Pith. sign in

REVIEW 3 major objections 3 minor 44 references

Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that interactive test-time query enrichment with a multimodal large language model can improve text-to-image person re-identification by vetting top-ranked candidates and merging the answers into an enriched query before…

desk verdict Solid MLLM reranking package for TIReID, but Eq. (6) and a missing control leave the actual source of the Rank-1 gain unproven. read the letter →

arxiv 2506.11036 v1 pith:Y54TWEQK submitted 2025-05-21 cs.LG cs.MM

Yang Qin , Chao Chen , Zhihang Fu , Dezhong Peng , Xi Peng , Peng Hu This is my paper
classification cs.LGcs.MM
keywords text-to-imagepersonre-identificationmultimodallargelanguagemodeltest-timeinteractionvisualquestionansweringqueryenrichmentdataaugmentationre-ranking
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 claims that text-to-image person re-identification (retrieving a person's image from a written description) can be made substantially more accurate by an interactive plug-and-play test-time module that consults a multimodal large language model before final ranking. The module, called the Test-time Human-centered Interaction (THI) module, asks the MLLM whether each top-ranked candidate matches the query, and once a candidate is accepted, asks human-centered questions about visible attributes such as clothing, carried objects, and shoes; the answers are merged into an enriched query that is used to re-rank the gallery. A second component, called the Reorganization Data Augmentation (RDA) strategy, enriches, decomposes, rewrites, and recombines training descriptions so that the offline model itself becomes more discriminative and generalizes better across datasets. On four benchmarks the paper reports Rank-1 gains of 1.50, 0.91, and 2.85 percentage points from THI on CUHK-PEDES, ICFG-PEDES, and RSTPReid, Rank-1 above 91% on UFine6926, and consistent transfer when THI is added to existing models such as CLIP, IRRA, and RDE. If this is right, trained re-identification models can be improved after the fact without retraining, and written queries that are too vague or incomplete can be repaired at query time.

What carries the argument

The load-bearing machinery is the anchor-localization loop together with the re-ranking similarity $\hat S_{q,v} = \lambda S_{q,v} + (1-\lambda) \bar S_{\hat q,v}$, where $\bar S_{\hat q,\hat v_1}=1$ for the MLLM-confirmed anchor image and $\lambda$ balances the raw query against the enriched query. In each of $K$ rounds the MLLM answers the binary question "Can this text accurately describe the image?" for the current top candidate; on a "Yes" answer, human-centered visual question answering collects fine-grained attributes, and the MLLM merges the answers into a single fluent enriched query $\hat q$. LoRA (low-rank adaptation) fine-tuning of the MLLM on positive and hard-negative image-text pairs makes the binary judgment reliable enough to act as an anchor. RDA is the training-time complement: enriched texts are decomposed into attribute sub-sentences, rewritten in multiple styles, randomly reordered, and recombined, increasing the diversity and discriminability of the training queries.

What would settle it

Take a test split with ground-truth person identities, record for each query which candidate the MLLM says "Yes" to, and check whether that anchor has the same identity as the query. If anchor accuracy is no better than chance, or if disabling the anchor promotion (replacing $\bar S_{\hat q,\hat v_1}=1$ with the raw similarity) leaves the Rank-1 gains unchanged, the observed improvements cannot be attributed to the claimed anchor-localization mechanism.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that the bottleneck in text-to-image person re-identification is not only the cross-modal encoder but also the query itself, and that external multimodal knowledge can repair both at different stages. At test time, THI locates the image the user is looking for by asking the MLLM a yes/no question over the top-ranked candidates; for the accepted anchor it collects attribute-level answers through visual question answering, merges them into a fluent enriched query, and re-ranks with a weighted blend of the original and enriched similarities while setting the anchor's similarity to 1. At training time, RDA uses the same enrichment pipeline and then decomposes enriched texts into attribute sub-sentences, rewrites each sub-sentence in multiple styles, and randomly reorders and recombines them, producing more discriminative and more diverse training queries. The paper reports that each component contributes independently, that THI transfers to several off-the-shelf models, and that the combination reaches the best reported Rank-1 on all four benchmarks among the compared methods.

