Pith. sign in

REVIEW 4 major objections 5 minor 100 references

Try Harder: Hard Sample Generation and Learning for Clothes-Changing Person Re-ID

T0 review · 4 major / 5 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 HSGL claims that explicitly defining hard samples by clothing and viewpoint labels, generating them with diffusion and try-on models, and re-weighting triplet distances achieves state-of-the-art clothes-changing person re-identification…

desk verdict Solid CC-ReID data-augmentation paper with real gains on PRCC, but the LTCC SOTA claim is overstated and the HSDA loss description is confused. read the letter →

arxiv 2507.11119 v1 pith:3DL7M5KK submitted 2025-07-15 cs.CV

Hankun Liu , Yujian Zhao , Guanglin Niu This is my paper
classification cs.CV
keywords clothes-changingpersonre-identificationhardsamplegenerationlearningmultimodaldiffusionmodelsvirtualtry-onmetrictripletloss
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

Clothes-changing person re-identification fails when two images of the same person look different and images of different people look alike. This paper claims the missing ingredient is an explicit, label-based definition of those 'hard' cases plus a way to manufacture more of them. It defines hard positives as same-identity pairs with different clothing or viewpoint and hard negatives as different-identity pairs wearing the same clothing, then generates both kinds with diffusion and virtual try-on models. A hardness-aware triplet loss that scales distances for these pairs lets a model trained with only 18–20% additional generated samples reach state-of-the-art accuracy on two benchmarks, with the learning scheme reaching baseline accuracy in about a tenth of the iterations.

What carries the argument

The load-bearing mechanisms are the explicit hard-sample definition, the dual-granularity generation module (DGHSG), and the hard-sample distance adjustment (HSDA). The definition converts vague 'hardness' into computable indicator matrices: for each batch, the Hard Sample Analyzer builds $\mathrm{IS\_HP}$ and $\mathrm{IS\_HN}$ from the label predicates. DGHSG couples a masked diffusion inpainting step that changes only the clothing region for volume, with a try-on/try-off pipeline that re-renders a person in a specified garment for label-controllable fine-grained samples. HSDA then multiplies pairwise feature distances by $(1+\alpha)$ for hard positives and $(1-\alpha)$ for hard negatives inside an aggregated triplet loss, $\mathcal{L}_{\text{hatrip}} = \mathcal{L}_{\text{triplet}}(d) + 0.5 \, \mathcal{L}_{\text{triplet}}(d')$, where $d$ is the original distance and $d'$ the adjusted distance. Together these components deliver targeted gradients for precisely the pairs the model finds confusing.

What would settle it

Train the same framework but, before using each generated image, verify identity preservation with a separate face-recognition or body-shape verifier and discard generated samples that fail. If the reported Rank-1 gains on PRCC vanish when only identity-verified generated samples are kept, the gains came from generator artifacts rather than from clothing-invariant hard samples.

Watch

Extended reading notes

Core claim

The paper establishes that hard samples in clothes-changing person re-identification can be defined explicitly from dataset labels: a pair is a hard positive when $(y_i = y_j) \wedge [(c_i \neq c_j) \vee (v_i \neq v_j)]$, and a hard negative when $(y_i \neq y_j) \wedge (c_i = c_j)$, where $y$, $c$, and $v$ are identity, clothing, and viewpoint labels. On top of this definition it builds a dual-granularity generation module: a latent diffusion model inpaints new clothing inside a clothing mask for coarse-grained, large-scale hard positives, while a multimodal try-on/try-off pipeline swaps garments from a shared clothing library to produce fine-grained hard positives and hard negatives with controllable clothing labels. The learning side, Hard Sample Adaptive Learning (HSAL), multiplies the distances of identified hard positive pairs by $1+\alpha$ and hard negative pairs by $1-\alpha$ inside an aggregated triplet loss, injecting targeted gradient signal for exactly the confusing pairs. On PRCC, combining coarse-to-fine generation with HSAL raises Rank-1 by 8.9% and mAP by 11.5% over the baseline, and the approach reaches new state-of-the-art results on PRCC and LTCC while the learning component needs roughly one-tenth of the training iterations.

Load-bearing premise

The generated hard samples genuinely preserve each person's identity while changing only clothing and viewpoint, so the model learns clothing invariance rather than matching generator artifacts.

