Pith. sign in

REVIEW 4 major objections 6 minor 64 references

Achieving Text-based Person Retrieval with Any Granularity

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read This paper argues that text-based person retrieval should handle any query granularity, and it supplies the benchmark, evaluation protocol, and model to make that work: CMAM, which outperforms prior methods at every granularity level.

desk verdict Solid benchmark-plus-baseline paper whose headline gains on the new benchmark are partly co-designed; worth reviewing, but the authors need to address test-set tuning and label circularity. read the letter →

arxiv 2607.21057 v1 pith:5DNHB4FE submitted 2026-07-23 cs.CV

classification cs.CV
keywords text-basedpersonretrievalanygranularityspectrumcross-identitylabelsprobabilisticalignmentorthogonalexpertsMG-EvalCMAM
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

The paper tackles a blind spot in text-based person retrieval: real-world queries vary in specificity from a vague phrase to a full paragraph, yet existing benchmarks and models assume a fixed granularity. It defines granularity as the number of distinct semantic attributes a description conveys (using an 18-attribute ontology), and on that basis builds UFine6926-MG, the first dataset with balanced, granularity-controlled texts across five levels, plus the MG-Eval benchmark, which treats coarse queries as legitimately matching multiple identities. It then proposes CMAM, a model that combines granularity-specialized visual experts, soft-label probabilistic alignment, and granularity-consistency reasoning. The reported result is that CMAM outperforms state-of-the-art methods at every granularity level, for example 50.40% versus 45.12% rank-1 on the new separate-granularity protocol.

What carries the argument

The load-bearing identity is the attribute-overlap soft label: for a coarse query, the overlap between its extracted attribute set and another identity's fine description becomes a probabilistic training target, with Gaussian noise added to model real-world ambiguity. Around that, CMAM uses an orthogonal-expert vision encoder that produces granularity-specific representations, and a granularity-consistency reasoning head that classifies the granularity of image and text and verifies they match. The granularity spectrum itself, five levels defined by counts from an 18-attribute ontology, is the yardstick that makes the whole pipeline trainable and measurable.

What would settle it

Take a held-out set of human-written coarse queries, have independent annotators judge which gallery identities are valid matches for each query, and compare CMAM's ranking against those human labels; if the model's lead over baselines shrinks or disappears when validity is not defined by the fixed attribute ontology, the central claim is undermined.

Watch

Extended reading notes

Core claim

The central claim is that retrieval quality is not one number but a curve over query specificity, and that specificity can be quantified as the count of distinct attributes from a fixed 18-attribute ontology. On that measure the paper builds a five-level granularity spectrum, a dataset with roughly uniform text coverage across those levels, an evaluation suite in which coarse queries are assigned cross-identity labels to reflect that several people may be valid matches, and a model that explicitly adapts to granularity. The reported outcome is that CMAM leads all compared methods at every granularity level, with a rank-1 accuracy of 50.40% on the separate-granularity protocol against 45.12%

Load-bearing premise

The load-bearing premise is that a description's granularity is faithfully captured by the number of distinct attributes it conveys from a fixed 18-attribute ontology, and that the large-model-generated descriptions and attribute extractions used to build the dataset and labels are accurate enough to serve as ground truth.

Editorial extensions

If this is right

  • Coarse queries stop being penalized: evaluation credits models for retrieving visually similar, semantically valid people, which matches how users actually search.
  • Training on a granularity-balanced dataset improves training stability and generalization, per the paper's controlled comparisons, not just accuracy on one granularity.
  • The progressive-granularity protocol gives a concrete way to measure iterative search, where users narrow results by adding detail.
  • The mSD metric, built on continuous similarity distributions, can reveal retrieval-quality differences that rank-based recall misses.
  • Choosing a granularity-appropriate visual representation at inference time yields gains on both the new multi-grained benchmark and standard fixed-granularity benchmarks.

