Pith. sign in

REVIEW 5 major objections 5 minor 57 references

Dual-Granularity Cross-Modal Identity Association for Weakly-Supervised Text-to-Person Image Matching

T0 review · 5 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 The paper claims that weakly supervised text-to-person image matching can reach near-supervised accuracy by mining identity links both within each training batch and against a global visual memory bank, then reinforcing those links with…

desk verdict Solid incremental weakly-supervised text-to-person retrieval paper with real gains and one legitimate robustness question about the fixed similarity threshold. read the letter →

arxiv 2507.06744 v1 pith:N7MQFNQ3 submitted 2025-07-09 cs.CV cs.LGcs.MM

Yafei Zhang , Yongle Shang , Huafeng Li This is my paper
classification cs.CVcs.LGcs.MM
keywords text-basedpersonsearchweaklysupervisedlearningcross-modalmatchingidentityassociationmemorybankconsistencyimage-textretrievalre-identification
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

This paper is trying to establish that weakly supervised text-to-person image matching, where only each image's paired caption is known and identity relations between different images are not labeled, can be made dramatically more accurate by constructing pseudo-identity associations at two granularities instead of relying on clustering. At the local level, the model combines image-image and text-text similarities inside each batch to decide which image-text pairs plausibly share an identity. At the global level, each pedestrian image serves as an anchor against a memory bank of all training images, and the resulting identity links are propagated to texts. An added consistency loss on artificially perturbed image-text pairs is meant to handle hard samples without mining them explicitly. If the claims hold, this would make text-based person search practical under weak supervision, cutting annotation cost while coming close to fully supervised performance.

What carries the argument

The core mechanism is the local-and-global dual-granularity identity association. Locally, the cross-modal identity matrix $\boldsymbol{M}_{v,t}=\tilde{\boldsymbol{M}}_v\odot\tilde{\boldsymbol{M}}_t$ is formed by thresholding image-image and text-text self-similarity matrices and combining them, then softened into a target distribution for a similarity-distribution-matching loss. Globally, image features are used as anchors to select a set $\boldsymbol{J}_i$ of memory-bank images whose similarity exceeds the threshold, and identity labels are exchanged between those images and their paired texts; a dynamic confidence weight in Eq. (14) down-weights weak associations. A third module builds information-asymmetric pairs by image augmentation and text masking, enforcing cross-modal consistency on these deliberately harder pairs. These components together replace clustering-based pseudo-labeling with explicit, thresholded relationship construction at two scales.

What would settle it

Run the same training pipeline with the threshold varied to 0.5, 0.6, and 0.8 on CUHK-PEDES while keeping everything else fixed; if Rank-1 accuracy or the measured accuracy of the constructed identity relations changes by more than a few points, the fixed-threshold assumption fails and the reported gains depend on an uncalibrated hyperparameter.

Watch

Extended reading notes

Core claim

The central claim is that a dual-granularity identity association mechanism, built on CLIP-pretrained encoders and trained with a similarity-distribution-matching loss, substantially improves weakly supervised text-to-person retrieval. Locally, the method binarizes intra-modal similarity matrices at a fixed threshold and takes their elementwise product to obtain a cross-modal identity correspondence matrix within the batch. Globally, it maintains memory banks for both modalities, treats images as visual anchors, selects memory-bank images whose similarity to the current image exceeds the same threshold, and propagates those identities to texts in both directions. A confidence-based dynamic weighting down-weights weakly related global pairs, and an information-asymmetric construction masks text tokens while augmenting images to create harder training pairs with preserved identity. The resulting model reports Rank-1 accuracy of 73.06% on CUHK-PEDES, 63.71% on ICFG-PEDES, and 61.30% on RSTPReid, outperforming the prior weakly supervised method by 3.03, 1.11, and 2.95 percentage points respectively, while exceeding unsupervised pseudo-caption methods by large margins.

Load-bearing premise

The load-bearing premise is that a single fixed similarity threshold of 0.7 correctly separates same-identity from different-identity pairs, both for the intra-batch matrices and for the global memory-bank selection, on all three datasets without per-dataset adjustment.

