Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

SCING:Towards More Efficient and Robust Person Re-Identification through Selective Cross-modal Prompt Tuning

T0 review · 3 major / 7 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 Selective cross-modal prompt tuning matches or beats heavy adapter designs in person re-identification while keeping inference cost essentially unchanged.

desk verdict A plausible prompt-tuning method for CLIP-based ReID with solid benchmark coverage, but internal number inconsistencies and an unverified transfer path for the consistency loss make the evidence weaker than the tone suggests. read the letter →

arxiv 2507.00506 v1 pith:G6MZWBPM submitted 2025-07-01 cs.CV

Yunfei Xie , Yuxuan Cheng , Juncheng Wu , Haoyu Zhang , Yuyin Zhou , Shoudong Han This is my paper
classification cs.CV
keywords PersonRe-IdentificationPromptTuningCross-ModalContrastiveLearningSelectiveVisualFusionPerturbation-DrivenConsistencyAlignmentOccludedCLIP
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

SCING sets out to show that a CLIP model can be adapted to person re-identification (ReID) without heavy adapter modules or extra segmentation labels. The paper argues that two lightweight mechanisms, selectively gating visual features into learnable text prompts and a perturbation-driven consistency loss on those fused prompts, close the image-text modality gap that plain prompt tuning leaves open. Across six ReID benchmarks, including four occluded ones, the method reports top or competitive Rank-1 and mAP numbers while keeping inference parameters and FLOPs essentially identical to the CLIP-ReID baseline. If this holds, the efficient path to robust ReID is prompt-side cross-modal interaction, not architecture growth.

What carries the argument

Selective Visual Prompt Fusion (SVIP) is a weighted gating module $\boldsymbol{\alpha} = \sigma(\mathbf{W}_s \mathbf{V} + \mathbf{b}_s)$ that selects which visual-condition dimensions matter and then adds the gated visual condition to the first $M$ learnable text tokens via $\mathbf{p}_i^{svip} = \mathbf{p}_i + \mathbf{c}_i \odot \boldsymbol{\alpha}_i$. Perturbation-Driven Consistency Alignment (PDCA) generates two perturbed views $I'$ and $I''$ of the same image and applies the loss $L_{con} = 1 - \tfrac{1}{3}\big(\cos(\mathbf{w},\mathbf{w}') + \cos(\mathbf{w},\mathbf{w}'') + \cos(\mathbf{w}',\mathbf{w}'')\big)$, pushing the fused text embeddings from original and perturbed images together. Stage 1 trains these mechanisms with the CLIP contrastive loss plus $L_{con}$, and Stage 2 freezes the text side while full-tuning the visual encoder with identity classification and triplet losses.

What would settle it

Train SCING with the PDCA loss exactly as specified and then, before Stage 2 visual tuning, freeze the visual encoder and extract the descriptor; if the text-side consistency loss produces no measurable change in those Stage-1 visual features, the claimed transfer from cross-modal consistency to visual-only inference is refuted. A simpler check is to ablate only $L_{con}$ while keeping SVIP and all other settings, in which case the reported +4.3 mAP gain on Occluded-Duke should disappear if PDCA is the cause.

Watch

Extended reading notes

Core claim

The central discovery is that indiscriminate image-conditioned prompts, of the CoCoOp style, hurt ReID because they let background noise into the text tokens, while a gating mechanism that selects only discriminative visual dimensions before fusing them into the first $M$ text tokens recovers the alignment. Adding a consistency term that maximizes cosine similarity among text embeddings produced from the original image and two perturbed views makes the descriptor stable under occlusion and viewpoint change. The trained text encoder and prompts are discarded at inference; only the CLIP visual encoder, fine-tuned in a second stage, serves as the global descriptor. On the reported numbers this yields 96.2% Rank-1 on Market1501, 71.1% Rank-1 on Occluded-Duke, and an average occluded-ReID Rank-1 of 86.20% versus 67.20% for the CLIP-ReID baseline, all at 126.95M parameters and 24.14 G FLOPs.