Reading between the lines

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

  • The same recipe, attribute-overlap soft labels plus granularity-disentangled experts, could transfer to other retrieval tasks with intrinsic query ambiguity, such as text-based video search or e-commerce product search.
  • If granularity is better measured by information-theoretic content rather than a fixed attribute ontology, the spectrum boundaries could shift; the paper itself flags the derivation as open.
  • The cross-identity labels in MG-Eval could support a set-retrieval evaluation, where the output is a short list of valid matches instead of a single identity, closer to practical search behavior.
  • Because the dataset and labels rely on large-model-generated annotations, their validity could be probed by independently re-annotating a random subset with human attribute counts and checking whether model rankings change.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces a new paradigm, text-based person retrieval with any granularity, and provides a systematic solution consisting of three parts. First, it proposes a five-level granularity spectrum over an 18-attribute ontology and constructs UFine6926-MG, a multi-grained extension of UFine6926 with balanced granularity-controlled text annotations generated by an LLM/VLM annotation engine. Second, it proposes MG-Eval, a multi-grained benchmark with cross-identity labels that reflect the one-to-many semantics of coarse queries, together with separate and progressive evaluation protocols and a new metric, mSD. Third, it proposes the CMAM framework, which combines orthogonal-expert multi-grained visual perception, probabilistic cross-identity alignment with attribute-overlap soft labels, and granularity-consistent reasoning. Experiments on conventional benchmarks and on MG-Eval are reported, with CMAM showing state-of-the-art or competitive results on CUHK-PEDES, ICFG-PEDES, and UFine6926, and a 5.28% R@1 advantage over the strongest baseline on MG-Eval.

Significance. If the central claims hold, the paper would make a substantial contribution: it provides a practically motivated granularity taxonomy, a balanced multi-granularity dataset, a benchmark with protocols that better reflect coarse-query ambiguity, and a method with several novel components. The diagnostic studies in Sec. 5 are informative, and the release of code and data will be valuable to the community. However, the current evidence for the paper's headline claim—that CMAM significantly outperforms all existing methods across granularity levels—is weakened by the co-design of the MG-Eval benchmark and the PCIA training objective, by test-set hyperparameter selection, and by the absence of statistical uncertainty measures. The conventional-benchmark results are competitive but not overwhelmingly superior (e.g., Table 2: 77.12 vs 75.94 R@1 on CUHK-PEDES against RDE). With additional independent validation of the multi-grained evaluation, this could be a strong and useful contribution.

major comments (4)
  1. [Sec. 4.1, Eq. (13), Table 3] MG-Eval's cross-identity relevance labels are generated by attribute matching between coarse queries and fine-grained texts of other identities, using the same attribute-overlap criterion that PCIA explicitly optimizes as soft labels in Eq. (13). Consequently, CMAM is trained to reproduce MG-Eval's own match definition, whereas baselines are trained with conventional identity-only losses. The reported 5.28% R@1 lead over FGCLIP on MG-Eval is therefore not an independent measure of retrieval quality; it partly measures how well the method internalizes the benchmark's match rule. Please add an evaluation on independently human-labeled cross-identity relevance (e.g., a held-out subset of MG-Eval with manual 'valid match' judgments) or train the baselines with the same PCIA soft-label objective to isolate the architectural contribution. Without such evidence, the abstract's 'significantly ou
  2. [Sec. 7.5, Table 7, Fig. 16] The PCIA hyperparameters θ and σ are selected by sweeping on MG-Eval itself: Table 7 reports average rank-1 on MG-Eval across θ values, and Fig. 16 reports performance on MG-Eval queries across σ values. This is effectively test-set tuning and applies to the final numbers in Table 3. Please select θ and σ on a validation split (or via cross-validation) and report the held-out results, or explicitly state that the reported MG-Eval numbers involve test-set selection. As written, a nontrivial part of the reported gap may reflect overfitting to the benchmark rather than generalizable improvement.
  3. [Sec. 3.1, 3.3, 4.1] The entire granularity construct—the five-level taxonomy, UFine6926-MG granularity labels, PCIA soft labels, and MG-Eval cross-identity judgments—rests on the accuracy and completeness of the 18-attribute ontology and the LLM/VLM-extracted attribute sets. Only 100 identities were human-validated (Sec. 3.3), and Sec. 3.1 explicitly acknowledges that the ontology and level boundaries are 'empirically motivated.' If attribute extraction is systematically wrong or incomplete, all downstream labels and training signals are compromised. Please provide a larger-scale validation of the attribute extraction against human annotations (e.g., per-attribute precision/recall on several hundred identities and inter-annotator agreement on granularity labels), and report quality by granularity level. This is a concrete correctness-risk concern that directly affects the benchmark's validity.
  4. [Tables 2–4, Sec. 7.3] No standard deviations or significance tests are reported. On conventional benchmarks, CMAM's improvements over the strongest baseline are modest (e.g., 77.12 vs 75.94 R@1 on CUHK-PEDES, 88.45 vs 87.60 on UFine6926, 68.13 vs 67.91 on ICFG-PEDES in Table 2) and are likely within run-to-run variation for this task. The claim that CMAM 'significantly outperforms' must be supported by multiple seeds and a paired significance test, especially for the central MG-Eval comparisons (Tables 3 and 4).
minor comments (6)
  1. [Algorithm 1, line 13] Typo: 'Dr = D s \ C s' should presumably be 'D_r = D_r \ C_s' (the set of remaining attribute descriptions must be updated by removing the sampled subset).
  2. [Fig. 6] The 'information entropy' used for granularity distributions is not defined. Please specify how entropy is computed (over words? attributes?) so the scatter plots are reproducible.
  3. [Sec. 5.2, Eq. (8)] The information-theoretic identity involving conditional mutual information is stated without formal definitions of the random variables or a proof. If it is intended as motivation rather than a theorem, please label it as such; otherwise provide a derivation.
  4. [Sec. 7.2] The expert dimensionalities (192, 384, 768, 1536, 3072) and hyperparameters λ1, λ2, τ are stated without ablations, despite the paper's careful sensitivity analysis for θ and σ. Please report at least one ablation or a reference for these choices.
  5. [Table 2] mAP values are missing for ICFG-PEDES in several rows and for UFine6926 in several upper-block rows. For consistency and comparability, please report the same metric set across all methods and datasets.
  6. [Sec. 1, Sec. 3.3] The claim 'first multi-grained dataset with granularity-controlled annotations' should be qualified given that the annotations are largely LLM-generated and only 100 identities were human-validated. Consider stating 'first in the text-based person retrieval literature' and describing the validation rate more prominently.