Load-bearing premise

The argument rests on the LoRA-tuned MLLM's binary text-matches-image judgment being accurate enough that the image it selects as the anchor is usually the person the user wants; if the anchor is wrong, the method promotes a wrong image to the top, and the paper's own numbers show Rank-5 and Rank-10 degrade in that case.

Editorial extensions

If this is right

  • If THI works as reported, any existing text-to-image person re-identification model can be upgraded at inference time without retraining or weight changes; the paper reports average Rank-1 gains of 2.57, 1.86, and 1.41 points on CLIP, IRRA, and RDE respectively.
  • If RDA works as reported, training on enriched, decomposed, and recombined descriptions lifts not only in-domain accuracy but also cross-domain transfer; for example, on UFine6926-to-CUHK-PEDES the method improves Rank-1 by about 17.5 points over the strongest baseline.
  • Multi-round interaction need not be run to exhaustion, because the reported gains concentrate in the first two rounds; the paper uses five rounds and notes that errors accumulate when the anchor is wrong, which motivates early or confidence-based stopping.
  • The similarity threshold used to decide which queries need interaction means the module can be applied to large galleries while controlling the number of MLLM calls.
  • Because the interactive module is separate from the offline model, it can be stacked on top of future, better text-to-image re-identification models rather than competing with them.

Reading between the lines

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

  • A natural but untested extension is to apply the same anchor-localize-then-enrich loop to other text-based visual retrieval tasks, such as attribute-based product search or surveillance search from witness descriptions; nothing in the mechanism is specific to person re-identification.
  • Because the re-ranking formula hard-codes the anchor's similarity to 1, queries where the MLLM selects a wrong anchor should hurt more than they help; a confidence-weighted variant that weakens or skips low-confidence anchors could recover the Rank-5 and Rank-10 drop the paper reports.
  • The paper leaves open how the gains scale with the MLLM's strength; swapping the MLLM for a weaker or stronger one and re-measuring Rank-1 would separate the contribution of the interaction protocol from the contribution of the underlying model.
  • RDA's rewritten texts may introduce hallucinated attributes, so a consistency filter that keeps only rewrites whose retrieval behavior agrees with the original text could make the augmentation safer, at some cost to diversity.
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 / 3 minor

Summary. The paper proposes an Interactive Cross-modal Learning framework (ICL) for text-to-image person re-identification (TIReID), consisting of a plug-and-play Test-time Human-centered Interaction (THI) module and a Reorganization Data Augmentation (RDA) strategy. THI uses a LoRA-fine-tuned MLLM (Qwen2-VL-7B) to locate a latent anchor image via binary 'Can this text accurately describe the image?' questions, then performs human-centered VQA on the anchor to extract fine-grained attributes, merges the answers into a refined query, and re-ranks gallery images with a similarity combination shown in Eq. (6). RDA enriches training texts via VQA and diversifies them through decomposition, rewriting, and recombination. Experiments on CUHK-PEDES, ICFG-PEDES, RSTPReid, and UFine6926 report Rank-1 improvements over the RDE baseline, transferability to CLIP and IRRA, and cross-dataset generalization gains.

Significance. If validated, the framework would offer a test-time, model-agnostic way to inject external MLLM knowledge into offline TIReID systems, with a training-time augmentation that improves cross-domain generalization. The transferability study in Table 5 and the cross-domain results in Tables 3-4 are useful empirical contributions, and the code is publicly available. However, the central claim that query refinement via VQA is what drives the Rank-1 gains is not yet separated from the hard similarity promotion in Eq. (6), and the headline CUHK-PEDES results are tuned on the test set; these issues make the current evidence inconclusive for the paper's main narrative.

