Pith. sign in

REVIEW 3 major objections 6 minor 33 references

Graph-Based Cross-Domain Knowledge Distillation for Cross-Dataset Text-to-Image Person Retrieval

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read The paper proposes GCKD, an unsupervised framework that jointly runs graph-based multi-domain propagation and contrastive momentum knowledge distillation, and reports state-of-the-art cross-dataset text-to-image person retrieval on three…

desk verdict Plausible UDA method for cross-dataset text-to-image person retrieval, but the central KD equations are internally inconsistent and need correction before the main claim is credible. read the letter →

arxiv 2501.15052 v1 pith:UQAYKYGP submitted 2025-01-25 cs.CV cs.AIcs.MM

classification cs.CVcs.AIcs.MM
keywords text-to-imagepersonretrievalunsuperviseddomainadaptationknowledgedistillationgraphneuralnetworkcross-modalmomentumvision-languagepre-training
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

Cross-dataset text-to-image person retrieval means finding a person in an image gallery from a text description when the target dataset has no image-text identity annotations. The paper proposes an unsupervised domain adaptation method, GCKD, that adapts a vision-language model using only unlabeled target images and texts together with a labeled source dataset. The method combines a graph module that propagates features across source and target domains with a momentum teacher that produces pseudo image-text similarity targets for contrastive and matching losses. On three public datasets the paper reports consistent gains over state-of-the-art transfer baselines, with an average Rank-1 improvement of 4.37 percentage points over the second-best method. If these results hold, labeling a new deployment dataset is unnecessary for adapting text-to-image person retrieval models.

What carries the argument

The machinery is a pair of interacting modules. The Graph-based Multi-domain Propagation (GMP) module stores recent image and text embeddings from source and target domains in four memory banks, builds a KNN graph over the current batch plus those memories, and applies a two-layer graph convolution so target features are refined by neighboring source and target samples. The Contrastive Momentum Knowledge Distillation (CMKD) module maintains an EMA teacher initialized from the source-pretrained ALBEF vision-language backbone; the teacher computes pseudo similarity targets and high-confidence positive pairs used in a cross-domain image-text contrastive loss and an image-text matching loss, while the student is trained with those targets. A masked-language-modeling loss from the backbone completes the objective.

What would settle it

Run the same training pipeline on ICFG-PEDES to RSTPReid but replace the teacher's pseudo similarity targets with uniform random weights or with the student's own similarities; if Rank-1 stays roughly the same, the momentum distillation is not the cause of the reported gain. A second check is to compute the precision of the teacher's high-confidence positive pairs, those above the threshold $\delta$, against known identity matches on a held-out labeled subset of the target data, since near-chance precision would mean the cross-domain matching loss is trained on noisy pseudo-labels.

Watch

Extended reading notes

Core claim

The central claim is that the two challenges of cross-dataset text-to-image person retrieval—domain shift between datasets and the modality gap between images and text—can be handled together by one framework in the unsupervised setting. The paper's GCKD builds a dynamic K-nearest-neighbor graph whose vertices are the current batch plus memory banks of source and target image and text embeddings, then runs a two-layer graph network to propagate information across domains. In parallel, a student model is trained with a teacher model updated by exponential moving average; the teacher supplies pseudo similarity targets and high-confidence positive pairs that drive cross-modal contrastive and fine-grained matching losses. The paper reports that the full method outperforms all compared baselines on ICFG-PEDES, RSTPReid, and CUHK-PEDES transfer tasks, and that each module contributes in the ablation.

Load-bearing premise

The method depends on the teacher model's automatically generated image-text similarity scores on unlabeled target data being accurate enough to supervise the student, and the paper does not directly measure that accuracy.

Editorial extensions

If this is right

  • A new target dataset needs only unpaired images and texts, not identity labels, to adapt a text-to-image person retrieval model.
  • The ablation indicates that the contrastive momentum distillation module is responsible for most of the improvement, with the graph propagation adding a smaller further margin.
  • The reported results set a new comparison point for cross-dataset text-to-image person retrieval, where the paper shows existing UDA baselines often trail single-domain retrieval baselines in the transfer setting.