Circularity Check

2 steps flagged · score 4.0 of 10

MG-Eval's cross-identity labels and CMAM's PCIA soft labels are the same attribute-overlap rule, so the headline MG-Eval gain is partly benchmark co-design; conventional-benchmark results remain independent.

  1. self definitional [Sec. 4.1 (Evaluation Set with Cross-Identity Labels) and Sec. 6.3, Eq. (13)-(15)]
    "The key innovation is assigning cross-identity labels through attribute matching: coarse-grained queries are linked to all identities whose fine-grained attributes semantically encompass them. ... r_{ij} = |A(T_i) ∩ A(T_j^fine)| / |A(T_i)|. We construct the soft similarity matrix S, where S_{ij} = r_{ij} if r_{ij} ≥ θ and 0 otherwise."

    MG-Eval's positive labels are computed by the attribute-overlap ratio between a coarse query and another identity's fine-grained text. PCIA uses the exact same ratio (Eq. 13) as the soft training label, and the model is explicitly optimized to maximize this quantity via L_stal. Therefore a high MG-Eval score partly reflects that CMAM was trained to reproduce the benchmark's own label-generation rule, while baselines trained with hard identity labels never see this rule. The claimed 5.28% R@1 lead on MG-Eval is thus not an independent measure of retrieval quality; the evaluation definition and the training objective are the same construction.

  2. fitted input called prediction [Sec. 7.5, Table 7 and Fig. 16]
    "Performance peaks at θ=0.6, which we fix as the default soft label configuration (denoted '+SF'). ... Additional noise injection peaks at σ=0.1."

    The PCIA hyperparameters θ and σ are selected by peaking rank-1 accuracy on MG-Eval itself (Table 7 and Fig. 16), and the final MG-Eval results in Table 3 are then reported with these selected values. The headline comparison therefore embeds test-set-selected hyperparameters; part of the reported gain is an artifact of tuning on the evaluation benchmark rather than a parameter-free prediction.

full rationale

CMAM's results on conventional benchmarks (Table 2) are trained either on original fixed-granularity data or augmented multi-granularity data and remain competitive (e.g., 77.12% R@1 on CUHK-PEDES), so the central claim is not entirely forced by self-citation or by construction. However, the headline multi-granularity claim rests on MG-Eval, and MG-Eval's cross-identity labels and PCIA's soft-label objective are the same attribute-overlap formula (Sec. 4.1 vs. Sec. 6.3, Eq. 13). This is a co-designed benchmark/objective: a model trained to maximize the attribute-overlap ratio is partially guaranteed to score well on a benchmark whose positives are defined by that same ratio. Additionally, θ and σ are tuned on MG-Eval and then reported as final MG-Eval numbers, further weakening the independence of the 'any granularity' superiority claim. These issues do not reduce the dataset-construction pipeline or the conventional-benchmark comparisons to a self-citation chain. The paper itself acknowledges the granularity ontology is empirically motivated and that only 100 identities were human-validated, which reinforces the concern that the benchmark and method share the same empirically chosen attribute model. Overall, this is partial benchmark circularity; the method still has substantial independent empirical content.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The paper's central claims rest primarily on empirical design choices: the granularity spectrum, the attribute ontology, the annotation pipeline, and the soft-label overlap model. All are plausible but are introduced for this paper and validated only indirectly through the benchmark that was built from the same choices.