major comments (3)
  1. [Section 3.2, Eq. (6) and Algorithm 1] The re-ranking formula sets \bar{S}_{\hat{q},v}\equiv 1 for v=\hat{v}_1, but \hat{v}_1 is defined in Eq. (1) as the original Top-1 candidate, not the MLLM-confirmed anchor. In the k>1 branch of Algorithm 1 (lines 9-12), the MLLM says 'Yes' to the k-th candidate while the Top-1 similarity is below \xi; applying Eq. (6) then force-promotes the original Top-1 even when the MLLM rejected it, contradicting the described mechanism. If the intended target is the located anchor (\hat{v}_k), the equation must say so explicitly. As written, the implementation is ambiguous, so the Rank-1 gains in Table 1 cannot be cleanly attributed to the proposed interaction.
  2. [Section 4.3, Eqs. (4)-(6)] No experiment isolates whether the reported Rank-1 improvement comes from the VQA-refined query \hat{q} or from the hard assignment of similarity 1 to a single image. Setting the anchor's similarity to 1 can by itself push that image to rank 1 even if the VQA-derived text is never used. A control that replaces \hat{q} with the raw query q in Eq. (6) while keeping the hard 1-assignment is needed to demonstrate that the refined query, rather than the binary anchor-localization decision, is responsible for the gains. Without this control, the paper's stated mechanism of 'refining user queries' is not established.
  3. [Section 4.6, Figure 5, Tables 1-2] The hyperparameters \xi and \lambda are selected by a parameter study on the full CUHK-PEDES benchmark, which is also the dataset on which the final results are reported. This makes the CUHK-PEDES numbers in Table 1 optimistic, and the chosen values are applied to the other benchmarks without a dedicated validation split or evidence of transfer. In addition, the paper reports no multi-seed repetitions, standard deviations, or significance tests, so the robustness of small differences (e.g., the 0.91% Rank-1 gain on ICFG-PEDES) is unclear.
minor comments (3)
  1. [Title and Abstract] The phrase 'Humane-centered' in the title and abstract should be 'Human-centered'; the same typo appears in the running text (e.g., Section 3.2 heading).
  2. [Section 3.2, Eq. (3) and Eq. (5)] There are typos in the text: 'Pormpt' should be 'Prompt' in Eq. (3), and 'Taggr is is the prompt template' in the sentence after Eq. (5) has a duplicated 'is'.
  3. [Section 4.1] The dataset name is written as 'CHUK-PEDES' in the first sentence; it should be 'CUHK-PEDES' to match the rest of the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; THI and RDA are self-contained empirical modules evaluated on held-out queries.

full rationale

The derivation chain does not reduce to the paper's own inputs. The LoRA-fine-tuned MLLM is trained on the training split (Eq. 3) and then applied to held-out test queries, so the anchor-localization and VQA responses are genuine test-time predictions rather than fitted quantities reused as outputs. Eq. (6) re-ranks by a convex combination of the original cross-modal similarity and the refined-query similarity; the special case that forces one candidate's similarity to 1 is an explicit reranking mechanism, not a definition of the target result, and it is not derived from or equivalent to the claim that query refinement improves accuracy. RDA augments training texts from existing image-text pairs via MLLM decomposition and reorganization; this is data augmentation, not a prediction equal to its own input. Self-citations such as RDE [21] serve as an externally published baseline/backbone and are not load-bearing support for the framework's core assumptions. The paper's admission that incorrect anchors degrade Rank-5/Rank-10 is a limitation of the interactive mechanism, and the absence of an ablation separating hard anchor promotion from VQA-refined query refinement is an experimental confound, not a circularity. No label leakage or fitted-parameter-renamed-as-prediction pattern is present.

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

The central innovation rests on the reliability of an external MLLM's visual judgment and on two hyperparameters (xi and lambda) tuned on a benchmark. No new physical or conceptual entities are introduced.

