    2015

  37. [37]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 47(3), 1362–1380 (2024)

    Zheng, W.S., Yan, J., Peng, Y.X.: A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence 47(3), 1362–1380 (2024)

    2024

  38. [38]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking person re-identification with k- reciprocal encoding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1318–1327 (2017)

    2017

  39. [39]

    In: Proceedings of the AAAI conference on artificial intelligence

    Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmenta- tion. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001–13008 (2020)

    2020