Pith. sign in

REVIEW 3 major objections 5 minor 64 references

DisenQ: Disentangling Q-Former for Activity-Biometrics

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DisenQ claims one Q-Former with three learnable query sets aligned to structured text can separate identity from appearance and motion, achieving state-of-the-art activity-biometrics without extra visual data.

desk verdict A solid three-query Q-Former with an honest supplement, but the paper's own causal intervention shows the text semantics are barely load-bearing, so the 'language-guided disentanglement' framing overstates the mechanism. read the letter →

arxiv 2507.07262 v1 pith:42GZ62FR submitted 2025-07-09 cs.CV

Shehreen Azad , Yogesh S Rawat This is my paper · ORCID
classification cs.CV
keywords activity-biometricspersonre-identificationfeaturedisentanglementQ-Formerlanguage-guidedsupervisionvision-languagemodelvideoidentification
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 argues that activity-biometrics—identifying a person from arbitrary daily activities—becomes tractable when the model deliberately separates three kinds of video information: biometric traits such as body shape and posture, motion patterns of the action being performed, and appearance details like clothing and accessories. Its proposed DisenQ architecture is a Q-Former with three independent sets of learnable queries, each aligning the same video features with a structured text description of one of those three aspects, so that identity matching can rely on biometrics and motion while ignoring appearance. This matters because existing activity-aware identification leans on silhouettes or pose, whose extraction errors limit real-world reliability, and because identity cues are otherwise entangled with motion and appearance, causing misidentifications across activities. The paper reports state-of-the-art rank-1 accuracy and mAP on NTU RGB-AB, PKU MMD-AB, and Charades-AB, along with competitive results on the walking-dominated MEVID benchmark, using text supervision only at training time.

What carries the argument

The load-bearing object is DisenQ, a modified Q-Former (the lightweight querying transformer introduced for BLIP-2) in which the single learnable query set is replaced by three independent query sets—biometrics $z_b$, motion $z_m$, and non-biometrics $\hat{z}_b$—all sharing one self-attention and cross-attention block. Each query set first refines itself in self-attention, then cross-attends with query, key, and value built from the global video feature $F$ concatenated with the matching structured text embedding ($T_b$, $T_m$, or $\hat{T}_b$), per Eqs. (1)–(3); the three streams therefore read the same video through different text-modulated keys and values and never exchange tokens. The text supervision itself comes from a frozen vision-language model that fills a fixed template describing physique and posture (biometrics), the action label and movement (motion), and clothing and accessories (non-biometrics), with biometrics descriptions averaged per identity for stability, and the whole framework is trained with an identity cross-entropy loss, a triplet loss, a motion cross-entropy loss, and an orthogonality term $L_{\mathrm{Orth}} = \left|F_b^T \hat{F}_b\right|$ on the pooled biometrics and non-biometrics vectors.

What would settle it

One decisive observation is the paper's own cross-feature leakage measurement (Supplement C.3): a linear classifier predicts the action from the biometrics stream at 14.4% accuracy on NTU RGB-AB, well above the 4.8% it reaches from non-biometrics. A reader could settle the independence claim by completing this audit—training probes to predict clothing from the motion stream $F_m$ and identity from the non-biometrics stream $\hat{F}_b$ on the same splits; if clothing is predictable from $F_m$ at well above chance, the reported motion benefit is appearance leakage rather than genuine motion disentanglement.

Watch

Extended reading notes

Core claim

The central discovery claimed is that explicit feature disentanglement is achievable with language as the only auxiliary signal: a single Q-Former transformer, given three separate learnable query sets $z_b$ (biometrics), $z_m$ (motion), and $\hat{z}_b$ (non-biometrics) that share the same self-attention and cross-attention layers, learns to route the same visual features into streams that are identity-specific, motion-specific, and appearance-specific, respectively. Each query stream cross-attends to the concatenation of the video-level feature $F$ with its corresponding frozen text embedding ($T_b$, $T_m$, $\hat{T}_b$) as defined in Eqs. (1)–(3). Retrieval uses only the pooled biometrics and motion vectors, combined by an adaptive MLP weight (Eq. 8), and an orthogonality penalty between the pooled biometrics and non-biometrics vectors (Eq. 6) reinforces separation; the motion stream is supervised by an action cross-entropy loss. The paper asserts this design yields state-of-the-art results on three activity-biometrics benchmarks (rank-1 of 82.2/80.9 on NTU RGB-AB, 89.2/84.1 on PKU MMD-AB, 49.9/48.4 on Charades-AB for same/cross activity) and competitive generalization to MEVID.