Load-bearing premise

The consistency loss is applied to text embeddings computed from original and perturbed images, but at inference the text encoder and prompts are discarded, so the method assumes this text-side alignment changes the visual encoder in a way that improves the visual-only descriptor.

Editorial extensions

If this is right

  • CLIP-based ReID systems can match or beat adapter-heavy methods using about 127M parameters and 24.14 G FLOPs at inference, removing a practical deployment barrier.
  • Occluded ReID improves without segmentation labels or part-specific prompts, because the gating mechanism suppresses identity-irrelevant visual content before it reaches the text tokens.
  • CoCoOp-style global visual-to-text fusion is not sufficient for ReID; a selective version is what prevents background noise from contaminating the text prompts.
  • Since only the visual encoder is used at inference, the learned prompts and text encoder can be discarded after training, so the train-time cross-modal machinery adds no runtime cost.

Reading between the lines

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

  • The paper leaves implicit that the same recipe could transfer to other CLIP-based retrieval tasks where one modality is unavailable at test time, such as vehicle re-identification or face verification, since the cross-modal machinery is train-only and discarded before deployment.
  • A testable extension is to measure how much of the Rank-1 gain is explained by the measured reduction of the image-text modality gap, using the UMAP-style alignment diagnostic as a predictor rather than just a visualization.
  • The paper does not isolate whether PDCA improves the visual descriptor directly or only sharpens the Stage-1 text prompts that anchor Stage-2 training; adding the same perturbation consistency directly to the Stage-2 visual-only loss would clarify the causal path.
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

3 major / 7 minor

Summary. The paper proposes SCING, a two-stage training framework that adapts CLIP to person re-identification without inference-time adapters. In Stage 1, learnable text prompts are fused with a gated visual condition (SVIP, Eqs. 9-11) and optimized with a CLIP loss plus a perturbation-driven consistency loss (PDCA, Eq. 14) that encourages text embeddings from original and perturbed views to agree. In Stage 2, the text side is frozen and the visual encoder is fine-tuned against the learned text prototypes (Eqs. 17-20); at inference only the visual descriptor is used. Experiments on six benchmarks report competitive or state-of-the-art results and inference cost essentially equal to CLIP-ReID.

Significance. If the central claim is correct, SCING would be an attractive result: a lightweight cross-modal interaction mechanism during training that improves the visual-only descriptor at inference, with no extra inference parameters or FLOPs. The paper compares against a broad set of baselines, and the gating idea behind SVIP is simple and plausible. The efficiency numbers, once corrected, would also be a useful point of comparison against adapter-based methods. However, the manuscript does not yet establish the load-bearing link between the training-time consistency loss and the final visual descriptor, and the reported baseline numbers are internally inconsistent. No code or checkpoints are provided, so precision in the training protocol is essential.