free parameters (4)
  • xi (similarity threshold) = 0.5 to 0.6
    Filters which queries and candidates trigger MLLM interaction; set based on a parameter study on CUHK-PEDES (Section 4.6, Figure 5a), likely on its test set.
  • lambda (rerank balance factor) = 0.8
    Balances original similarity and refined-query similarity in Eq. (6); set via parameter study on CUHK-PEDES (Figure 5b).
  • K (interaction rounds) = 5
    Number of anchor localization rounds; fixed across experiments. The paper notes that errors accumulate if the anchor is wrong.
  • Nq (number of VQA questions) = not stated
    Number of human-centered questions per anchor is not specified in the main text; prompt templates are deferred to the supplementary.
assumptions (4)
  • domain assumption The MLLM (Qwen2-VL-7B-Instruct) after LoRA SFT can reliably answer whether a text accurately describes a ReID image and answer fine-grained person attribute questions.
    Central to anchor localization in Eq. (2) and VQA in Eq. (4); no independent evaluation of MLLM answer accuracy is provided beyond downstream Rank-1 changes.
  • domain assumption The cross-modal similarity S_{q,v} encapsulates useful ranking information for anchor selection, and high-similarity top candidates are more likely to be correct.
    Used to set the xi threshold and to decide which images to interact with in Algorithm 1, supported only by the empirical density plots in Figure 3.
  • ad hoc to paper The values of xi and lambda chosen on CUHK-PEDES generalize to ICFG-PEDES, RSTPReid, and UFine6926.
    The paper applies one setting across all benchmarks and does not describe a separate validation split for hyperparameter selection.
  • domain assumption MLLM-generated enriched, decomposed, and reorganized texts preserve identity-relevant semantics and do not introduce hallucinated attributes that mislead training.
    RDA relies on MLLM outputs as training text; no human verification or automatic filtering of generated texts is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification." pith.science (2026). https://pith.science/paper/Y54TWEQK

@misc{pith2026250611036,
  author       = {Pith},
  title        = {Pith review of: Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y54TWEQK}},
  note         = {Machine review of arXiv:2506.11036}
}
read the original abstract

Despite remarkable advancements in text-to-image person re-identification (TIReID) facilitated by the breakthrough of cross-modal embedding models, existing methods often struggle to distinguish challenging candidate images due to intrinsic limitations, such as network architecture and data quality. To address these issues, we propose an Interactive Cross-modal Learning framework (ICL), which leverages human-centered interaction to enhance the discriminability of text queries through external multimodal knowledge. To achieve this, we propose a plug-and-play Test-time Humane-centered Interaction (THI) module, which performs visual question answering focused on human characteristics, facilitating multi-round interactions with a multimodal large language model (MLLM) to align query intent with latent target images. Specifically, THI refines user queries based on the MLLM responses to reduce the gap to the best-matching images, thereby boosting ranking accuracy. Additionally, to address the limitation of low-quality training texts, we introduce a novel Reorganization Data Augmentation (RDA) strategy based on information enrichment and diversity enhancement to enhance query discriminability by enriching, decomposing, and reorganizing person descriptions. Extensive experiments on four TIReID benchmarks, i.e., CUHK-PEDES, ICFG-PEDES, RSTPReid, and UFine6926, demonstrate that our method achieves remarkable performance with substantial improvement.

Figures

Figures reproduced from arXiv: 2506.11036 by the authors.

Figure 1
Figure 1. The illustration of our motivation. When perform [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The illustration of our Test-time Human-centered Interaction (THI) module. THI includes [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The similarity statistics of the Top-1 retrieved items in the test sets of the three benchmarks. It is obvious that the items with higher similarity are more likely to be correct retrievals. question answering (VQA), which is expressed as: r_{\bar {v}}=\mathcal {M}(\mathcal {T}_\text {vqa} (\{c_i\}^{N_q}_{i=1},\bar {v})), \label {eq5} (4) where {ci} Nq i=1 are Nq questions directed at the detail char￾acteristics (e.… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The illustration of our RDA. The purpose of RDA is [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Performance (mAP) versus rounds on three datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Top-10 retrieved results on CUHK-PEDES dataset be [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 5
Figure 5. Figure 5: Variation of performance with different ξ and λ. 4.7. Interactive Study This section further explores the interactive module, i.e. THI. We report the performance (mAP) changes after mul￾tiple rounds of interactions in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 33 canonical work pages

  1. [1]

    Rasa: relation and sensitivity aware representation learning for text-based person search

    Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second Interna- tional Joint Conference on Artificial Intelligence, pages 555– 563, 2023. 2, 6

    work page 2023

  2. [2]

    Language and vision based person re-identification for surveillance systems using deep learning with lip layers

    Maryam Bukhari, Sadaf Yasmin, Sheneela Naz, Muazzam Maqsood, Jehyeok Rew, and Seungmin Rho. Language and vision based person re-identification for surveillance systems using deep learning with lip layers. Image and Vision Com- puting, 132:104658, 2023. 1

    work page 2023

  3. [3]