Editorial extensions

If this is right

  • Datasets that carry clothing and camera labels can use HSAL directly without mining heuristics, because hard positives and negatives are identified by the labels themselves.
  • The coarse-to-fine generation pipeline provides a way to create the hard-negative pairs that real datasets essentially never contain, since it reuses a shared clothing library across identities.
  • Hard sample learning alone speeds convergence: on PRCC, models reach baseline-level accuracy at the 5th–10th epoch rather than the 50th–60th.
  • The gains come from a small injection of synthetic data (18–20%), so the method does not require doubling or tripling the training set.
  • HSGL is plug-and-play over existing methods: CAL, AIM, and FIRe2/CSSC all improve when the hard-sample loss is added, without replacing their own loss terms.

Reading between the lines

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

  • If the gains are driven mainly by dataset scarcity, the explicit definition and generation machinery should transfer to any ReID variant with label-defined ambiguity, such as cross-resolution or cross-pose matching, where hard positives and negatives can be defined by meta-labels.
  • The identity-preservation premise is testable: a face-verification or body-shape check on the generated images would show whether the Rank-1 gain comes from genuine clothing swap or from generator style artifacts; the paper does not report such a check.
  • Because HSDA operates on pairwise distance matrices, it could be applied to batch-all or batch-hard triplet mining, or to other metric losses such as circle loss, which would extend the idea beyond the aggregated-triplet formulation used here.
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 / 5 minor

Summary. The paper proposes HSGL, a framework for clothes-changing person re-identification (CC-ReID) that combines (i) dual-granularity hard sample generation—coarse-grained diffusion-based clothing replacement (CHPSG) and fine-grained virtual try-on/try-off generation (SHSG)—with (ii) a hard sample adaptive learning loss (HSAL) that uses textual clothing/camera labels to modulate triplet distances. Experiments are reported on PRCC and LTCC using CAL, AIM, FIRe2, and CSSC as baselines. The central claims are that adding roughly 18–20% generated hard samples together with HSAL improves accuracy and accelerates convergence, and that the method achieves state-of-the-art performance on both PRCC and LTCC.

Significance. If the results hold, the explicit definition of hard samples via clothing and camera labels is a useful conceptual step, and the plug-and-play nature of HSAL makes the approach easy to integrate into existing ReID losses. The ablations on three baselines consistently show that the combination of generation and learning gives larger gains than either component alone, which is a credible empirical signal. The release of code is also a positive for reproducibility. However, the headline SOTA claim is directly contradicted by the paper's own LTCC table, and the description of the loss mechanism in the abstract and Section 3.3.2 is inverted relative to what Eq. (8) actually optimizes. These issues are central to the paper's message and must be corrected before the contribution can be fully assessed.