free parameters (9)
  • Granularity spectrum thresholds = UC: 1-5, C: 6-8, M: 9-11, F: 12-14, UF: 15-18
    The five-level taxonomy is chosen by hand based on an 18-attribute ontology; the paper explicitly calls it empirically motivated and leaves an information-theoretic derivation open (Sec. 3.1).
  • Attribute ontology (18 attributes) = Gender, skin color, age, height, physique, hairstyle, hair color, top type, top length, top color, bottom type, bottom l
    The ontology defines what counts as an attribute and therefore defines the granularity label; no independent evidence for completeness is provided beyond 'consistency with person ReID ontologies'.
  • Annotation engine constants N_c, N_iv, N_f = 5, 3, 15
    Set to match the granularity spectrum; used to control how many attribute subsets are sampled for generated texts (Alg. 1).
  • PCIA attribute-overlap threshold theta = 0.6
    Selected by sweeping theta on the MG-Eval benchmark and choosing the value that maximizes performance (Table 7). This is test-set hyperparameter selection.
  • PCIA noise scale sigma = 0.1
    Selected by sweeping sigma on MG-Eval results (Fig. 16); peak performance occurs at 0.1.
  • Soft-label temperature tau = 0.015
    Set by hand as a fixed temperature; no sensitivity analysis is reported.
  • Loss weights lambda1, lambda2 = 1, 1
    Both empirically set to 1 (Sec. 7.2), no ablation or sensitivity analysis reported.
  • Expert hidden dimensions = 192, 384, 768, 1536, 3072
    Chosen to give progressive capacity from coarse to fine experts; no principled derivation is provided.
  • mSD scaling factor k = 1
    Arbitrary constant in PNR = 1 - e^{-kx}; no derivation or sensitivity analysis is given (Eq. 3).
assumptions (5)
  • domain assumption The 18-attribute ontology and five-level attribute-count granularity spectrum are a valid and sufficient characterization of textual granularity for person retrieval.
    Invoked throughout Sec. 3 to assign granularity labels to training and evaluation texts; the paper acknowledges the ontology and taxonomy are empirically motivated and that an optimal derivation remains open.
  • domain assumption LLM/VLM-generated multi-grained descriptions and their extracted attribute lists are semantically accurate and attribute-complete.
    The annotation engine (Sec. 3.2) generates the dataset and evaluation queries; only 100 identities were human-validated, with a 94.7% pass rate (Sec. 3.3), so the remaining annotations are assumed correct.
  • domain assumption Attribute-overlap ratio between a coarse text and another identity's fine text is a valid measure of semantic match, used both for evaluation labels and training soft labels.
    Eq. 13 defines r_ij as the match strength; Sec. 4.1 uses the same idea for cross-identity evaluation labels, and Sec. 6.3 uses it for training. No independent user study validates this match model.
  • ad hoc to paper The optimal image-text similarity should be higher for fine-grained texts than for coarse-grained texts, by an amount equal to conditional mutual information of additional attributes.
    Eq. 8 is presented as an information-theoretic motivation for granularity-dependent similarity, but it is not derived from the model and is not used to set any constants; it is a motivational assumption.
  • domain assumption Pretrained FG-CLIP/ViT-Base features provide an adequate backbone for the multi-grained visual and textual encoding.
    The architecture inherits the pretrained cross-modal ViT-Base backbone (Sec. 7.2) and assumes its representation space supports granularity-specific experts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Achieving Text-based Person Retrieval with Any Granularity." pith.science (2026). https://pith.science/paper/5DNHB4FE

@misc{pith2026260721057,
  author       = {Pith},
  title        = {Pith review of: Achieving Text-based Person Retrieval with Any Granularity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DNHB4FE}},
  note         = {Machine review of arXiv:2607.21057}
}
read the original abstract

Text-based person retrieval faces a critical but under-explored challenge: the inherent uncertainty of query granularity in real-world scenarios. This paper introduces a new paradigm, Text-based Person Retrieval with Any Granularity, and provides a systematic solution. First, we formalize a five-level granularity spectrum and construct UFine6926-MG, a high-quality multi-grained dataset annotated comprehensively at all granularities via a novel Multi-grained Text Annotation Engine. Second, acknowledging that coarse queries naturally correspond to multiple valid candidates, we propose MG-Eval, a holistic evaluation benchmark with progressively detailed texts and cross-identity labels that reflect real-world semantics, alongside tailored evaluation metrics and protocols. Third, after a comprehensive diagnosis reveals the systemic limitations of existing research, we propose the Cross-modal Multi-grained Aligning and Matching (CMAM) framework. CMAM achieves granularity-aware retrieval through: 1) orthogonal-expert perception to disentangle granularity-specific features; 2) probabilistic alignment to model many-to-many matches under query uncertainty; and 3) granularity-consistent reasoning to steer feature learning via joint cross-modal granularity verification. Experiments demonstrate that CMAM significantly outperforms state-of-the-art methods across all granularity levels. This work establishes a foundational benchmark and a robust baseline, paving the way for more practical person retrieval systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 3 linked inside Pith