Editorial extensions

If this is right

  • If correct, weak supervision can substitute for manual identity labeling in text-based person search, yielding a large drop in annotation cost with only a modest performance gap relative to fully supervised methods.
  • The visual-anchor strategy implies that image features are more reliable carriers of identity than text features, a design choice that could transfer to other cross-modal retrieval tasks.
  • The information-asymmetric consistency learning offers a way to create hard samples synthetically, avoiding the need to mine unreliable hard pairs from the training set.
  • The reported gains over clustering-based weakly supervised baselines suggest that explicit local and global relationship construction is a more effective pseudo-labeling strategy than dataset-wide clustering.
  • Single-GPU training at batch size 64 with CLIP backbones suggests the approach is practical to reproduce and deploy without large-scale compute.

Reading between the lines

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

  • A natural extension the paper leaves implicit is that the fixed threshold of 0.7 may need per-dataset calibration; if the visual-anchor principle generalizes, an adaptive threshold based on similarity distributions could enlarge the gains further.
  • Beyond person search, the same dual-granularity identity association—local batch-level correspondence plus global memory-bank anchoring—could be applied to weakly supervised image-text retrieval in domains such as e-commerce or surveillance, where identity labels are scarce.
  • The text-masking strategy in the consistency module could be viewed as a form of dropout in semantic space; an untested variant would mask entire attribute phrases rather than random tokens, potentially preserving identity while making even harder samples.
  • The paper's ablation suggests each module contributes independently; a reader might reasonably test whether the global module helps most on datasets with many images per identity, which would identify when the method is most valuable.
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

5 major / 5 minor

Summary. The paper proposes a weakly supervised text-to-person image matching method composed of three modules: local intra-batch cross-modal relationship construction (LRC), global memory-bank-based relationship construction with feature refinement (GSRC-FR), and information-asymmetric sample-pair construction with consistency learning (IASC-CL). Training uses CLIP backbones and combines contrastive losses with similarity-distribution-matching losses. Experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid report Rank-1 improvements over CPCL of 3.03, 1.11, and 2.95 percentage points, together with module ablations and an analysis of the accuracy of the constructed global relations.

Significance. If the results hold, the paper is a solid contribution to weakly supervised text-based person search: it demonstrates that pseudo-identity construction at both batch and global levels can substantially narrow the gap with supervised methods. The empirical evaluation is broad, covering three benchmarks and comparisons with supervised, unsupervised, and weakly supervised methods, and the module-level ablations are informative. The paper does not release code or models, and no machine-checked proofs are involved; the main strengths are the scope of the experiments and the clearly specified training objectives. However, the central claim is currently only partially supported because the pseudo-identity construction relies on a single fixed threshold that is not analyzed, and the dynamic confidence adjustment mechanism that is listed as a contribution is not isolated in the ablations.