    An empirical study of clip for text-based person search

    Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 465–473, 2024. 1, 2, 6

    work page 2024

  4. [4]

    Tipcb: A simple but effective part-based convolutional baseline for text-based person search

    Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neuro- computing, 494:171–181, 2022

    work page 2022

  5. [5]

    Semantically self-aligned network for text-to- image part-aware person re-identification

    Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to- image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 5, 6

    arXiv 2021

  6. [6]

    Person retrieval in surveillance using textual query: a review

    Hiren Galiyawala and Mehul S Raval. Person retrieval in surveillance using textual query: a review. Multimedia Tools and Applications, 80(18):27343–27383, 2021. 1

    work page 2021

  7. [7]

    Semi-supervised text-based person search

    Daming Gao, Yang Bai, Min Cao, Hao Dou, Mang Ye, and Min Zhang. Semi-supervised text-based person search. arXiv preprint arXiv:2404.18106, 2024. 1

    arXiv 2024

  8. [8]

    Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-rerank pipeline

    Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-rerank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2, 3

    arXiv 2024

Show all 44 references
  1. [9]

    Chatting with interactive memory for text-based person retrieval

    Chen He, Shenshen Li, Zheng Wang, Hua Chen, Fumin Shen, and Xing Xu. Chatting with interactive memory for text-based person retrieval. Multimedia Systems, 31(1):31,

  2. [10]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2

    2016

  3. [11]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 6

    2021 arXiv

  4. [12]

    Cross-modal implicit relation rea- soning and aligning for text-to-image person retrieval

    Ding Jiang and Mang Ye. Cross-modal implicit relation rea- soning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2787–2797, 2023. 1, 2, 6, 7

    2023

  5. [13]

    Chatting makes perfect: Chat-based image retrieval

    Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischin- ski. Chatting makes perfect: Chat-based image retrieval. Ad- vances in Neural Information Processing Systems, 36, 2024. 2, 3

    2024

  6. [14]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 2

    2022

  7. [15]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In In- ternational conference on machine learning , pages 19730– 19742. PMLR, 2023. 3

    2023

  8. [16]

    Person search with natural lan- guage description

    Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural lan- guage description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1970–1979,

    1970

  9. [17]

    Image clustering with external guidance

    Yunfan Li, Peng Hu, Dezhong Peng, Jiancheng Lv, Jianping Fan, and Xi Peng. Image clustering with external guidance. In Forty-first International Conference on Machine Learn- ing. 2

  10. [18]

    Visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3

    2023

  11. [19]

    Deep evidential learning with noisy correspondence for cross-modal retrieval

    Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948–4956,

  12. [20]

    Cross-modal active complementary learning with self-refining correspondence.Advances in neu- ral information processing systems, 36:24829–24840, 2023

    Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence.Advances in neu- ral information processing systems, 36:24829–24840, 2023. 1

    2023

  13. [21]

    Noisy-correspondence learning for text-to-image person re-identification

    Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197–27206, 2024. 1, 2, 6, 7

    2024

  14. [22]

    Dual consensus anchor learning for fast multi-view cluster- ing

    Yalan Qin, Chuan Qin, Xinpeng Zhang, and Guorui Feng. Dual consensus anchor learning for fast multi-view cluster- ing. IEEE Transactions on Image Processing, 2024. 1

    2024

  15. [23]

    Latent space learning based ensemble clustering

    Yalan Qin, Nan Pu, Nicu Sebe, and Guorui Feng. Latent space learning based ensemble clustering. IEEE Transac- tions on Image Processing, 2025. 1

    2025

  16. [24]

    Learning transferable visual models from natural language supervi- sion

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

    2021

  17. [25]

    Learning granularity-unified representations for text-to-image person re-identification

    Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6

    2022

  18. [26]

    See finer, see more: Implicit modality alignment for text-based person retrieval

    Xiujun Shu, Wei Wen, Haoqian Wu, Keyu Chen, Yiran Song, Ruizhi Qiao, Bo Ren, and Xiao Wang. See finer, see more: Implicit modality alignment for text-based person retrieval. In European Conference on Computer Vision , pages 624–

  19. [27]

    Robust multi-view clustering with noisy cor- respondence

    Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy cor- respondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 1

    2024

  20. [28]

    Harnessing the power of mllms for transferable text-to-image person reid

    Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yib- ing Zhan, and Dapeng Tao. 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, pages 17127–17137, 2024. 1, 2, 3, 6

    2024

  21. [29]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Jun- yang Lin. Qwen2-vl: Enhancing vision-language model’s ...