major comments (4)
  1. [Abstract, §1, §4.1, Table 1] The abstract and Section 1 claim state-of-the-art performance on both PRCC and LTCC, but Table 1 shows this is not the case: on LTCC the proposed method obtains Rank-1=45.4 and mAP=19.2, whereas CCFA (CVPR'23) reports Rank-1=45.3 and mAP=22.1. Since the paper itself uses both Rank-1 and mAP as evaluation metrics (Section 4), the method is not state-of-the-art on LTCC; the 0.1 Rank-1 advantage is within likely run-to-run noise, and no error bars or multiple-run statistics are provided. Please revise the claim (e.g., SOTA on PRCC only) or add significance testing and a more recent LTCC baseline comparison.
  2. [§3.3.2, Eq. (8), Abstract] The mechanism described in the abstract and Section 3.3.2 is the opposite of what Eq. (8) implements. The abstract says HSAL 'encourages the separation of hard positives and drawing hard negatives closer,' and Section 3.3.2 says it 'increases the distance between hard positive pairs and decreases it for hard negative pairs during training.' However, in Eq. (8), HP_M=1+α multiplies the positive distance term in a triplet loss, which amplifies the gradient that decreases d(a,p); HN_M=1−α multiplies the negative distance term, which attenuates the gradient that increases d(a,n). Thus the implemented loss pulls hard positives together and weakens the separation of hard negatives. Please reconcile the prose with the equations, or change the loss if the intended behavior is the one described in the text.
  3. [§4.3, Table 2] The AIM baseline in Table 2 is reported as Rank-1=54.8, mAP=55.7, but Table 1 lists AIM as Rank-1=57.9, mAP=58.3. If the ablation uses a re-implementation, this must be stated explicitly and the settings (e.g., backbone, resolution, loss weights) given; otherwise the claimed gains for AIM (+4.7/+3.9) are not anchored to the same reference point as the literature comparison. The same issue affects the interpretation of the CAL baseline (54.8/54.7 vs 55.2/55.8 in Table 1).
  4. [§3.2, §4.5, Fig. 5] The framework's validity rests on the assumption that generated hard samples preserve identity (face, body shape, posture) while changing only clothing. The only evidence provided is a set of visual examples in Fig. 5. Please add a quantitative identity-preservation check, for example retrieval accuracy on the generated images against the original identity gallery, or a manual/automatic failure-rate measurement of the generation pipeline. Without this, the Rank-1 gains could partly reflect matching of generation artifacts rather than true clothing invariance.
minor comments (5)
  1. [§3.3.2] The sentence 'f(ax) and a f(x) yield the same gradient upon differentiation' is not true for general nonlinear functions such as the triplet loss in Eq. (8), which contains a max and a margin. This justification should be removed or replaced with a correct statement about loss weighting.
  2. [Eq. (3)] The formulas for N_HP and N_HN appear to double-count pairs. For an identity with K original images and mn generated images, the number of within-identity cross-garment/cross-view pairs is (K+mn)(K+mn−1)/2, not mn(K+mn−1). Since these counts are not used in training, this is a presentation issue, but the formulas should be corrected or removed.
  3. [Table 2] The first column labeled 'Baseline' is used differently for FIRe2 (where it indicates CHPSG-based pretraining) than for CAL/AIM; please rename it (e.g., 'CHPSG pretrain') and explain in the caption.
  4. [§4.2 and §4.4] The claims about convergence speed (1/10 of training iterations) and hyperparameter sensitivity are based on Figures 3 and 4, but the text does not provide the numerical epoch/accuracy values or the alpha/lambda values used. Please report these numbers so the claims can be verified.
  5. [References] References [43] and [44] refer to the same paper (Cloth-aware Augmentation) in two venues, and [70] and [71] are also duplicated; please consolidate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical data-generation plus metric-learning pipeline validated on external benchmarks.

full rationale

The paper's derivation chain is an empirical construction, not a formal derivation. Eq. (1) defines hard positive/negative pairs from dataset labels (identity, clothing, viewpoint). The HSDA matrices in Eqs. (6)-(7) and the loss in Eq. (8) are deliberately built from these same indicators (Eq. 5), so the training signal is by construction aligned with the paper's own hard-sample definition. This is a design choice for a supervised objective, not a circular prediction: the Rank-1/mAP results in Table 1 are measured on held-out test data from PRCC and LTCC, which are independent of the definition and of the generated training samples. The ablation and convergence results are similarly empirical. The self-citations present (IFD [68], Pose2ID [77], CCUP [84]) appear only as comparison baselines, related work, or data-generation references; none is load-bearing for the central claims. Hyper-parameter tuning of alpha and lambda on PRCC is a fitting/soundness concern, not circularity. The identity-preservation assumption of the generators is supported only by visual examples, which is an evidence-quality concern, not a circularity concern. Separately, the abstract's claim of 'state-of-the-art on both PRCC and LTCC' is internally inconsistent with Table 1, where LTCC mAP (19.2) is below CCFA (22.1); this is a correctness/consistency problem, not a circularity problem. No step in the paper reduces to its own inputs by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework's gains rest on two unverified pillars: synthetic images preserve identity, and label-defined hard samples can be generated and weighted usefully. Several hyperparameters are hand-set or tuned without reported values, and the identity-preservation claim is supported only by visual examples. No new physical entities are introduced.