major comments (3)
  1. [§3.5.1, §3.5.2, §4.2, Eq. (14)] The paper contradicts itself on what Stage 1 optimizes, and this determines whether PDCA can affect the inference descriptor. Section 3.5.1 says Stage 1 'jointly optimize[s] all parameters in both visual and textual streams,' but Section 4.2 says 'only the learnable text tokens [X]... are optimized' in Stage 1 and 'only the visual encoder' in Stage 2. Because Eq. (14) is defined entirely on text embeddings w, w', w'', under the Section 4.2 protocol dL_con/dF_v = 0 and the robustness benefit must flow indirectly through the text prototypes used in Stage 2. The paper neither reports the actual trainable parameter sets per stage nor isolates this indirect route. I ask the authors to (i) reconcile the two descriptions, (ii) report lambda and gamma, and (iii) provide an ablation or analysis showing that PDCA changes the final visual descriptor beyond what Stage-2 full fine-tuning already provides, for example by comparing with and without PDCA while holding Stage-2 settings and initialization fixed and running multiple seeds.
  2. [Tables 2 and 3] The CLIP-ReID numbers on Occluded-Duke are inconsistent: Table 2 lists 67.2 Rank-1 / 60.3 mAP, while Table 3 lists 59.1 mAP / 65.7 Rank-1 for the same baseline. These numbers are close but not identical, so the discrepancy cannot be dismissed as a simple column swap; it suggests a different protocol or a typo. Since the ablation story in Section 4.6 is built on increments relative to this baseline, the authors must report the exact training/evaluation protocol used in Table 3 and state which numbers are correct.
  3. [Table 4] The efficiency table's baseline average is internally inconsistent. Table 4 states that Avg. Rank-1 and Avg. mAP are computed over Occluded-Duke, Occluded-ReID, and P-DukeMTMC, but CLIP-ReID has no Occluded-ReID entry in Table 2, and its listed average 67.20/60.30 is exactly its Occluded-Duke entry. Averaging the available entries (Occluded-Duke 67.2/60.3 and P-DukeMTMC 91.3/83.7) gives 79.25/72.0. Please correct the averaging procedure or state which datasets are actually used; this directly affects the 'optimal trade-off' claim.
minor comments (7)
  1. [§3.5.2] 'freezing all textual components (text encoder F_v, ...)' should read 'text encoder F_t'; F_v is the visual encoder elsewhere in the paper.
  2. [§3.3.1, §4.2] The values of L and M in Eq. (6) are never reported, and Eq. (9) outputs d gate values while the text says alpha contains M weights; please clarify how the d-dimensional gate is sliced onto M prompt tokens.
  3. [References] Reference [26] is cited for Market1501, but the listed reference is 'A Survey of Pruning Methods for Efficient Person Re-identification Across Domains'; the Market1501 citation appears to be missing.
  4. [§4.2] The perturbation schedule (random flipping, erasing, cropping, occlusion, feature dropout at 50%) is described only qualitatively; specify the composition, intensity, and any scheduling of these perturbations, since PDCA results may depend on them.
  5. [Abstract and §1] The abstract lists five benchmarks while the body evaluates six (Occluded-Market is omitted from the abstract); align the count.
  6. [Table 4] Table 4 reports inference parameters and FLOPs, but the paper does not state this scope explicitly; add a sentence clarifying that training cost (including Stage-2 full visual fine-tuning) is not part of the efficiency comparison.
  7. [§3.4.3 and Abstract] The abstract and Section 1 describe PDCA as regularizing 'original and augmented cross-modal embeddings,' but Eq. (14) compares only text embeddings w, w', w''; the image representation is not part of the consistency loss. Please reword the description to match the actual objective.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SCING is an empirical prompt-tuning framework validated on external benchmarks, with no load-bearing step reducing to its own inputs.

full rationale

SCING is an empirical method: SVIP (Eqs. 9-11) and PDCA (Eq. 14) are proposed training-time components, and the paper's claims are benchmark numbers (Tables 1-2) and an ablation study (Table 3). No equation is defined in terms of the quantity it is said to predict, and no fitted parameter is renamed as a prediction. The inference descriptor is the CLIP visual backbone output only (Sec. 3.5.3), while SVIP/PDCA act during training; whether PDCA's text-side consistency transfers to the visual descriptor is an empirical, testable question, and the Table 3 ablation is the paper's evidence for it. An unverified transfer would be a correctness or reproducibility concern, not circularity. The paper includes one co-authored prior work ([51]) but only as a dataset/related-work reference, and it is not load-bearing for the central derivation. Internal inconsistencies (CLIP-ReID reported as 67.2/60.3 in Table 2 versus 59.1/65.7 in Table 3, and the Table 4 averages) are correctness risks rather than circular reasoning. The derivation chain is therefore self-contained with respect to circularity, so the score is 0.

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

The central claim rests on standard CLIP assumptions plus several unstated hyperparameters. No new physical or architectural entities are introduced beyond learned parameters. The most consequential choices are the unreported loss weights and the assumption that text-side consistency improves visual-only retrieval.