Reading between the lines

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

  • A direct test of the distillation mechanism would be to replace the teacher's pseudo similarity targets with the student's own soft similarities; if the gains vanish, the teacher is doing the work, and if they do not, the gain is largely contrastive learning on target data.
  • The pseudo-label threshold $\delta$ and memory size $C$ are not reported, so their sensitivity is unknown; reporting them would let practitioners calibrate the precision-recall trade-off on new datasets.
  • The graph propagation module is not specific to person images or text, so it could be transplanted to cross-modal retrieval in other domains, such as video-text or sketch-image search.
  • For very small target galleries, the memory banks and KNN graph may not contain enough neighboring structure, so the method likely needs adaptation for few-shot target domains.
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 / 6 minor

Summary. The paper proposes GCKD, an unsupervised domain adaptation method for cross-dataset text-to-image person retrieval. It combines a graph-based multi-domain propagation (GMP) module, which propagates features over a dynamic KNN graph connecting source and target memory banks, with a contrastive momentum knowledge distillation (CMKD) module, which uses an EMA-updated teacher to produce pseudo similarity targets on unpaired target data. The authors evaluate on ICFG-PEDES, RSTPReid, and CUHK-PEDES, including both cross-dataset and intra-dataset transfer settings, and report consistent improvements over source-only and UDA baselines.

Significance. If the reported results are reproducible, the paper would provide a strong unsupervised domain adaptation baseline for text-to-image person retrieval and one of the first applications of vision-language pretraining in this cross-dataset setting. The experimental work is substantial: three datasets, multiple transfer directions, stepwise ablations, and comparisons to a reasonable set of existing methods. However, the central CMKD equations are ambiguous about which network produces the pseudo-targets, and several key hyperparameters are omitted. Because the ablation in Table 4 attributes all improvement over the baseline to CMKD, resolving this ambiguity is essential before the claimed knowledge-distillation mechanism can be accepted. The paper also ships no code and omits reproducibility details such as the pseudo-label threshold and memory-bank size.

major comments (3)
  1. [§3.3, Eqs. (7)–(8)] The text states that "the student model is used to generate the pseudo target domain image features \hat f_{TI} and pseudo target domain text features \hat f_{TT}", but Eqs. (7)–(8) are described in the following sentence as "pseudo similarity targets generated by the teacher model". If the \hat f features are student outputs, then the soft targets in Eq. (7) are computed from the student's own embeddings, making Eq. (6) a self-training or self-distillation objective rather than knowledge distillation; the teacher would play no role in this loss, and the ablation improvement in Table 4 could not be attributed to momentum distillation. If the \hat f features are teacher outputs, then the main text is incorrect and the required teacher forward pass on target samples is nowhere described, nor is it stated whether gradients are stopped at \hat f. Please clarify which network computes \hat f_{TI} and \hat f_{TT}, specify the relationship between the \hat f and f variables, and state the gradient flow explicitly.
  2. [§4, Implementation Details; Eq. (9)] The implementation details report L, K, τ, α, λ1, λ2, λ3, batch size, and learning rate, but omit the pseudo-label threshold δ used in Eq. (9) and the memory bank size C used in graph construction (Eq. (2)) and in Eqs. (6)–(8). Since δ controls the number of positive pairs entering L_cd−itm and C defines the graph vertex set and memory capacity, these values are necessary to reproduce the reported results. The paper also provides no analysis of the accuracy of the high-confidence pseudo labels (for example, precision at the chosen threshold), which is important because the teacher is initialized from source-pretrained ALBEF and applied to an unpaired target domain.
  3. [Table 4 vs. Table 1; §4 Ablation Study] The ablation baseline in Table 4 reports exactly the same numbers as the RaSa row in Table 1 for both transfer tasks (55.00/73.65/81.55/46.18 for ICFG-PEDES→RSTPReid and 48.65/69.90/76.53/42.03 for ICFG-PEDES→CUHK-PEDES). However, the text describes the baseline as "the backbone only", while RaSa is a published method with additional relation- and sensitivity-aware components. If the baseline is in fact RaSa, then the statement that adding CMKD improves the baseline is relative to a strong published method, not to a plain backbone; if it is a plain ALBEF backbone, the exact numerical coincidence requires an explanation. Please clarify the baseline training objective and reconcile this inconsistency, as it directly affects the ablation attribution in Table 4.
minor comments (6)
  1. [Abstract] The abstract contains "acExtensive experiments"; this appears to be a typo and should read "Extensive experiments".
  2. [§3.2, Graph Construction; Eq. (2)] The symbol K is used both for the number of graph neighbors in Eq. (3) and for the feature dimension in F_m ∈ R^{B×K} and X ∈ R^{(B+2C)×K}; please use a different symbol, such as d, for the feature dimension.
  3. [Eq. (6)] The sums over f_TI and f_TT are not indexed; please specify that the loss is summed over the target-domain image and text samples in the current batch.
  4. [Eq. (5) and §4 Implementation Details] Equation (5) uses m for the momentum coefficient, while the implementation details report α = 0.999; please unify the notation.
  5. [Table 3] The evaluation section states that Rank-K (K = 1, 5, 10) and mAP are used, but Table 3 reports only Rank-1 and Rank-5 for the intra-dataset experiments; please clarify why Rank-10 and mAP are omitted.
  6. [Eq. (3)] The adjacency matrix A is defined asymmetrically (A_ij = 1 if v_i is one of the K nearest neighbors of v_j); if the GNN assumes an undirected graph, please define the symmetrization explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