  1. [1]

    Person search with natural language description,

    S. Li, T. Xiao, H. Li, B. Zhou, D. Yue, and X. Wang, “Person search with natural language description,” inCVPR, 2017, pp. 1970–1979

  2. [2]

    Scalable person re-identification: A benchmark,

    L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian, “Scalable person re-identification: A benchmark,” inICCV, 2015, pp. 1116– 1124

  3. [3]

    Person transfer gan to bridge domain gap for person re-identification,

    L. Wei, S. Zhang, W. Gao, and Q. Tian, “Person transfer gan to bridge domain gap for person re-identification,” inCVPR, 2018, pp. 79–88

  4. [4]

    Transreid: Transformer-based object re-identification,

    S. He, H. Luo, P . Wang, F. Wang, H. Li, and W. Jiang, “Transreid: Transformer-based object re-identification,” inICCV, 2021, pp. 15 013–15 022

  5. [5]

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

    J. Zuo, Y. Nie, H. Zhou, H. Zhang, H. Wang, T. Guo, N. Sang, and C. Gao, “Cross-video identity correlating for person re- identification pre-training,”NeurIPS, vol. 37, pp. 25 228–25 250, 2024

  6. [6]

    Deep learning for person re-identification: A survey and outlook,

    M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. Hoi, “Deep learning for person re-identification: A survey and outlook,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 6, pp. 2872–2893, 2021

  7. [7]

    Transformer for object re-identification: A survey,

    M. Ye, S. Chen, C. Li, W.-S. Zheng, D. Crandall, and B. Du, “Transformer for object re-identification: A survey,”International Journal of Computer Vision, vol. 133, no. 5, pp. 2410–2440, 2025

  8. [8]

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

    Z. Ding, C. Ding, Z. Shao, and D. Tao, “Semantically self-aligned network for text-to-image part-aware person re-identification,” arXiv preprint arXiv:2107.12666, 2021

Show all 64 references
  1. [9]

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

    A. Zhu, Z. Wang, Y. Li, X. Wan, J. Jin, T. Wang, F. Hu, and G. Hua, “Dssl: Deep surroundings-person separation learning for text-based person retrieval,” inACMMM, 2021, pp. 209–217

  2. [10]

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

    J. Zuo, H. Zhou, Y. Nie, F. Zhang, T. Guo, N. Sang, Y. Wang, and C. Gao, “Ufinebench: Towards text-based person retrieval with ultra-fine granularity,” inCVPR, 2024, pp. 22 010–22 019

  3. [11]

    Chat-based person retrieval via dialogue-refined cross-modal alignment,

    Y. Bai, Y. Ji, M. Cao, J. Wang, and M. Ye, “Chat-based person retrieval via dialogue-refined cross-modal alignment,” inCVPR, 2025, pp. 3952–3962. 17

  4. [12]

    LLaVA- reID: Selective multi-image questioner for interactive person re- identification,

    Y. Lu, M. Yang, D. Peng, P . Hu, Y. Lin, and X. Peng, “LLaVA- reID: Selective multi-image questioner for interactive person re- identification,” inICML, 2025

  5. [13]

    Human- centered interactive learning via mllms for text-to-image person re-identification,

    Y. Qin, C. Chen, Z. Fu, D. Peng, X. Peng, and P . Hu, “Human- centered interactive learning via mllms for text-to-image person re-identification,” inCVPR, 2025, pp. 14 390–14 399

  6. [14]

    Qwen2.5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P . Wang, S. Wang, J. Tanget al., “Qwen2.5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

  7. [15]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

  8. [16]

    Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval,

    D. Jiang and M. Ye, “Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval,” inCVPR, 2023, pp. 2787–2797

  9. [17]

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

    Y. Qin, Y. Chen, D. Peng, X. Peng, J. T. Zhou, and P . Hu, “Noisy-correspondence learning for text-to-image person re- identification,” inCVPR, 2024, pp. 27 197–27 206

  10. [18]