free parameters (5)
  • lambda (stage-1 consistency loss weight)
    Balances L_CLIP and L_con in Eq. 16; value not reported, chosen by hand.
  • gamma (stage-2 triplet loss weight)
    Balances L_ce and L_trp in Eq. 20; value not reported.
  • triplet margin alpha = 0.2
    Set in Sec. 3.5.2 without a sensitivity analysis.
  • L and M (prompt token counts)
    The numbers of total prompt tokens L and fused tokens M affect text prompt capacity; neither is reported.
  • perturbation schedule = 50% probability for flipping, erasing, cropping; 50% feature dropout
    Hand-chosen augmentation set in Sec. 4.2 with no ablation over perturbation types or strengths.
assumptions (4)
  • domain assumption CLIP's pretrained encoders provide a shared image-text embedding space suitable for person re-identification.
    Assumed from Sec. 3.1 and used throughout; the paper tunes prompts and later the visual encoder rather than re-training CLIP.
  • domain assumption A learnable text prompt of the form 'a photo of a [X_1..X_L] person' can represent a person identity.
    Inherited from CLIP-ReID [20] and used in Eq. 11 and Eq. 17; not validated independently.
  • domain assumption Maximizing cross-modal consistency under perturbations improves robustness to occlusions.
    PDCA premise in Sec. 3.4; no analysis of why text-embedding consistency transfers to visual-only retrieval.
  • domain assumption Text prompts are not needed at inference; visual-only cosine distance suffices for retrieval.
    Sec. 3.5.3 assumes Stage-2 visual tuning creates a standalone descriptor.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCING:Towards More Efficient and Robust Person Re-Identification through Selective Cross-modal Prompt Tuning." pith.science (2026). https://pith.science/paper/G6MZWBPM

@misc{pith2026250700506,
  author       = {Pith},
  title        = {Pith review of: SCING:Towards More Efficient and Robust Person Re-Identification through Selective Cross-modal Prompt Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G6MZWBPM}},
  note         = {Machine review of arXiv:2507.00506}
}
read the original abstract

Recent advancements in adapting vision-language pre-training models like CLIP for person re-identification (ReID) tasks often rely on complex adapter design or modality-specific tuning while neglecting cross-modal interaction, leading to high computational costs or suboptimal alignment. To address these limitations, we propose a simple yet effective framework named Selective Cross-modal Prompt Tuning (SCING) that enhances cross-modal alignment and robustness against real-world perturbations. Our method introduces two key innovations: Firstly, we proposed Selective Visual Prompt Fusion (SVIP), a lightweight module that dynamically injects discriminative visual features into text prompts via a cross-modal gating mechanism. Moreover, the proposed Perturbation-Driven Consistency Alignment (PDCA) is a dual-path training strategy that enforces invariant feature alignment under random image perturbations by regularizing consistency between original and augmented cross-modal embeddings. Extensive experiments are conducted on several popular benchmarks covering Market1501, DukeMTMC-ReID, Occluded-Duke, Occluded-REID, and P-DukeMTMC, which demonstrate the impressive performance of the proposed method. Notably, our framework eliminates heavy adapters while maintaining efficient inference, achieving an optimal trade-off between performance and computational overhead. The code will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2507.00506 by the authors.

Figure 1
Figure 1. UMAP visualization [30] shows the image-text modality gap using red dots for image embeddings and blue dots for paired text embeddings connected by grey lines. We comparing modality gap among CLIP-ReID, CoCoOp, and our approach. CLIP-ReID (a) and CoCoOp (b) exhibit a significant modality gap between person images and their correspond￾ing textual descriptions, while our model (c) significantly reduces this gap, achie… view at source ↗
Figure 2
Figure 2. Performance comparison of different methods [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Three approaches for integrating visual features into text prompts. CoCoOp (a) compresses image information into [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Model framework. For a given image and its learnable text description, we first use Selective Visual Prompt Fusion [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Visualization of different prompt learning models [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Reliability-Aware 3D Geometric Injection for Universal Person Re-identification

    cs.CV 2026-07 conditional novelty 5.0 of 10

    UniGeo adds monocular SMPL body geometry to a 2D person-ReID model through a learned gate that suppresses unreliable 3D, improving occlusion, clothing-change, and cross-modality benchmarks without hurting clean ones.