free parameters (5)
  • Distance adjustment factor alpha = not reported (tuned on PRCC with CAL; Fig. 4)
    Controls how much hard positive/negative distances are rescaled in Eqs. 6-8; sensitivity analysis shows strong dependence.
  • Aggregated triplet loss weight lambda = 0.1 for CAL/AIM; original baseline coefficient for FIRe2/CSSC
    Weight in Eq. 12; set per baseline, not justified by a principled criterion.
  • Half-weight for adjusted triplet loss = 0.5
    Fixed heuristic in Eq. 10 balancing original and adjusted triplet losses.
  • Generation counts m and n = not reported
    Determines number of synthetic hard positives/negatives in Eq. 3; only total 18-20% augmentation is mentioned.
  • Frontal pose and quality thresholds = 0.7 visibility, epsilon_y and epsilon_v not quantified
    Hand-set filters in Eq. 2 that decide which images enter the generation pipeline.
assumptions (4)
  • domain assumption Pre-trained generative models (IDM-VTON, TryOffAnyone, Stable Diffusion, SCHP) preserve identity and clothing semantics when applied to pedestrian images.
    Invoked throughout Section 3.2; if false, synthetic hard samples would teach the model wrong cues. Only qualitative visual evidence is provided.
  • domain assumption Clothing labels in PRCC (indoor/outdoor as clothing) and LTCC (clothing IDs) are accurate proxies for visual clothing state.
    Sections 3.1 and 3.3.1 rely on these labels to define hard positives/negatives; mislabeled clothing would create incorrect supervision.
  • standard math Scaling distances in the triplet loss by constants yields gradients equivalent to a harder triplet loss.
    Section 3.3.2 appeals to f(ax) versus a f(x) gradient equivalence; this is standard but only approximate for hinge loss.
  • ad hoc to paper Increasing the learning pressure on hard sample pairs transfers to better retrieval on real test data.
    This is the core design belief of HSAL; not derived from first principles and depends on the generated data distribution matching test hard samples.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Try Harder: Hard Sample Generation and Learning for Clothes-Changing Person Re-ID." pith.science (2026). https://pith.science/paper/3DL7M5KK

@misc{pith2026250711119,
  author       = {Pith},
  title        = {Pith review of: Try Harder: Hard Sample Generation and Learning for Clothes-Changing Person Re-ID},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3DL7M5KK}},
  note         = {Machine review of arXiv:2507.11119}
}
read the original abstract

Hard samples pose a significant challenge in person re-identification (ReID) tasks, particularly in clothing-changing person Re-ID (CC-ReID). Their inherent ambiguity or similarity, coupled with the lack of explicit definitions, makes them a fundamental bottleneck. These issues not only limit the design of targeted learning strategies but also diminish the model's robustness under clothing or viewpoint changes. In this paper, we propose a novel multimodal-guided Hard Sample Generation and Learning (HSGL) framework, which is the first effort to unify textual and visual modalities to explicitly define, generate, and optimize hard samples within a unified paradigm. HSGL comprises two core components: (1) Dual-Granularity Hard Sample Generation (DGHSG), which leverages multimodal cues to synthesize semantically consistent samples, including both coarse- and fine-grained hard positives and negatives for effectively increasing the hardness and diversity of the training data. (2) Hard Sample Adaptive Learning (HSAL), which introduces a hardness-aware optimization strategy that adjusts feature distances based on textual semantic labels, encouraging the separation of hard positives and drawing hard negatives closer in the embedding space to enhance the model's discriminative capability and robustness to hard samples. Extensive experiments on multiple CC-ReID benchmarks demonstrate the effectiveness of our approach and highlight the potential of multimodal-guided hard sample generation and learning for robust CC-ReID. Notably, HSAL significantly accelerates the convergence of the targeted learning procedure and achieves state-of-the-art performance on both PRCC and LTCC datasets. The code is available at https://github.com/undooo/TryHarder-ACMMM25.

Figures

Figures reproduced from arXiv: 2507.11119 by the authors.

Figure 1
Figure 1. Conceptual comparison of hard samples and solu [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed Hard Sample Generation And Learning(HSGL) framework. The architecture comprises [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Accuracy improvement curves of CAL and AIM [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Rank-1 Accuracy Heatmap Based on the Dis [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Samples generated by the two strategies of Dual [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

100 extracted references · 60 canonical work pages

  1. [1]

    Vaibhav Bansal, Gian Luca Foresti, and Niki Martinel. 2022. Cloth-changing person re-identification with self-attention. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 602–610

    2022

  2. [2]

    Mikhail Belkin and Partha Niyogi. 2003. Laplacian eigenmaps for dimensionality reduction and data representation. Neural computation 15, 6 (2003), 1373–1396