major comments (5)
  1. [Sections 3.2 and 3.3, Eqs. (3) and (12)] The single fixed threshold th=0.7 is the sole gating mechanism for both the local binary correspondence matrix M_vt and the global match selection J_i, yet the paper reports no sensitivity analysis for th, no distributions of the underlying intra-modal similarities, and no per-dataset adjustment. Because CLIP similarity scales can differ between image and text modalities and the three datasets have different description lengths, the claimed Rank-1 gains of 3.03/1.11/2.95 points could change materially at other threshold values. The authors should report Rank-1 accuracy as a function of th on all three datasets, or otherwise justify why one universal threshold is appropriate.
  2. [Section 3.3, Eqs. (14)-(15) and Table 4] The 'dynamic confidence adjustment' described around Eq. (14) is a stated contribution, but it is never isolated in the ablation study. The comparison 'B+LRC' versus 'B+LRC+GSRC-FR' varies the entire GSRC-FR module, which includes global relation construction, memory-bank matching, and Eq. (14) jointly, so the reported improvements cannot be attributed to the adjustment mechanism itself. An ablation that toggles Eq. (14) on and off is needed to support the contribution claim.
  3. [Section 3.3, Eq. (14) and surrounding text] Equation (14) assigns lower target probabilities to the most similar positive samples and higher probabilities to weakly similar ones, because Q'(i,j) = 1 - softmax(sim) and the subsequent softmax in Eq. (15) preserves that ordering. This behavior matches the introduction's statement that samples with weaker associations should receive increased attention, but the sentence immediately after Eq. (14) says the weighting 'strengthens the associations between strongly correlated identity samples while suppressing the impact of weakly related pairs.' These two statements are contradictory and should be reconciled.
  4. [Section 3.2, Eqs. (5)-(6)] With 0 < lambda < 1, the matrix Q defined in Eq. (5) takes values 1 on the diagonal, lambda for accepted off-diagonal correspondences, and -lambda for rejected ones; it never takes the value 0. Consequently, the branch in Eq. (6) that sets Q(i,j) to -infinity when Q(i,j)=0 never executes, so the claimed mechanism to 'further emphasize strong identity correspondences while suppressing weak ones' is vacuous in the presented formulation. The authors should either correct the definition of Q, for example by binarizing before applying Eq. (6), or remove Eq. (6) and describe the actual softening procedure.
  5. [Section 4.2, Section 4.4, and Figure 4] The text masking ratio is selected based on the performance curves in Figure 4 on the three datasets, and the other hyperparameters (th, lambda, alpha, tau) are fixed without any validation-based selection being described. Tuning hyperparameters on the test sets makes the reported improvements less conservative. The authors should state which split was used for any hyperparameter selection, or report the results with a validation-based protocol.
minor comments (5)
  1. [Figure 3 caption] The caption reads 'GSCS-FR' but the module is called GSRC-FR elsewhere; this is a typo.
  2. [Section 3.3, Eq. (14)] Equation (14) uses sim(f^v_i, f^v_j) for all j in J'_i, but for columns of S' that come from the memory bank, the stored feature is the memory-bank feature and not the current batch feature; the notation should distinguish current-batch and memory-bank features.
  3. [Table 2] On ICFG-PEDES the proposed method's mINP (6.20) is slightly below CPCL's mINP (6.31), but the text describes consistent improvements without mentioning this exception.
  4. [Section 3.3, after Eq. (12)] The text says the top-k indices are selected from M^v_m, but the value of k is never specified anywhere in the paper, which affects reproducibility; please state k or clarify that Eq. (12) alone defines the selection.
  5. [Tables 4 and 5] The B+GSRC-FR result on CUHK-PEDES (Table 4) and B+LRCI result on CUHK-PEDES (Table 5) are both 67.56; please verify that this coincidence is correct.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the method's components are derived from external CLIP features and evaluated on fixed test sets, with self-citations only in related work and not load-bearing.

full rationale

The central derivation chain is self-contained. The pseudo-identity construction (Eqs. 3-4, 12-15) thresholds cosine similarities from frozen CLIP encoders; th=0.7 and the masking ratio are stated hyperparameters, while the claimed improvements are measured against fixed test splits in Tables 1-3. The SDM losses (Eqs. 9, 17) align predicted similarities to target distributions built from those thresholds, but no predicted retrieval metric is made equal to a fitted input by construction. The self-citations [14, 15, 37] appear only in related-work context and supervised baseline tables, and are not invoked to justify Eqs. 3-22 or any uniqueness claim. The fixed threshold across datasets and the omitted top-k value are robustness/hyperparameter concerns, not circular steps: they do not make the central result reduce to its inputs by definition. Therefore no circular step can be exhibited.

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

The method's central contribution relies on domain assumptions about similarity and identity (threshold-based matching, visual anchoring, masking preserving identity) and on hyperparameters that are either set by hand or unspecified. No new entities (particles, forces, dimensions) are introduced, so the ledger contains only parameters and assumptions.