Load-bearing premise

The argument hinges on the assumption that three query sets sharing the same attention layers—with an orthogonality penalty applied only between the pooled biometrics and non-biometrics vectors, and none at all on the motion stream—are enough to keep the three information streams genuinely independent; the paper's own leakage test shows that action identity is still readable from biometrics features at 14.4% accuracy.

Editorial extensions

If this is right

  • Activity-biometrics identification can be improved without silhouettes or pose estimates, since the only auxiliary supervision is text generated once per identity at training time and discarded at inference.
  • Identity matching becomes largely clothing-invariant: the non-biometrics stream carries near-random identity and action information and is excluded from the similarity computation.
  • The adaptive weighting between biometrics and motion similarity stabilizes identification across activities, preventing highly distinctive or subtle actions from dominating the decision.
  • The framework is robust to the choice of vision-language model, so its gains do not depend on a specific captioner.
  • The same design transfers to a traditional walking-based benchmark, suggesting the disentangling front-end generalizes beyond diverse-activity settings.

Reading between the lines

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

  • The reported 14.4% action-predictability of the biometrics stream (Supplement C.3) and the absence of any orthogonality constraint on the motion stream imply the three-way separation is partial, not complete; a direct test of whether tightening the motion constraint helps or hurts cross-activity identification would pin down whether residual motion information in the biometrics stream is a bug or
  • The causal-intervention result—swapping prompt semantics across branches changes accuracy by only 1–2%—suggests the visual supervision and architecture, rather than the text content, do the main disentangling work; this is consistent with an interpretation in which the three query sets act as learned projectors and the text mainly stabilizes training.
  • The paper acknowledges (Supplement C.4) that demographic bias may persist from the pre-trained vision-language model; an audit of retrieval accuracy across demographic groups would show whether text-guided separation inherits or amplifies that bias.
  • A natural extension would be to replace per-identity averaged prompts with category-level text, turning DisenQ into a reusable disentangling front-end for zero-shot video understanding beyond person identification.
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 / 5 minor

Summary. The paper introduces DisenQ, a query-transformer module with three learnable query sets intended to capture biometrics, motion, and non-biometrics features for activity-based person identification. The queries cross-attend to visual features concatenated with text embeddings generated by a frozen VLM from key-frames, and the model is trained with identity and action cross-entropy, triplet, and orthogonality losses. Evaluation is performed on three activity-biometrics benchmarks (NTU RGB-AB, PKU MMD-AB, Charades-AB) and on MEVID, with reported state-of-the-art or competitive results. The ablation and analysis sections argue that the three-way separation is responsible for the gains, and the paper claims that structured language guidance is the active mechanism enabling this disentanglement.

Significance. If the language-guided disentanglement mechanism were established, this would be a useful contribution: it replaces error-prone silhouettes with text, requires no additional visual data at inference, and is among the first uses of a Q-Former for feature disentanglement. The architectural recipe is simple and the empirical results are consistently above published baselines. However, the paper's own causal intervention (Supp. C.3) shows that the semantic content of the text is almost irrelevant to the outcome, which removes the proposed mechanism from the central claim. The contribution is therefore better described as a three-branch Q-Former with task-specific losses and adaptive motion weighting, and the paper needs substantial re-analysis or re-framing before the stated claims are supported.