    FG-CLIP: Fine-grained visual and textual alignment,

    C. Xie, B. Wang, F. Kong, J. Li, D. Liang, G. Zhang, D. Leng, and Y. Yin, “FG-CLIP: Fine-grained visual and textual alignment,” in ICML, 2025

  11. [19]

    Plip: Language-image pre-training for person represen- tation learning,

    J. Zuo, J. Hong, F. Zhang, C. Yu, H. Zhou, C. Gao, N. Sang, and J. Wang, “Plip: Language-image pre-training for person represen- tation learning,”NeurIPS, vol. 37, pp. 45 666–45 702, 2024

  12. [20]

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

    S. Yang, Y. Zhou, Z. Zheng, Y. Wang, L. Zhu, and Y. Wu, “Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark,” inACMMM, 2023, pp. 4492– 4501

  13. [21]

    Deep cross-modal projection learning for image-text matching,

    Y. Zhang and H. Lu, “Deep cross-modal projection learning for image-text matching,” inECCV, 2018, pp. 686–701

  14. [22]

    Dual-path convolutional image-text embeddings with instance loss,

    Z. Zheng, L. Zheng, M. Garrett, Y. Yang, M. Xu, and Y.-D. Shen, “Dual-path convolutional image-text embeddings with instance loss,”ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), vol. 16, no. 2, pp. 1–23, 2020

  15. [23]

    Improving text-based person search by spatial matching and adaptive threshold,

    T. Chen, C. Xu, and J. Luo, “Improving text-based person search by spatial matching and adaptive threshold,” inWACV. IEEE, 2018, pp. 1879–1887

  16. [24]

    Language person search with mutually connected classification loss,

    Y. Wang, C. Bo, D. Wang, S. Wang, Y. Qi, and H. Lu, “Language person search with mutually connected classification loss,” in ICASSP. IEEE, 2019, pp. 2057–2061

  17. [25]

    Pose-guided multi-granularity attention network for text-based person search,

    Y. Jing, C. Si, J. Wang, W. Wang, L. Wang, and T. Tan, “Pose-guided multi-granularity attention network for text-based person search,” inAAAI, vol. 34, no. 07, 2020, pp. 11 189–11 196

  18. [26]

    Vitaa: Visual-textual attributes alignment in person search by natural language,

    Z. Wang, Z. Fang, J. Wang, and Y. Yang, “Vitaa: Visual-textual attributes alignment in person search by natural language,” in ECCV. Springer, 2020, pp. 402–420

  19. [27]

    Improving description-based person re-identification by multi-granularity image-text alignments,

    K. Niu, Y. Huang, W. Ouyang, and L. Wang, “Improving description-based person re-identification by multi-granularity image-text alignments,”IEEE Transactions on Image Processing, vol. 29, pp. 5542–5556, 2020

  20. [28]

    Caibc: Capturing all-round information beyond color for text- based person retrieval,

    Z. Wang, A. Zhu, J. Xue, X. Wan, C. Liu, T. Wang, and Y. Li, “Caibc: Capturing all-round information beyond color for text- based person retrieval,” inACMMM, 2022, pp. 5314–5322

  21. [29]

    Lapscore: language-guided person search via color reasoning,

    Y. Wu, Z. Yan, X. Han, G. Li, C. Zou, and S. Cui, “Lapscore: language-guided person search via color reasoning,” inICCV, 2021, pp. 1624–1633

  22. [30]

    An empirical study of clip for text-based person search,

    M. Cao, Y. Bai, Z. Zeng, M. Ye, and M. Zhang, “An empirical study of clip for text-based person search,” inAAAI, vol. 38, no. 1, 2024, pp. 465–473

  23. [31]

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

    S. Yan, N. Dong, L. Zhang, and J. Tang, “Clip-driven fine-grained text-image person re-identification,”IEEE Transactions on Image Processing, vol. 32, pp. 6032–6046, 2023

  24. [32]

    P-clip: Pro- gressive discrepancy learning for one-shot text-to-image person re-identification,

    C. Wang, M. Dong, M. Ye, H. Sun, and X. Jiang, “P-clip: Pro- gressive discrepancy learning for one-shot text-to-image person re-identification,”IEEE Transactions on Image Processing, 2026

  25. [33]

    I2id: Disentangling identity features via synchronized masking for zero-shot composed person retrieval,

    G. Yu, D. Wang, C. Yan, F. Yan, N. Luo, Y. Wang, and Q. Wang, “I2id: Disentangling identity features via synchronized masking for zero-shot composed person retrieval,”Pattern Recognition, p. 113654, 2026

  26. [34]