    2003

  3. [3]

    Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al . 2023. Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing 31 (2023), 2523–2533

    2023

  4. [4]

    Patrick PK Chan, Xiaoman Hu, Haorui Song, Peng Peng, and Keke Chen. 2023. Learning disentangled features for person re-identification under clothes chang- ing. ACM Transactions on Multimedia Computing, Communications and Applica- tions 19, 6 (2023), 1–21

    2023

  5. [5]

    Jiaxing Chen, Xinyang Jiang, Fudong Wang, Jun Zhang, Feng Zheng, Xing Sun, and Wei-Shi Zheng. 2021. Learning 3d shape feature for texture-insensitive person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8146–8155

    2021

  6. [6]

    Jiaxing Chen, Wei-Shi Zheng, Qize Yang, Jingke Meng, Richang Hong, and Qi Tian. 2021. Deep shape-aware person re-identification for overcoming moderate clothing changes. IEEE Transactions on Multimedia 24 (2021), 4285–4300

    2021

  7. [7]

    Kezhou Chen, Yang Chen, Chuchu Han, Nong Sang, and Changxin Gao. 2020. Hard sample mining makes person re-identification more efficient and accurate. Neurocomputing 382 (2020), 259–267. doi:10.1016/j.neucom.2019.11.094

    work page doi:10.1016/j.neucom.2019.11.094 2020

  8. [8]

    Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. 2021. Viton- hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14131–14140

    2021

Show all 100 references
  1. [9]

    Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin

  2. [10]

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah

  3. [11]

    Zhenyu Cui, Jiahuan Zhou, Yuxin Peng, Shiliang Zhang, and Yaowei Wang

  4. [12]

    Weijian Deng, Liang Zheng, Qixiang Ye, Guoliang Kang, Yi Yang, and Jianbin Jiao

  5. [13]

    Weijian Deng, Liang Zheng, Qixiang Ye, Yi Yang, and Jianbin Jiao. 2018. Similarity- preserving image-image domain adaptation for person re-identification. arXiv preprint arXiv:1811.10551 (2018)

    2018 arXiv

  6. [14]

    IEEE transactions on circuits and systems for video technology 33, 8 (2023), 4415–4428

    Dcr-reid: Deep component reconstruction for cloth-changing person re- identification. IEEE transactions on circuits and systems for video technology 33, 8 (2023), 4415–4428

    2023

  7. [15]

    Zan Gao, Shengxun Wei, Weili Guan, Lei Zhu, Meng Wang, and Shengyong Chen. 2023. Identity-guided collaborative learning for cloth-changing person reidentification. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 5 (2023), 2819–2837

    2023

  8. [16]

    Shuze Geng, Yifan Liu, Zijin Wang, Gang Yan, Yu Yang, and Yingchun Guo

  9. [17]

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. Advances in neural information processing systems 27 (2014)

    2014

  10. [18]

    Zan Gao, Hongwei Wei, Weili Guan, Weizhi Nie, Meng Liu, and Meng Wang

  11. [19]

    Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. 2022. Clothes-changing person re-identification with rgb modality only. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 1060–1069

    2022

  12. [21]

    Peini Guo, Hong Liu, Jianbing Wu, Guoquan Wang, and Tao Wang. 2023. Semantic-aware consistency network for cloth-changing person re-identification. In Proceedings of the 31st ACM international conference on multimedia. 8730–8739

    2023

  13. [22]

    Ke Han, Shaogang Gong, Yan Huang, Liang Wang, and Tieniu Tan. 2023. Clothing- change feature augmentation for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 22066–22075

    2023

  14. [23]

    Ke Han, Yan Huang, Shaogang Gong, Liang Wang, and Tieniu Tan. 2022. 3d shape temporal aggregation for video-based clothing-change person re-identification. In Proceedings of the Asian Conference on Computer Vision . 2371–2387

    2022

  15. [24]

    Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang

  16. [25]

    In Proceedings of the 31st ACM International Conference on Multimedia

    Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia. 7599–7607

  17. [26]

    Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, et al. 2024. Instruct-reid: A multi-purpose person re-identification task with instructions. InProceedings of the IEEE/CVF Conference on Computer Vision an...