major comments (3)
  1. [§3.2; Supplement C.3] Supplement C.3 reports a causal intervention in which the biometrics, motion, and non-biometrics text embeddings are deliberately swapped across branches while all losses are kept unchanged; identity and action accuracy on NTU RGB-AB drop only by 1-2%. This is the authors' own evidence that the structured language descriptions, which the abstract and Section 3.2 identify as the mechanism that 'explicitly disentangles' the feature spaces, are not causally responsible for the separation. The observed separation could be fully explained by the three query streams plus LID, LAct, and LOrth. Because the central claim of the paper is language-guided disentanglement, this result is load-bearing rather than a minor caveat. The authors should either add experiments that isolate the causal role of text semantics (e.g., random versus structured versus no text at equal query count, or a control with fixed dummy prompts), or substantially re-frame the title, abstract, and introduction around the three-branch architecture.
  2. [§3.2, Eq. (6); Supplement C.3] The independence claim is only weakly enforced and only partially measured. The orthogonality loss in Eq. (6) is applied to the pooled biometrics and non-biometrics vectors only; there is no orthogonality or mutual-information penalty between Fb and Fm, and the motion branch is not constrained to be identity-free. The paper's own leakage test shows that action can be predicted from Fb at 14.4% accuracy, and no analogous number is reported for identity leakage from Fm or appearance leakage from Fb. Given the abstract's claim that identity cues remain independent of appearance and motion variations, the authors should report the full leakage matrix (action-from-Fb, identity-from-Fm, clothing-from-Fb, action-from-Fhatb) and either add a constraint on the Fb-Fm pair or weaken the independence claim accordingly.
  3. [Tables 1 and 2; §4.1] All reported numbers come from a single training run, and the margins over the best baselines are small in several cells (e.g., MEVID R@1 60.7 versus 59.5 for ShARc; NTU Cross R@1 80.9 versus 77.0 for ABNet). Without multiple seeds or a significance test, it is not possible to determine whether these gains are real, and the absence of released code or generated prompts makes the text-generation pipeline difficult to reproduce independently. The authors should provide error bars over at least three seeds for the main tables and release the prompt templates and dataset splits used to build the text supervision.
minor comments (5)
  1. [Eq. (4)] Equation (4) uses the symbol '≃' to define LID, which is not a standard way to define a loss; the two cross-entropy terms should be written out explicitly for LID and LAct.
  2. [§3.2] Section 3.2 states that the three query sets share self-attention and cross-attention layers while 'without interaction' attending to different information; these statements sit in tension and should be clarified, since shared parameters do not by themselves prevent interaction through the shared attention weights.
  3. [Section 5] Section 5 contains typos ('asses', 'exisiting'), and Table 4's caption places the word 'Motion' and Figure 5 inside the table; this formatting should be fixed.
  4. [Supplement A; §3.1] The motion prompt template takes the ground-truth action label as input, so Tm is a transform of the label; the paper should state whether this makes the VLM-generated text redundant with the action label in LAct, and whether the same results hold when the action label is omitted from the prompt.
  5. [§4.1] Section 4.1 sets all loss weights λi to 0.01 and gives no sensitivity analysis; a short ablation over λ3 and λ4 would help the reader judge the contribution of LOrth and LAct.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential text-supervision loop: the language guidance is generated from the same key-frames and action labels it is claimed to replace; otherwise the reported benchmark results are empirical and not circular.

  1. self definitional [Section 3.1 'Prompt generation and textual feature extraction'; Supplement A prompt template; Eqs. 1-3]
    "To generate structured and semantically consistent language description, we use a frozen VLM to generate prompts from the key-frame of the input video during only training, without requiring the VLM during inference. ... Analyze the given image where action label is <action label> and extract the following details: Biometrics: A <physique/body shape> person with <posture>... Motion: Performing the action of <action label> by <action description>. Non-biometrics: A <color, type of clothing> and <other accessories>."

    The text embeddings Tb/Tm/Tˆb are not independent supervision: they are generated from the same key-frame that produces the visual feature F, and the motion prompt is templated around the ground-truth action label. In Eqs. 1-3 the cross-attention keys and values are W[F, T], so the 'language guidance' is a re-encoding of the same video content plus the action label. The paper presents this as replacing visual data with textual supervision, but the text is a transformed version of the visual input itself, making the disentanglement loop self-referential. This does not invalidate the empirical rank-1/mAP results, but it weakens the claimed causal role of language guidance.

full rationale

The paper's reported results are benchmark accuracies from a concrete trained model; no predicted quantity is a fitted constant or an equation-level restatement of an input. Self-citations to ABNet [3] are used for dataset splits, evaluation protocols, and as the prior activity-biometrics baseline, not as a load-bearing uniqueness or justification argument, so they do not constitute circularity. The main circular content is the self-supervision loop: structured text is generated by a VLM from the same key-frames and with the ground-truth action label in the prompt, then used as the 'language guidance' in Eqs. 1-3, so the text is not an independent modality. Supplement C.3's causal intervention adds a separate, non-circular validity concern: swapping prompt semantics across branches changes accuracy by only 1-2%, indicating the semantic content of the prompts is not the active driver of the reported separation. That is a causal-attribution weakness, not a circularity, and is better weighed as correctness risk. Overall, the core empirical derivation is self-contained against external benchmarks, so the circularity score is low.