Reference graph

Works this paper leans on

68 extracted references · 54 canonical work pages · cited by 1 Pith paper

  1. [1]

    Peixian Chen, Wenfeng Liu, Pingyang Dai, Jianzhuang Liu, Qixiang Ye, Mingliang Xu, Qi’an Chen, and Rongrong Ji. 2021. Occlude them all: Occlusion-aware attention network for occluded person re-id. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 11833–11842

    work page 2021

  2. [2]

    Wei Chen, Yu Liu, Weiping Wang, Erwin M Bakker, Theodoros Georgiou, Paul Fieguth, Li Liu, and Michael S Lew. 2022. Deep learning for instance retrieval: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 6 (2022), 7270–7292

    work page 2022

  3. [3]

    De Cheng, Zhipeng Xu, Xinyang Jiang, Nannan Wang, Dongsheng Li, and Xinbo Gao. 2024. Disentangled prompt representation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 23595–23604

    work page 2024

  4. [4]

    Xinhua Cheng, Mengxi Jia, Qian Wang, and Jian Zhang. 2022. More is better: Multi-source Dynamic Parsing Attention for Occluded Person Re-identification. In Proceedings of the ACM International Conference on Multimedia . 6840–6849

    work page 2022

  5. [5]

    Can Cui, Siteng Huang, Wenxuan Song, Pengxiang Ding, Min Zhang, and Donglin Wang. 2024. ProFD: Prompt-Guided Feature Disentangling for Occluded Person Re-Identification. In Proceedings of the 32nd ACM International Conference on Multimedia. 1583–1592

    work page 2024

  6. [6]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    arXiv 2020

  7. [7]

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2024. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision 132, 2 (2024), 581–595

    2024

  8. [8]

    Shang Gao, Jingya Wang, Huchuan Lu, and Zimo Liu. 2020. Pose-guided visible part matching for occluded person reid. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 11744–11752

    work page 2020

Show all 68 references
  1. [9]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

    2016

  2. [10]

    Lingxiao He and Wu Liu. 2020. Guided saliency feature learning for person re-identification in crowded scenes. In Proceedings of the European Conference on Computer Vision. Springer, 357–373

    2020

  3. [11]

    Shuting He, Weihua Chen, Kai Wang, Hao Luo, Fan Wang, Wei Jiang, and Henghui Ding. 2023. Region generation and assessment network for occluded person re-identification. IEEE Transactions on Information Forensics and Security (2023)

    2023

  4. [12]

    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

  5. [13]

    Houjing Huang, Dangwei Li, Zhang Zhang, Xiaotang Chen, and Kaiqi Huang

  6. [14]

    Jiaxing Huang, Kai Jiang, Jingyi Zhang, Han Qiu, Lewei Lu, Shijian Lu, and Eric Xing. 2024. Learning to prompt segment anything models. arXiv preprint arXiv:2401.04651 (2024)

    2024 arXiv

  7. [15]

    Mengxi Jia, Xinhua Cheng, Shijian Lu, and Jian Zhang. 2022. Learning dis- entangled representation implicitly via transformer for occluded person re- identification. IEEE Transactions on Multimedia 25 (2022), 1294–1305

    2022

  8. [16]

    Mengxi Jia, Yifan Sun, Yunpeng Zhai, Xinhua Cheng, Yi Yang, and Ying Li. 2023. Semi-attention partition for occluded person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 998–1006

    2023

  9. [17]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  10. [18]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning . PMLR, 12888–12900

    2022

  11. [19]

    Qi Li, Runpeng Yu, and Xinchao Wang. 2024. Encapsulating Knowledge in One Prompt. In European Conference on Computer Vision . Springer, 215–232

    2024

  12. [20]

    Siyuan Li, Li Sun, and Qingli Li. 2023. 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. 1405–1413

    2023

  13. [21]

    Wen Li, Cheng Zou, Meng Wang, Furong Xu, Jianan Zhao, Ruobing Zheng, Yuan Cheng, and Wei Chu. 2023. Dc-former: Diverse and compact transformer for person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 1415–1423

    2023

  14. [22]

    Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu

  15. [23]

    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

  16. [24]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning. Advances in neural information processing systems 36 (2023), 34892–34916

    2023

  17. [25]

    Liping Lu, Zihao Fu, Duanfeng Chu, Wei Wang, and Bingrong Xu. 2025. CLIP- SENet: CLIP-based Semantic Enhancement Network for Vehicle Re-identification. arXiv preprint arXiv:2502.16815 (2025)