    Learning transfer- able visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P . Mishkin, J. Clarket al., “Learning transfer- able visual models from natural language supervision,” inICML. PmLR, 2021, pp. 8748–8763

  27. [35]

    Fine-grained semantics-aware representation learning for text-based person retrieval,

    D. Wang, F. Yan, Y. Wang, L. Zhao, X. Liang, H. Zhong, and R. Zhang, “Fine-grained semantics-aware representation learning for text-based person retrieval,” inICMR, 2024, p. 92–100

  28. [36]

    Clip-based synergistic knowledge transfer for text-based person retrieval,

    Y. Liu, Y. Li, Z. Liu, W. Yang, Y. Wang, and Q. Liao, “Clip-based synergistic knowledge transfer for text-based person retrieval,” in ICASSP. IEEE, 2024, pp. 7935–7939

  29. [37]

    Enhancing visual representation for text-based person searching,

    W. Shen, M. Fang, Y. Wang, J. Xiao, D. Li, H. Chen, L. Xu, and W. Zhang, “Enhancing visual representation for text-based person searching,”Knowledge-Based Systems, vol. 309, p. 112893, 2025

  30. [38]

    Vse++: Improving visual-semantic embeddings with hard negatives,

    F. Faghri, D. J. Fleet, J. R. Kiros, and S. Fidler, “Vse++: Improving visual-semantic embeddings with hard negatives,” 2018

  31. [39]

    Align before fuse: Vision and language representation learning with momentum distillation,

    J. Li, R. Selvaraju, A. Gotmare, S. Joty, C. Xiong, and S. C. H. Hoi, “Align before fuse: Vision and language representation learning with momentum distillation,”NeurIPS, vol. 34, pp. 9694–9705, 2021

  32. [40]

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

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language- image pre-training for unified vision-language understanding and generation,” inICML. PMLR, 2022, pp. 12 888–12 900

  33. [41]

    FILIP: Fine-grained interactive language- image pre-training,

    L. Yao, R. Huang, L. Hou, G. Lu, M. Niu, H. Xu, X. Liang, Z. Li, X. Jiang, and C. Xu, “FILIP: Fine-grained interactive language- image pre-training,” inICML, 2022

  34. [42]

    Long-clip: Unlocking the long-text capability of clip,

    B. Zhang, P . Zhang, X. Dong, Y. Zang, and J. Wang, “Long-clip: Unlocking the long-text capability of clip,” inECCV. Springer, 2024, pp. 310–325

  35. [43]

    Imram: Iterative matching with recurrent attention memory for cross- modal image-text retrieval,

    H. Chen, G. Ding, X. Liu, Z. Lin, J. Liu, and J. Han, “Imram: Iterative matching with recurrent attention memory for cross- modal image-text retrieval,” inCVPR, 2020, pp. 12 655–12 663

  36. [44]

    Fine-grained image-text matching by cross-modal hard aligning network,

    Z. Pan, F. Wu, and B. Zhang, “Fine-grained image-text matching by cross-modal hard aligning network,” inCVPR, 2023, pp. 19 275– 19 284

  37. [45]

    Flair: Vlm with fine-grained language-informed image representations,

    R. Xiao, S. Kim, M.-I. Georgescu, Z. Akata, and S. Alaniz, “Flair: Vlm with fine-grained language-informed image representations,” inCVPR, 2025, pp. 24 884–24 894

  38. [46]

    Benchmark granularity and model robustness for image-text retrieval: A reproducibility study,

    M. Hendriksen, S. Zhang, R. Reinanda, M. Yahya, E. Meij, and M. de Rijke, “Benchmark granularity and model robustness for image-text retrieval: A reproducibility study,” inProceedings of the 48th International ACM SIGIR Conference on Research and Develop- ment in Information R...

  39. [47]

    Multi-grained vision language pre- training: Aligning texts with visual concepts,

    Y. Zeng, X. Zhang, and H. Li, “Multi-grained vision language pre- training: Aligning texts with visual concepts,” inICML. PMLR, 2022, pp. 25 994–26 009

  40. [48]

    Pixclip: Achieving fine-grained visual language understanding via any-granularity pixel-text alignment learning,

    Y. Xiao, Y. Chen, H. Ma, J. Hong, C. Li, L. Wu, H. Guo, and J. Wang, “Pixclip: Achieving fine-grained visual language understanding via any-granularity pixel-text alignment learning,”arXiv preprint arXiv:2511.04601, 2025

  41. [49]

    Mulclip: A multi-level alignment framework for enhancing fine-grained long-context clip,

    C. Truong, H. T. Quang, and D. D. Le, “Mulclip: A multi-level alignment framework for enhancing fine-grained long-context clip,”arXiv preprint arXiv:2512.07128, 2025