free parameters (5)
  • th = 0.7
    Used in Eq. (3) and Eq. (12) to binarize similarities into identity relationships; chosen by hand, sensitivity not reported.
  • lambda = unspecified
    Balancing factor in Eq. (5); no value given in the paper.
  • alpha = unspecified
    Memory bank momentum in Eq. (11); no value given in the paper.
  • masking_ratio = 0.5
    Text masking ratio in IASC-CL; selected based on Figure 4 sweeps.
  • temperature_tau = 0.02
    Temperature in contrastive losses (Eq. 1); fixed, likely tuned.
assumptions (4)
  • domain assumption CLIP pretrained weights provide sufficiently well-aligned image-text features as a starting point.
    The entire method is built on fine-tuning a CLIP pretrained model; no evidence is given that the method would work from scratch.
  • domain assumption Intra-modal similarity above a fixed threshold indicates same identity.
    Used in Eqs. (3)-(4) and Eq. (12) to build pseudo identity correspondences; this is critical for local and global relation construction.
  • domain assumption Visual appearance is a more reliable identity cue than textual descriptions.
    Motivates using images as anchors in global construction; if false, the memory-bank propagation could propagate errors.
  • domain assumption Masking text or augmenting images preserves the underlying identity.
    IASC-CL assumes that asymmetric pairs still share identity; this is required for the consistency loss to be meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual-Granularity Cross-Modal Identity Association for Weakly-Supervised Text-to-Person Image Matching." pith.science (2026). https://pith.science/paper/N7MQFNQ3

@misc{pith2026250706744,
  author       = {Pith},
  title        = {Pith review of: Dual-Granularity Cross-Modal Identity Association for Weakly-Supervised Text-to-Person Image Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7MQFNQ3}},
  note         = {Machine review of arXiv:2507.06744}
}
read the original abstract

Weakly supervised text-to-person image matching, as a crucial approach to reducing models' reliance on large-scale manually labeled samples, holds significant research value. However, existing methods struggle to predict complex one-to-many identity relationships, severely limiting performance improvements. To address this challenge, we propose a local-and-global dual-granularity identity association mechanism. Specifically, at the local level, we explicitly establish cross-modal identity relationships within a batch, reinforcing identity constraints across different modalities and enabling the model to better capture subtle differences and correlations. At the global level, we construct a dynamic cross-modal identity association network with the visual modality as the anchor and introduce a confidence-based dynamic adjustment mechanism, effectively enhancing the model's ability to identify weakly associated samples while improving overall sensitivity. Additionally, we propose an information-asymmetric sample pair construction method combined with consistency learning to tackle hard sample mining and enhance model robustness. Experimental results demonstrate that the proposed method substantially boosts cross-modal matching accuracy, providing an efficient and practical solution for text-to-person image matching.

Figures

Figures reproduced from arXiv: 2507.06744 by the authors.

Figure 1
Figure 1. Comparison between the proposed method and ex [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of the proposed method. The input image-text pairs are first processed by the BFE module to extract [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Analysis of the impact of masking ratio on model [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages

  1. [1]

    Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. 2023. RaSa: Relation and Sensitivity Aware Representa- tion Learning for Text-based Person Search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence . 555–563

    work page 2023

  2. [2]

    Yang Bai, Jingyao Wang, Min Cao, Chen Chen, Ziqiang Cao, Liqiang Nie, and Min Zhang. 2023. Text-based person search without parallel image-text data. In Proceedings of the 31st ACM International Conference on Multimedia . 757–767

    work page 2023

  3. [3]

    Tianlang Chen, Chenliang Xu, and Jiebo Luo. 2018. Improving text-based person search by spatial matching and adaptive threshold. In2018 IEEE Winter Conference on Applications of Computer Vision (W ACV). IEEE, 1879–1887

    work page 2018

  4. [4]

    Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng

    work page

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1 (long and short papers). 4171–4186

    work page 2019

  6. [6]

    Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. 2021. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666 (2021)

    arXiv 2021

  7. [7]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ICLR (2021)

    work page 2021

  8. [8]

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. 1996. A density- based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. 226–231

    1996

Show all 57 references
  1. [9]

    Zhang, Shaoqing Ren, and Jian Su

    Kaiming He, X. Zhang, Shaoqing Ren, and Jian Su. 2016. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 770–778

    2016

  2. [10]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.Neural Computation 9, 8 (1997), 1735–1780

    1997

  3. [11]

    Ding Jiang and Mang Ye. 2023. Cross-modal implicit relation reasoning and align- ing for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 2787–2797

    2023

  4. [12]

    Ya Jing, Chenyang Si, Junbo Wang, Wei Wang, Liang Wang, and Tieniu Tan

  5. [13]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. In ICLR

    2015

  6. [14]