    2024

  18. [27]

    Alexander Hermans, Lucas Beyer, and Bastian Leibe. 2017. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)

    2017 arXiv

  19. [28]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

    2020

  20. [29]

    Peixian Hong, Tao Wu, Ancong Wu, Xintong Han, and Wei-Shi Zheng. 2021. Fine-grained shape-appearance mutual learning for cloth-changing person re- identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10513–10522

    2021

  21. [30]

    Yan Huang, Qiang Wu, JingSong Xu, Yi Zhong, and ZhaoXiang Zhang. 2021. Clothing status awareness for long-term person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 11895–11904

    2021

  22. [31]

    Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. 2018. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition . 7543–7552

    2018

  23. [32]

    Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. 2021. Tran- sreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision . 15013–15022

    2021

  24. [33]

    Asaf Joseph and Shmuel Peleg. 2025. Clothes-Changing Person Re-identification Based On Skeleton Dynamics. arXiv preprint arXiv:2503.10759 (2025)

    2025 arXiv

  25. [34]

    Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. 2024. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8176–8185

    2024

  26. [35]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classifi- cation with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)

    2012

  27. [36]

    Qingming Leng, Mang Ye, and Qi Tian. 2019. A survey of open-world person re-identification. IEEE Transactions on Circuits and Systems for Video Technology 30, 4 (2019), 1092–1108

    2019

  28. [37]

    Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. 2020. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 6 (2020), 3260–3271

    2020

  29. [38]

    Xuemei Jia, Xian Zhong, Mang Ye, Wenxuan Liu, and Wenxin Huang. 2022. Complementary data augmentation for cloth-changing person re-identification. IEEE Transactions on Image Processing 31 (2022), 4227–4239

    2022

  30. [39]

    Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. 2022. Cloth-changing person re-identification from a single image with gait prediction and regulariza- tion. In Proceedings of the IEEE/CVF confere...

    2022

  31. [40]

    Feng Liu, Minchul Kim, ZiAng Gu, Anil Jain, and Xiaoming Liu. 2023. Learning clothing and pose invariant 3d shape representation for long-term person re- identification. In Proceedings of the IEEE/CVF international conference on computer vision. 19617–19626

    2023

  32. [41]

    Feng Liu, Minchul Kim, Zhiyuan Ren, and Xiaoming Liu. 2024. Distilling CLIP with dual guidance for learning discriminative human body shape representa- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 256–266

    2024

  33. [42]

    Fangyi Liu, Mang Ye, and Bo Du. 2023. Dual level adaptive weighting for cloth- changing person re-identification. IEEE Transactions on Image Processing 32 (2023), 5075–5086

    2023

  34. [44]

    Fangyi Liu, Mang Ye, and Bo Du. 2024. Cloth-aware Augmentation for Cloth- generalized Person Re-identification. InProceedings of the 32nd ACM International Conference on Multimedia. 4053–4062

    2024

  35. [45]

    Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. 2023. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

    2023

  36. [46]

    Xin Liang and Yogesh S Rawat. 2025. DIFFER: Disentangling Identity Fea- tures via Semantic Cues for Clothes-Changing Person Re-ID. arXiv preprint arXiv:2503.22912 (2025)

    2025 arXiv

  37. [47]

    Huy Nguyen, Kien Nguyen, Akila Pemasiri, Feng Liu, Sridha Sridharan, and Clin- ton Fookes. 2025. AG-VPReID: A Challenging Large-Scale Benchmark for Aerial- Ground Video-based Person Re-Identification. arXiv preprint arXiv:2503.08121 (2025)

    2025 arXiv

  38. [48]

    Vuong D Nguyen, Pranav Mantini, and Shishir K Shah. 2024. Contrastive clothing and pose generation for cloth-changing person re-identification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 7541–7549

    2024

  39. [49]

    Vuong D Nguyen, Pranav Mantini, and Shishir K Shah. 2024. Temporal 3d shape modeling for video-based cloth-changing person re-identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 173–182

    2024

  40. [50]

    C Peng, B Wang, D Liu, N Wang, R Hu, and X Gao. [n. d.]. Masked attribute description embedding for cloth-changing person re-identification (2024). arXiv preprint arXiv:2401.05646 ([n. d.])