  42. [50]

    X-clip: End-to- end multi-grained contrastive learning for video-text retrieval,

    Y. Ma, G. Xu, X. Sun, M. Yan, J. Zhang, and R. Ji, “X-clip: End-to- end multi-grained contrastive learning for video-text retrieval,” in ACMMM, 2022, pp. 638–647

  43. [51]

    Mgsgm: Multi-granularity selective graph mamba for image-text retrieval,

    Y. Huang, Y. Bu, K. Guo, Z. Liu, X. Song, and S. Sun, “Mgsgm: Multi-granularity selective graph mamba for image-text retrieval,” inProceedings of the 2025 International Conference on Multimedia Retrieval, 2025, pp. 1983–1987

  44. [52]

    Fairmot: On the fairness of detection and re-identification in multiple object tracking,

    Y. Zhang, C. Wang, X. Wang, W. Zeng, and W. Liu, “Fairmot: On the fairness of detection and re-identification in multiple object tracking,”International journal of computer vision, vol. 129, no. 11, pp. 3069–3087, 2021

  45. [53]

    Rgb-infrared cross-modality person re-identification,

    A. Wu, W.-S. Zheng, H.-X. Yu, S. Gong, and J. Lai, “Rgb-infrared cross-modality person re-identification,” inICCV, 2017, pp. 5380– 5389

  46. [54]

    Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared per- son re-identification,

    Y. Zhang and H. Wang, “Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared per- son re-identification,” inCVPR, 2023, pp. 2153–2162

  47. [55]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” NeurIPS, vol. 30, 2017

  48. [56]

    Har- nessing the power of mllms for transferable text-to-image person reid,

    W. Tan, C. Ding, J. Jiang, F. Wang, Y. Zhan, and D. Tao, “Har- nessing the power of mllms for transferable text-to-image person reid,” inCVPR, 2024, pp. 17 127–17 137

  49. [57]

    Modeling thousands of human annotators for generalizable text-to-image person re-identification,

    J. Jiang, C. Ding, W. Tan, J. Wang, J. Tao, and X. Xu, “Modeling thousands of human annotators for generalizable text-to-image person re-identification,” inCVPR, 2025, pp. 9220–9230

  50. [58]

    Learn- ing granularity-unified representations for text-to-image person re-identification,

    Z. Shao, X. Zhang, M. Fang, Z. Lin, J. Wang, and C. Ding, “Learn- ing granularity-unified representations for text-to-image person re-identification,” inACMMM, 2022, pp. 5566–5574

  51. [59]

    Bilma: Bidirectional local-matching for text-based person re-identification,

    T. Fujii and S. Tarashima, “Bilma: Bidirectional local-matching for text-based person re-identification,” inICCV, 2023, pp. 2786–2790. 18

  52. [60]

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

    Y. Bai, M. Cao, D. Gao, Z. Cao, C. Chen, Z. Fan, L. Nie, and M. Zhang, “Rasa: relation and sensitivity aware representation learning for text-based person search,” inIJCAI, 2023, pp. 555–563

  53. [61]

    Hier- archical knowledge-guided reasoning for text-based person re- identification,

    R. Zeng, W. Ma, T. Zhou, S. Zhao, X. Mao, and J. Liu, “Hier- archical knowledge-guided reasoning for text-based person re- identification,”Neural Networks, p. 107888, 2025

  54. [62]

    Instance-level feature bias calibration learning for text-to-image person re- identification,

    Y. Gou, Z. Li, J. Zhang, Y. Wang, and Y. Ge, “Instance-level feature bias calibration learning for text-to-image person re- identification,”Knowledge-Based Systems, vol. 315, p. 113251, 2025. Jialong Zuoreceived B.S. degree in Huazhong University of Science and Technology, Chi...

  55. [2020]

    degree in the School of Artificial Intelligence and Au- tomation, Huazhong University of Science and Technology, supervised by Prof

    He is currently pursuing the Ph.D. degree in the School of Artificial Intelligence and Au- tomation, Huazhong University of Science and Technology, supervised by Prof. Changxin Gao. His research interests include dataset pruning, network pruning, training acceleration and mult...

  56. [2023]

    degree in the School of Artificial Intelligence and Au- tomation, Huazhong University of Science and Technology, supervised by Prof

    He is currently pursuing the Ph.D. degree in the School of Artificial Intelligence and Au- tomation, Huazhong University of Science and Technology, supervised by Prof. Changxin Gao. His research interest is person re-identification and text-based person retrieval. Dongyue Wure...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.