CMKD's pseudo-target equations are self-referential: the student is said to generate f_hat features that Eqs. (7)-(8) then call teacher-generated targets, so the core distillation signal reduces to the student's own output as written.

  1. self definitional [Section 'Contrastive Momentum Knowledge Distillation' -> 'Cross-modal Contrastive Learning', Eqs. (6)-(8)]
    "Specifically, given a batch of paired source domain samples (vs, ts), and unpaired target domain visual and textual samples vt and tt, the student model is used to generate the pseudo target domain image features \hat f_{TI} and pseudo target domain text features \hat f_{TT}. ... are the pseudo similarity targets generated by the teacher model"

    Equation (6) minimizes the contrastive loss against soft targets si2t and st2i defined in Eqs. (7)-(8). The sentence introducing those equations states that the student model generates \hat f_{TI} and \hat f_{TT}; the sentence after Eqs. (7)-(8) calls the same quantities 'pseudo similarity targets generated by the teacher model.' If the student supplies \hat f, the soft targets are the student's own cosine similarities over the same target memory that appears in the denominator of Eq. (6), so the 'teacher' signal reduces to the student's own output and the distillation is self-referential. If the teacher supplies \hat f, no teacher forward pass on target samples is specified and Eq. (5) alone does not define the targets.

full rationale

The load-bearing circularity is in the CMKD module. The pseudo-target definitions in Eqs. (7)-(8) are explicitly said to be generated by the student model in the same paragraph that calls them teacher-generated, so the contrastive distillation target in Eq. (6) reduces to the student's own similarities as written. Because Table 4 attributes the primary performance gain to CMKD, the central claim is affected by this ambiguity. I did not find circularity in the GMP module or in the benchmark comparisons: the source-only and UDA baselines are external methods, and the graph construction does not presuppose the retrieval result. The remainder of the derivation is self-contained against external benchmarks, so the circularity is partial rather than total.

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

All free parameters are hand-chosen hyperparameters that directly affect the reported results; delta and C are not even reported, which is a reproducibility gap. The axioms are domain assumptions about transferability of VLP features, shared attribute vocabulary, EMA teacher quality, and the correspondence assumption between unpaired target images and texts. No new physical or conceptual entities are introduced beyond the algorithmic graph and memory modules.

free parameters (9)
  • EMA momentum coefficient m = 0.999
    Sets teacher lag; chosen by hand, no sensitivity analysis.
  • temperature tau = 0.07
    Scales contrastive logits; standard ALBEF value, not tuned here.
  • KNN neighbor count K = 10
    Defines graph adjacency; chosen by hand.
  • GNN layers L = 2
    Depth of propagation; chosen by hand.
  • loss weights lambda_1, lambda_2, lambda_3 = 0.5, 0.5, 1
    Balance contrastive, matching, and MLM losses; no sensitivity study.
  • batch size = 4
    Affects contrastive negatives and graph construction; chosen by compute constraints.
  • learning rate = 1e-5
    AdamW initial learning rate; chosen by hand.
  • pseudo-label threshold delta
    Used to accept positive pairs in cross-domain image-text matching; value not reported.
  • memory bank size C
    Number of source and target embeddings stored per modality; never specified.
assumptions (5)
  • domain assumption ALBEF source-pretrained weights provide transferable visual and textual representations.
    The whole method initializes student and teacher from ALBEF pretrained on the source dataset and uses these features for graph construction, memory banks, and pseudo labels.
  • domain assumption Source and target person descriptions share a common attribute vocabulary.
    Cross-domain pseudo labels and hard negatives are meaningful only if clothing, color, and gender semantics transfer across datasets.
  • ad hoc to paper The EMA teacher retains source knowledge and provides better pseudo labels than the student.
    No empirical analysis supports this; it is the core premise of momentum distillation in this paper.
  • domain assumption Unpaired target images and texts correspond to overlapping identities in expectation.
    The contrastive loss weights every target image-text pair as a potential match; if target images and texts describe disjoint people, the objective is ill-posed.
  • domain assumption KNN graph and two-layer GNN propagation on cosine similarities align cross-domain features.
    The GMP module assumes local graph structure in embedding space is meaningful across domains despite the domain shift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph-Based Cross-Domain Knowledge Distillation for Cross-Dataset Text-to-Image Person Retrieval." pith.science (2026). https://pith.science/paper/UQAYKYGP