    2024 arXiv

  41. [51]

    Chunlei Peng, Boyu Wang, Decheng Liu, Nannan Wang, Ruimin Hu, and Xinbo Gao. 2024. Masked Attribute Description Embedding for Cloth-Changing Person Re-identification. IEEE Transactions on Multimedia (2024)

    2024

  42. [52]

    Mengmeng Liu, Zhi Ma, Tao Li, Yanfeng Jiang, and Kai Wang. 2022. Long- term person re-identification with dramatic appearance change: Algorithm and benchmark. In Proceedings of the 30th ACM International Conference on Multimedia. 6406–6415

    2022

  43. [53]

    Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Conference ACM MM ’2025, October 27–31, 2025, Dublin, Ireland Hankun Liu et al. et al. 2019. Mediapipe: A framework for building ...

    2019 arXiv

  44. [54]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  45. [55]

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. 2022. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings . 1–10

    2022

  46. [56]

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural infor...

    2022

  47. [57]

    Hao Sheng, Yanwei Zheng, Wei Ke, Dongxiao Yu, Xiuzhen Cheng, Weifeng Lyu, and Zhang Xiong. 2020. Mining Hard Samples Globally and Efficiently for Person Reidentification. IEEE Internet of Things Journal 7, 10 (2020), 9611–9622. doi:10.1109/JIOT.2020.2980549

    2020

  48. [58]

    Nyle Siddiqui, Florinel Alin Croitoru, Gaurav Kumar Nayak, Radu Tudor Ionescu, and Mubarak Shah. 2024. DLCR: A Generative Data Expansion Framework via Diffusion for Clothes-Changing Person Re-ID. arXiv preprint arXiv:2411.07205 (2024)

    2024 arXiv

  49. [59]

    Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yanwei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. 2020. Long-term cloth-changing person re-identification. In Proceedings of the Asian conference on computer vision

    2020

  50. [60]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen

  51. [61]

    arXiv preprint arXiv:2204.06125 1, 2 (2022), 3

    Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1, 2 (2022), 3

    2022 arXiv

  52. [62]

    Hongbin Tu, Chao Liu, Yuanyuan Peng, Haibo Xiong, and Haotian Wang. 2024. Clothing-change person re-identification based on fusion of RGB modality and gait features. Signal, Image and Video Processing 18, 3 (2024), 2367–2376

    2024

  53. [63]

    Qizao Wang, Xuelin Qian, Bin Li, Lifeng Chen, Yanwei Fu, and Xiangyang Xue

  54. [64]

    Qizao Wang, Xuelin Qian, Bin Li, Xiangyang Xue, and Yanwei Fu. 2024. Exploring fine-grained representation and recomposition for cloth-changing person re- identification. IEEE Transactions on Information Forensics and Security (2024)

    2024

  55. [65]

    Yubin Wang, Huimin Yu, Yuming Yan, Shuyi Song, Biyang Liu, and Yichong Lu

  56. [66]

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen

  57. [67]

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli

  58. [68]

    Haoxuan Xu, Bo Li, and Guanglin Niu. 2025. Identity-aware Feature Decoupling Learning for Clothing-change Person Re-identification. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . 1–5. doi:10.1109/ICASSP49660.2025.10888292

    2025

  59. [69]

    Yang Song and Stefano Ermon. 2019. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)

    2019

  60. [70]

    Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid

  61. [71]

    Qize Yang, Ancong Wu, and Wei-Shi Zheng. 2019. Person re-identification by contour sketch under moderate clothing change. IEEE transactions on pattern analysis and machine intelligence 43, 6 (2019), 2029–2046

    2019

  62. [72]

    Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. 2024. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 7017–7026

    2024

  63. [73]

    Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. 2023. Good is bad: Causality inspired cloth-debiasing for cloth-changing person re- identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1472–1481

    2023

  64. [74]

    In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Content and salient semantics collaboration for cloth-changing person re-identification. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 1–5

    2025

  65. [75]

    Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi

  66. [76]

    Hong-Xing Yu, Wei-Shi Zheng, Ancong Wu, Xiaowei Guo, Shaogang Gong, and Jian-Huang Lai. 2019. Unsupervised person re-identification by soft multilabel learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2148–2157

    2019

  67. [77]

    In Proceedings of the 31st ACM International Conference on Multimedia

    Exploring shape embedding for cloth-changing person re-identification via 2d-3d correspondences. In Proceedings of the 31st ACM International Conference on Multimedia. 7121–7130