    2025 arXiv

  18. [26]

    Hugo Masson, Amran Bhuiyan, Le Thanh Nguyen-Meidine, Mehrsan Javan, Parthipan Siva, Ismail Ben Ayed, and Eric Granger. 2019. A Survey of Pruning Methods for Efficient Person Re-identification Across Domains. arXiv preprint arXiv:1907.02547 (2019)

    2019 arXiv

  19. [27]

    Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. 2019. Pose-guided feature alignment for occluded person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 542–551

    2019

  20. [28]

    Yunjie Peng, Jinlin Wu, Boqiang Xu, Chunshui Cao, Xu Liu, Zhenan Sun, and Zhiqiang He. 2023. Deep Learning Based Occluded Person Re-Identification: A Survey. ACM Transactions on Multimedia Computing, Communications and Applications 20, 3 (2023), 1–27

    2023

  21. [29]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  22. [30]

    Tim Sainburg, Leland McInnes, and Timothy Q Gentner. 2021. Parametric UMAP embeddings for representation and semisupervised learning. Neural Computation 33, 11 (2021), 2881–2907

    2021

  23. [31]

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, and Dhruv Batra. 2020. Grad-CAM: visual explanations from deep networks via gradient-based localization. International journal of computer vision 128 (2020), 336–359

    2020

  24. [32]

    Vladimir Somers, Christophe De Vleeschouwer, and Alexandre Alahi. 2023. Body part-based representation learning for occluded person Re-Identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 1613–1623

    2023

  25. [33]

    Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. 2020. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6398–6407

    2020

  26. [34]

    Yifan Sun, Qin Xu, Yali Li, Chi Zhang, Yikang Li, Shengjin Wang, and Jian Sun

  27. [35]

    Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. 2018. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In Proceedings of the European Conference on Computer Vision. 480–496. Conference’17, July 2017, Washington,...

    2018

  28. [36]

    Hongchen Tan, Xiuping Liu, Baocai Yin, and Xin Li. 2022. MHSA-Net: Multihead self-attention network for occluded person re-identification. IEEE Transactions on Neural Networks and Learning Systems (2022)

    2022

  29. [37]

    Lei Tan, Pingyang Dai, Rongrong Ji, and Yongjian Wu. 2022. Dynamic pro- totype mask for occluded person re-identification. In Proceedings of the ACM International Conference on Multimedia . 531–540

    2022

  30. [38]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  31. [39]

    Guan’an Wang, Shuo Yang, Huanyu Liu, Zhicheng Wang, Yang Yang, Shuliang Wang, Gang Yu, Erjin Zhou, and Jian Sun. 2020. High-order information matters: Learning relation and topology for occluded person re-identification. In Pro- ceedings of the IEEE/CVF Conference on Computer ...

    2020

  32. [40]

    Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. 2018. Learning discriminative features with multiple granularities for person re-identification. In Proceedings of the ACM International Conference on Multimedia . 274–282

    2018

  33. [41]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191 (2024)

    2024 arXiv

  34. [42]

    Ran Wang, Hua Zuo, Zhen Fang, and Jie Lu. 2024. Prompt-Based Memory Bank for Continual Test-Time Domain Adaptation in Vision-Language Models. In2024 International Joint Conference on Neural Networks (IJCNN) . IEEE, 1–8

    2024

  35. [43]

    Tao Wang, Hong Liu, Pinhao Song, Tianyu Guo, and Wei Shi. 2022. Pose-guided feature disentangling for occluded person re-identification based on transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 2540–2549

    2022

  36. [44]

    Zhikang Wang, Feng Zhu, Shixiang Tang, Rui Zhao, Lihuo He, and Jiangning Song

  37. [45]

    Xiu-Shen Wei, Yi-Zhe Song, Oisin Mac Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. 2021. Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence 44, 12 (2021), 8927–8948

    2021

  38. [46]

    Jiayu Jiang Fei Wang Yibing Zhan Dapeng Tao Wentao Tan, Changxing Ding

  39. [47]

    Boqiang Xu, Lingxiao He, Jian Liang, and Zhenan Sun. 2022. Learning feature recovery transformer for occluded person re-identification. IEEE Transactions on Image Processing 31 (2022), 4651–4662

    2022

  40. [48]

    Jinrui Yang, Jiawei Zhang, Fufu Yu, Xinyang Jiang, Mengdan Zhang, Xing Sun, Ying-Cong Chen, and Wei-Shi Zheng. 2021. Learning to know where to see: A visibility-aware approach for occluded person re-identification. In Proceedings of the IEEE/CVF International Conference on Com...