@misc{pith2026250115052,
  author       = {Pith},
  title        = {Pith review of: Graph-Based Cross-Domain Knowledge Distillation for Cross-Dataset Text-to-Image Person Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQAYKYGP}},
  note         = {Machine review of arXiv:2501.15052}
}
read the original abstract

Video surveillance systems are crucial components for ensuring public safety and management in smart city. As a fundamental task in video surveillance, text-to-image person retrieval aims to retrieve the target person from an image gallery that best matches the given text description. Most existing text-to-image person retrieval methods are trained in a supervised manner that requires sufficient labeled data in the target domain. However, it is common in practice that only unlabeled data is available in the target domain due to the difficulty and cost of data annotation, which limits the generalization of existing methods in practical application scenarios. To address this issue, we propose a novel unsupervised domain adaptation method, termed Graph-Based Cross-Domain Knowledge Distillation (GCKD), to learn the cross-modal feature representation for text-to-image person retrieval in a cross-dataset scenario. The proposed GCKD method consists of two main components. Firstly, a graph-based multi-modal propagation module is designed to bridge the cross-domain correlation among the visual and textual samples. Secondly, a contrastive momentum knowledge distillation module is proposed to learn the cross-modal feature representation using the online knowledge distillation strategy. By jointly optimizing the two modules, the proposed method is able to achieve efficient performance for cross-dataset text-to-image person retrieval. acExtensive experiments on three publicly available text-to-image person retrieval datasets demonstrate the effectiveness of the proposed GCKD method, which consistently outperforms the state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2501.15052 by the authors.

Figure 1
Figure 1. An illustration of cross-dataset text-to-image per [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The main framework of the proposed Graph-based Cross-domain Knowledge Distillation (GCKD) method. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 21 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bai, Y.; Cao, M.; Gao, D.; Cao, Z.; Chen, C.; Fan, Z.; Nie, L.; and Zhang, M. 2023. RaSa: relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, 555--563

  4. [4]

    Chen, Y.; Huang, R.; Chang, H.; Tan, C.; Xue, T.; and Ma, B. 2021. Cross-modal knowledge adaptation for language-based person search. IEEE Transactions on Image Processing, 30: 4057--4069

  5. [5]

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

  6. [6]

    Fan, Z.; Cong, W.; Wen, K.; Wang, K.; Zhang, J.; Ding, X.; Xu, D.; Ivanovic, B.; Pavone, M.; Pavlakos, G.; Wang, Z.; and Wang, Y. 2024. InstantSplat: Unbounded Sparse-view Pose-free Gaussian Splatting in 40 Seconds. arXiv:2403.20309

  7. [7]

    Gu, X.; Fan, Z.; Zhu, S.; Dai, Z.; Tan, F.; and Tan, P. 2020. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2495--2504

  8. [8]

    Hao, X.; Zhang, W.; Wu, D.; Zhu, F.; and Li, B. 2023. Dual alignment unsupervised domain adaptation for video-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 18962--18972

Show all 33 references
  1. [9]

    He, S.; Luo, H.; Wang, P.; Wang, F.; Li, H.; and Jiang, W. 2021. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, 15013--15022

  2. [10]

    Hu, X.; Zhang, K.; Xia, L.; Chen, A.; Luo, J.; Sun, Y.; Wang, K.; Qiao, N.; Zeng, X.; Sun, M.; et al. 2024. ReCLIP: Refine contrastive language image pre-training with source free domain adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer V...