  68. [78]

    Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An- An Liu. 2024. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8372–8382

    2024

  69. [79]

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

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 1206–1217

  70. [80]

    Ioannis Xarchakos and Theodoros Koukopoulos. 2024. TryOffAnyone: Tiled Cloth Generation from a Dressed Person. arXiv preprint arXiv:2412.08573 (2024)

    2024 arXiv

  71. [81]

    Lei Zhang, Xiaowei Fu, Fuxiang Huang, Yi Yang, and Xinbo Gao. 2024. An open-world, diverse, cross-spatial-temporal benchmark for dynamic wild person re-identification. International Journal of Computer Vision 132, 9 (2024), 3823– 3846

    2024

  72. [82]

    Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. 2023. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 18381–18391

    2023

  73. [84]

    Yujian Zhao, Chengru Wu, Yinong Xu, Xuanzheng Du, Ruiyu Li, and Guanglin Niu. 2025. CCUP: A Controllable Synthetic Data Generation Pipeline for Pretrain- ing Cloth-Changing Person Re-Identification Models. arXiv:2410.13567 [cs.CV] https://arxiv.org/abs/2410.13567

    2025 arXiv

  74. [85]

    Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz

  75. [86]

    Zhun Zhong, Liang Zheng, Zhedong Zheng, Shaozi Li, and Yi Yang. 2018. Camera style adaptation for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition . 5157–5166

    2018

  76. [87]

    Zhengwei Yang, Xian Zhong, Zhun Zhong, Hong Liu, Zheng Wang, and Shin’Ichi Satoh. 2023. Win-win by competition: Auxiliary-free cloth-changing person re-identification. IEEE Transactions on Image Processing 32 (2023), 2985–2999

    2023

  77. [91]

    Chao Yuan, Guiwei Zhang, Changxiao Ma, Tianyi Zhang, and Guanglin Niu

  78. [92]

    In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)

    From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR). 24409–24418

  79. [94]

    Guoqing Zhang, Jin Li, Yuhui Zheng, and Ruili Wang. 2024. InfinitePerson: Innovating Synthetic Data Creation for Generalization Person Re-Identification. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

  80. [95]

    Haiyang Zhang and Xinshuang Wang. 2025. Cloth-debiasing with Stable Dif- fusion in Cloth-changing Person Re-identification. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 1–5

    2025

  81. [97]

    Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. 2021. Unrealperson: An adaptive pipeline towards costless person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11506–11515

    2021

  82. [98]

    Huazhong Zhao, Lei Qi, and Xin Geng. 2024. CLIP-DFGS: A Hard Sample Mining Method for CLIP in Generalizable Person Re-Identification. ACM Transactions on Multimedia Computing, Communications and Applications 21, 1 (2024), 1–20

    2024

  83. [101]

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

    Joint discriminative and generative learning for person re-identification. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2138–2147

  84. [103]

    Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Saharia, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. 2023. Tryondif- fusion: A tale of two unets. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4606–4615

    2023

  85. [2015]

    In International conference on machine learning

    Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning . pmlr, 2256–2265

  86. [2018]

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

    Image-image domain adaptation with preserved self-similarity and domain- dissimilarity for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition . 994–1003

  87. [2019]

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

    Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision . 7464– 7473

  88. [2021]

    IEEE transactions on pattern analysis and machine intelligence 44, 6 (2021), 2872–2893

    Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence 44, 6 (2021), 2872–2893

    2021

  89. [2022]

    In Proceedings of the 30th ACM international conference on multi- media

    Multigranular visual-semantic embedding for cloth-changing person re- identification. In Proceedings of the 30th ACM international conference on multi- media. 3703–3711

  90. [2023]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 9 (2023), 10850–10869

    Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 9 (2023), 10850–10869

    2023

  91. [2024]

    In European Conference on Computer Vision

    Improving diffusion models for authentic virtual try-on in the wild. In European Conference on Computer Vision . Springer, 206–235

  92. [2025]

    IEEE Transactions on Circuits and Systems for Video Technology (2025)

    Pose-Skeleton Guided Cross-Attention Representation Fusion for Occluded Pedestrian Re-Identification. IEEE Transactions on Circuits and Systems for Video Technology (2025)

    2025

Pith tools

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