    2021

  41. [49]

    Yajing Zhai, Yawen Zeng, Zhiyong Huang, Zheng Qin, Xin Jin, and Da Cao. 2024. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 6979–6987

    2024

  42. [50]

    Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. 2024. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 17117–17126

    2024

  43. [51]

    Ziwen Zhang, Shoudong Han, Donghaisheng Liu, and Delie Ming. 2024. Focus and imagine: Occlusion suppression and repairing transformer for occluded person re-identification. Neurocomputing 578 (2024), 127442

    2024

  44. [52]

    Zhimeng Zhang, Jianan Wu, Xuan Zhang, and Chi Zhang. 2017. Multi- Target, Multi-Camera Tracking by Hierarchical Clustering: Recent Progress on DukeMTMC Project. arXiv preprint arXiv:1712.09531 (2017)

    2017 arXiv

  45. [53]

    Liming Zhao, Xi Li, Yueting Zhuang, and Jingdong Wang. 2017. Deeply-learned part-aligned representations for person re-identification. In Proceedings of the IEEE international conference on computer vision . 3219–3228

    2017

  46. [54]

    Kecheng Zheng, Cuiling Lan, Wenjun Zeng, Jiawei Liu, Zhizheng Zhang, and Zheng-Jun Zha. 2021. Pose-guided feature learning with knowledge distillation for occluded person re-identification. In Proceedings of the ACM International Conference on Multimedia. 4537–4545

    2021

  47. [55]

    Liang Zheng, Zhi Bie, Yifan Sun, Jingdong Wang, Chi Su, Shengjin Wang, and Qi Tian. 2016. Mars: A video benchmark for large-scale person re-identification. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Nether- lands, October 11-14, 2016, Proceedings, P...

    2016

  48. [56]

    Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. 2017. Re-ranking per- son re-identification with k-reciprocal encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1318–1327

    2017

  49. [57]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2021. Learning to Prompt for Vision-Language Models. International Journal of Computer Vision 130 (2021), 2337 – 2348. https://api.semanticscholar.org/CorpusID:237386023

    2021

  50. [58]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional Prompt Learning for Vision-Language Models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2022), 16795–16804. https: //api.semanticscholar.org/CorpusID:247363011

    2022

  51. [59]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 16816–16825

    2022

  52. [60]

    Kuan Zhu, Haiyun Guo, Zhiwei Liu, Ming Tang, and Jinqiao Wang. 2020. Identity- guided human semantic parsing for person re-identification. In Proceedings of the European Conference on Computer Vision . Springer, 346–363

    2020

  53. [61]

    Jiaxuan Zhuo, Zeyu Chen, Jianhuang Lai, and Guangcong Wang. 2018. Occluded Person Re-Identification. Proceedings of the IEEE International Conference on Multimedia and Expo (2018), 1–6. https://api.semanticscholar.org/CorpusID: 4713514

    2018

  54. [62]

    Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. 2024. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 22010–22019

    2024

  55. [2018]

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

    Adversarially occluded samples for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5098– 5107

  56. [2019]

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

    Perceive where to focus: Learning visibility-aware part-level features for partial person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 393–402

  57. [2021]

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

    Diverse part discovery: Occluded person re-identification with part-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2898–2907

  58. [2022]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Feature erasing and diffusion network for occluded person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. 4754–4763

  59. [2023]

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

    Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision. 4015–4026

  60. [2024]

    CVPR (2024)

    Harnessing the Power of MLLMs for Transferable Text-to-Image Person ReID. CVPR (2024)

    2024

Pith tools

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