  3. [11]

    Jiang, D.; and Ye, M. 2023. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2787--2797

  4. [12]

    Jing, Y.; Wang, W.; Wang, L.; and Tan, T. 2020. Cross-modal cross-domain moment alignment network for person search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10678--10686

  5. [13]

    Lee, K.-H.; Chen, X.; Hua, G.; Hu, H.; and He, X. 2018. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), 201--216

  6. [14]

    Li, J.; Selvaraju, R.; Gotmare, A.; Joty, S.; Xiong, C.; and Hoi, S. C. H. 2021. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34: 9694--9705

  7. [15]

    Li, S.; Cao, M.; and Zhang, M. 2022. Learning semantic-aligned feature representation for text-based person search. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2724--2728. IEEE

  8. [16]

    Li, S.; Xiao, T.; Li, H.; Zhou, B.; Yue, D.; and Wang, X. 2017. Person search with natural language description. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1970--1979

  9. [17]

    Li, S.; Xu, X.; Shen, F.; and Yang, Y. 2023. Multi-granularity Separation Network for Text-Based Person Retrieval with Bidirectional Refinement Regularization. In Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, 307--315

  10. [18]

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

  11. [19]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PMLR

  12. [20]

    G.; and Gao, J

    Shao, Z.; Xi, H.; Lu, H.; Wang, Z.; Bell, M. G.; and Gao, J. 2024. STLLM-DF: A Spatial-Temporal Large Language Model with Diffusion for Enhanced Multi-Mode Traffic System Forecasting. arXiv preprint arXiv:2409.05921

  13. [21]

    Shao, Z.; Zhang, X.; Fang, M.; Lin, Z.; Wang, J.; and Ding, C. 2022. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th acm international conference on multimedia, 5566--5574

  14. [22]

    Shu, X.; Wen, W.; Wu, H.; Chen, K.; Song, Y.; Qiao, R.; Ren, B.; and Wang, X. 2022. See finer, see more: Implicit modality alignment for text-based person retrieval. In European Conference on Computer Vision, 624--641. Springer

  15. [23]

    Tanwisuth, K.; Zhang, S.; Zheng, H.; He, P.; and Zhou, M. 2023. POUF: Prompt-oriented unsupervised fine-tuning for large pre-trained models. In International Conference on Machine Learning, 33816--33832. PMLR

  16. [24]

    Tzeng, E.; Hoffman, J.; Saenko, K.; and Darrell, T. 2017. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 7167--7176

  17. [25]

    D.; Hensher, D

    Xi, H.; Nelson, J. D.; Hensher, D. A.; Hu, S.; Shao, X.; and Xie, C. 2024 a . Evaluating travel behavior resilience across urban and rural areas during the COVID-19 pandemic: contributions of vaccination and epidemiological indicators. Transportation research part A: policy an...

  18. [26]

    Xi, H.; Wang, Y.; Shao, Z.; Zhang, X.; and Waller, T. 2024 b . Optimizing mobility resource allocation in multiple MaaS subscription frameworks: a group method of data handling-driven self-adaptive harmony search algorithm. Annals of Operations Research, 1--29

  19. [27]

    Yan, S.; Dong, N.; Zhang, L.; and Tang, J. 2023. Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing

  20. [28]

    Yang, S.; Zhou, Y.; Zheng, Z.; Wang, Y.; Zhu, L.; and Wu, Y. 2023. 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

  21. [29]

    Zhang, Y.; and Lu, H. 2018. Deep cross-modal projection learning for image-text matching. In Proceedings of the European conference on computer vision (ECCV), 686--701

  22. [30]

    Zheng, K.; Liu, W.; Liu, J.; Zha, Z.-J.; and Mei, T. 2020. Hierarchical gumbel attention network for text-based person search. In Proceedings of the 28th ACM International Conference on Multimedia, 3441--3449

  23. [31]

    Zhong, Z.; Zheng, L.; Luo, Z.; Li, S.; and Yang, Y. 2019. Invariance matters: Exemplar memory for domain adaptive person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 598--607

  24. [32]

    Zhu, A.; Wang, Z.; Li, Y.; Wan, X.; Jin, J.; Wang, T.; Hu, F.; and Hua, G. 2021. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, 209--217

  25. [33]

    Zhu, A.; Wang, Z.; Xue, J.; Wan, X.; Jin, J.; Wang, T.; and Snoussi, H. 2024. Improving Text-Based Person Retrieval by Excavating All-Round Information Beyond Color. IEEE Transactions on Neural Networks and Learning Systems

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.