    2024 arXiv

  22. [30]

    Spectrum-bert: pre-training of deep bidirectional transformers for spectral classification of chi- nese liquors

    Yansong Wang, Yundong Sun, Yansheng Fu, Dongjie Zhu, and Zhaoshuo Tian. Spectrum-bert: pre-training of deep bidirectional transformers for spectral classification of chi- nese liquors. IEEE Transactions on Instrumentation and Measurement, 2024. 2

    2024

  23. [31]

    Multimodal llm en- hanced cross-lingual cross-modal retrieval

    Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm en- hanced cross-lingual cross-modal retrieval. arXiv preprint arXiv:2409.19961, 2024. 3

    2024 arXiv

  24. [32]

    Demo: Decoupled feature-based mixture of experts for multi-modal object re-identification

    Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Demo: Decoupled feature-based mixture of experts for multi-modal object re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1

    2025

  25. [33]

    Vi- taa: Visual-textual attributes alignment in person search by natural language

    Zhe Wang, Zhiyuan Fang, Jun Wang, and Yezhou Yang. Vi- taa: Visual-textual attributes alignment in person search by natural language. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XII 16, pages 402–420. Springer, 2020. 1, 2

    2020

  26. [34]

    Caibc: Capturing all-round in- formation beyond color for text-based person retrieval

    Zijie Wang, Aichun Zhu, Jingyi Xue, Xili Wan, Chao Liu, Tian Wang, and Yifeng Li. Caibc: Capturing all-round in- formation beyond color for text-based person retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5314–5322, 2022. 1

    2022

  27. [35]

    Lapscore: language- guided person search via color reasoning

    Yushuang Wu, Zizheng Yan, Xiaoguang Han, Guanbin Li, Changqing Zou, and Shuguang Cui. Lapscore: language- guided person search via color reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 1624–1633, 2021. 2

    2021

  28. [36]

    Learning comprehensive representations with richer self for text-to-image person re-identification

    Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In Pro- ceedings of the 31st ACM international conference on multi- media, pages 6202–6211, 2023. 6

    2023

  29. [37]

    Clip-driven fine-grained text-image person re-identification

    Shuanglin Yan, Neng Dong, Liyan Zhang, and Jinhui Tang. Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing, 2023. 1, 6

    2023

  30. [38]

    Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark

    Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. 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, pages 4492–4501, 2023. 1, 2, 6

    2023

  31. [39]

    Magic tokens: Select diverse tokens for multi-modal object re-identification

    Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17117–17126, 2024. 1

    2024

  32. [40]

    Llamafac- tory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafac- tory: Unified efficient fine-tuning of 100+ language models. arXiv preprint arXiv:2403.13372, 2024. 6

    2024 arXiv

  33. [41]

    Dssl: Deep surroundings-person separation learning for text-based per- son retrieval

    Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based per- son retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 209–217, 2021. 1, 2, 5

    2021

  34. [42]

    Plip: Language-image pre-training for person representation learning, 2023

    Jialong Zuo, Changqian Yu, Nong Sang, and Changxin Gao. Plip: Language-image pre-training for person representation learning, 2023. 1

    2023

  35. [43]

    Cross- video identity correlating for person re-identification pre- training

    Jialong Zuo, Ying Nie, Hanyu Zhou, Huaxin Zhang, Haoyu Wang, Tianyu Guo, Nong Sang, and Changxin Gao. Cross- video identity correlating for person re-identification pre- training. arXiv preprint arXiv:2409.18569, 2024. 1

    2024 arXiv

  36. [44]

    Ufinebench: Towards text-based person retrieval with ultra- fine granularity

    Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra- fine granularity. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 2201...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.