    Fan Li, Hang Zhou, Huafeng Li, Yafei Zhang, and Zhengtao Yu. 2025. Person text-image matching via text-feature interpretability embedding and external attack node implantation. IEEE Transactions on Emerging Topics in Computational Intelligence 9, 2 (2025), 1202–1215

    2025

  7. [15]

    Huafeng Li, Shedan Yang, Yafei Zhang, Dapeng Tao, and Zhengtao Yu. 2025. Progressive Feature Mining and External Knowledge-Assisted Text-Pedestrian Image Retrieval. IEEE Transactions on Multimedia 27 (2025), 1973–1987

    2025

  8. [16]

    Jiayi Li, Min Jiang, Jun Kong, Xuefeng Tao, and Xi Luo. 2024. Learning semantic polymorphic mapping for text-based person retrieval. IEEE Transactions on Multimedia 26 (2024), 10678–10691

    2024

  9. [17]

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. 2021. Align before fuse: Vision and language rep- resentation learning with momentum distillation. In Proceedings of the 35th International Conference on Neural Information...

    2021

  10. [18]

    Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang

  11. [19]

    Zongyi Li, Jianbo Li, Yuxuan Shi, Hefei Ling, Jiazhong Chen, Runsheng Wang, and Shijuan Huang. 2024. Cross-modal generation and alignment via attribute- guided prompt for unsupervised text-based person retrieval. In Proceedings of the Thirty-Third International Joint Conferenc...

    2024

  12. [20]

    Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. 2019. 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 . 1487–1495

    2019

  13. [21]

    Kai Niu, Yan Huang, Wanli Ouyang, and Liang Wang. 2020. Improving description- based person re-identification by multi-granularity image-text alignments. IEEE Transactions on Image Processing 29 (2020), 5542–5556

    2020

  14. [22]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  15. [23]

    Jicheol Park, Dongwon Kim, Boseung Jeong, and Suha Kwak. 2024. PLOT: Text- based person search with part slot attention for corresponding part discovery. In Computer Vision – ECCV 2024: 18th European Conference . Springer, 474–490

    2024

  16. [24]

    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 Proceedings of the 38th International Conf...

    2021

  17. [25]

    Zhiyin Shao, Xinyu Zhang, Changxing Ding, Jian Wang, and Jingdong Wang. 2023. Unified pre-training with pseudo texts for text-to-image person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 11174– 11184

    2023

  18. [26]

    Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. 2022. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia. 5566–5574

    2022

  19. [27]

    Karen Simonyan and Andrew Zisserman. 2015. Very Deep Convolutional Net- works for Large-Scale Image Recognition. In ICLR

    2015

  20. [28]

    A Venkata Subramanyam, Vibhu Dubey, Niranjan Sundararajan, and Brejesh Lall

  21. [29]

    Jintao Sun, Hao Fei, Zhedong Zheng, and Gangyi Ding. 2024. From Data Deluge to Data Curation: A Filtering-WoRA Paradigm for Efficient Text-based Person Search. arXiv preprint arXiv:2404.10292 (2024)

    2024 arXiv

  22. [30]

    Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. 2024. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 17127–17137

    2024

  23. [31]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems. 6000–6010

    2017

  24. [32]

    Chengji Wang, Zhiming Luo, Yaojin Lin, and Shaozi Li. 2021. Text-based person search via multi-granularity embedding learning. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence . 1068–1074

    2021

  25. [33]

    Xiaogang Wang, Gianfranco Doretto, Thomas Sebastian, Jens Rittscher, and Peter Tu. 2007. Shape and appearance context modeling. In2007 IEEE 11th International Conference on Computer Vision . IEEE, 1–8

    2007

  26. [34]

    Yuyu Wang, Chunjuan Bo, Dong Wang, Shuang Wang, Yunwei Qi, and Huchuan Lu. 2019. Language person search with mutually connected classification loss. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2057–2061

    2019

  27. [35]

    Zhe Wang, Zhiyuan Fang, Jun Wang, and Yezhou Yang. 2020. Vitaa: Visual- textual attributes alignment in person search by natural language. In Computer vision–ECCV 2020: 16th European Conference . Springer, 402–420

    2020

  28. [36]

    Yu Wu, Haiguang Wang, Mengxia Wu, Min Cao, and Min Zhang. 2024. LAIP: learning local alignment from image-phrase modeling for text-based person search. In 2024 IEEE International Conference on Multimedia and Expo (ICME) . IEEE, 1–10

    2024

  29. [37]

    Shuanglin Yan, Neng Dong, Shuang Li, and Huafeng Li. 2025. TriMatch: Triple Matching for Text-to-Image Person Re-Identification. IEEE Signal Processing Letters 32 (2025), 806–810

    2025

  30. [38]

    Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. 2023. Learn- ing comprehensive representations with richer self for text-to-image person re-identification. In Proceedings of the 31st ACM International Conference on Mul- timedia. 6202–6211

    2023

  31. [39]

    Shuanglin Yan, Neng Dong, Liyan Zhang, and Jinhui Tang. 2023. Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing 32 (2023), 6032–6046

    2023

  32. [40]

    Shuanglin Yan, Jun Liu, Neng Dong, Liyan Zhang, and Jinhui Tang. 2024. Proto- typical Prompting for Text-to-image Person Re-identification. In Proceedings of the 32nd ACM International Conference on Multimedia . 2331–2340

    2024

  33. [41]

    Shuanglin Yan, Hao Tang, Liyan Zhang, and Jinhui Tang. 2024. Image-specific information suppression and implicit local alignment for text-based person search. IEEE Transactions on Neural Networks and Learning Systems 35, 12 (2024), 17973– 17986

    2024

  34. [42]

    Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu

  35. [43]

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

  36. [44]

    Ying Zhang and Huchuan Lu. 2018. Deep cross-modal projection learning for image-text matching. In Proceedings of the European Conference on Computer Vision (ECCV). 686–701

    2018

  37. [45]

    Shizhen Zhao, Changxin Gao, Yuanjie Shao, Wei-Shi Zheng, and Nong Sang

  38. [46]

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian

  39. [47]

    In Proceedings of the 31st ACM International Conference on Multimedia

    Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia. 4492–4501

  40. [48]

    Zhedong Zheng, Liang Zheng, Michael Garrett, Yi Yang, Mingliang Xu, and Yi-Dong Shen. 2020. Dual-path convolutional image-text embeddings with instance loss. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 16, 2 (2020), 1–23

    2020

  41. [49]

    Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. 2021. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia. 209–217

    2021

  42. [52]

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

    Weakly supervised text-based person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 11395–11404

  43. [55]

    Yanwei Zheng, Xinpeng Zhao, Chuanlin Lan, Xiaowei Zhang, Bowen Huang, Jibin Yang, and Dongxiao Yu. 2024. CPCL: Cross-Modal Prototypical Contrastive Learning for Weakly Supervised Text-based Person Re-Identification. arXiv preprint arXiv:2401.10011 (2024)

    2024 arXiv

  44. [2015]

    In Proceedings of the IEEE ACM MM, 2025, Dublin, Ireland Yafei Zhang, Yongle Shang, and Huafeng Li International Conference on Computer Vision

    Scalable person re-identification: A benchmark. In Proceedings of the IEEE ACM MM, 2025, Dublin, Ireland Yafei Zhang, Yongle Shang, and Huafeng Li International Conference on Computer Vision . 1116–1124

    2025

  45. [2017]

    2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017), 5187–5196

    Person Search with Natural Language Description. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017), 5187–5196

    2017

  46. [2020]

    In Proceedings of the AAAI Conference on Artificial Intelligence , Vol

    Pose-guided multi-granularity attention network for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 34. 11189–11196

  47. [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

  48. [2022]

    Neurocomputing 494 (2022), 171–181

    TIPCB: A simple but effective part-based convolutional baseline for text- based person search. Neurocomputing 494 (2022), 171–181

    2022

  49. [2023]

    In Proceedings of the Fourteenth Indian Conference on Computer Vision, Graphics and Image Processing

    Dense captioning for Text-Image ReID. In Proceedings of the Fourteenth Indian Conference on Computer Vision, Graphics and Image Processing . 1–8

Pith tools

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