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

No new physical entities are posited; DisenQ is an architectural module, not an invented entity. The method's numerical results rest on a handful of hand-set hyperparameters, primarily the loss weights (all 0.01) and triplet margin (0.3), plus the engineered prompt template. The claims additionally depend on the availability of ground-truth action labels during training and on the accuracy of a frozen VLM's descriptions, both domain assumptions rather than derived facts.

free parameters (4)
  • loss weights lambda1..lambda4 = all set to 0.01
    Global loss weights in Eq. 7 chosen by hand; no sensitivity analysis reported.
  • triplet margin m = 0.3
    Margin in Eq. 5 set to 0.3; no ablation or sensitivity study.
  • VLM prompt template = structured template in Supp. A
    The hand-designed template determines the text supervision content; random replacement of non-biometric text drops R@1 by 9.2%.
  • running average update for biometrics prompt = unspecified
    Iterative refinement of stored biometrics description (Sec 3.1) has no stated update rule or rate.
assumptions (4)
  • domain assumption Frozen VLM descriptions generated from key-frames are sufficiently accurate to supervise feature disentanglement.
    Section 3.1 relies on LLaVA-1.5 7B outputs; random replacement of non-biometrics text drops R@1 by 9.2%, so accuracy matters.
  • domain assumption Ground-truth action labels are available at training and provided to the VLM prompt and LAct loss.
    Sec 3.1 and Eq. 4; MEVID uses a coarse 'walking' label for all tracklets.
  • ad hoc to paper Orthogonality between pooled Fb and Fhatb ensures independence of the underlying feature distributions.
    Eq. 6 penalizes only the dot product of mean-pooled vectors; no constraint on distributions or on the motion branch.
  • ad hoc to paper Shared self-attention/cross-attention weights across the three query sets do not themselves entangle the branches.
    Sec 3.2 states query sets 'share the same self-attention and cross-attention layers'; separation is expected from different queries and text keys, but is not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DisenQ: Disentangling Q-Former for Activity-Biometrics." pith.science (2026). https://pith.science/paper/42GZ62FR

@misc{pith2026250707262,
  author       = {Pith},
  title        = {Pith review of: DisenQ: Disentangling Q-Former for Activity-Biometrics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42GZ62FR}},
  note         = {Machine review of arXiv:2507.07262}
}
read the original abstract

In this work, we address activity-biometrics, which involves identifying individuals across diverse set of activities. Unlike traditional person identification, this setting introduces additional challenges as identity cues become entangled with motion dynamics and appearance variations, making biometrics feature learning more complex. While additional visual data like pose and/or silhouette help, they often struggle from extraction inaccuracies. To overcome this, we propose a multimodal language-guided framework that replaces reliance on additional visual data with structured textual supervision. At its core, we introduce \textbf{DisenQ} (\textbf{Disen}tangling \textbf{Q}-Former), a unified querying transformer that disentangles biometrics, motion, and non-biometrics features by leveraging structured language guidance. This ensures identity cues remain independent of appearance and motion variations, preventing misidentifications. We evaluate our approach on three activity-based video benchmarks, achieving state-of-the-art performance. Additionally, we demonstrate strong generalization to complex real-world scenario with competitive performance on a traditional video-based identification benchmark, showing the effectiveness of our framework.

Figures

Figures reproduced from arXiv: 2507.07262 by the authors.

Figure 1
Figure 1. Comparison of existing and our activity-aware per￾son identification framework: (a) Existing multimodal methods [32] lacking activity-awareness leads to misidentifications due to entangled biometrics and motion features. (b) Our model disentan￾gles biometrics and motion features using language guidance, en￾abling activity-awareness and more accurate identification across diverse activities while being appearance inv… view at source ↗
Figure 2
Figure 2. Framework overview: (a) Given an RGB video, our model performs language-guided activity-aware person identification using the proposed Disentangling Q-Former (DisenQ). (b) DisenQ disentangles biometrics, motion, and non-biometrics visual features using structured language guidance and dedicated learnable queries for each feature type. (c) The frozen VLM generates structured textual descriptions from the key-frame (r… view at source ↗
Figure 3
Figure 3. Impact of DisenQ on feature disentanglement. The left plot shows the feature space of vision encoder with no feature disentan￾glement, resulting in poor identity clustering. The middle plot shows an ineffective disentanglement using cross-attention and projection, where biometrics features remain mixed with non-biometrics, causing improper clustering. In contrast, the right plot demonstrates DisenQ￾enabled disentang… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Performance analysis across activities (top 5 best and worst) on NTU RGB-AB. Here, bars and dots respectively represent person identification and action recognition accuracy. by comparing feature spaces before and after its application. Without DisenQ, feature separati…
Figure 5
Figure 5. Figure 5: Feature space of gener￾ated motion prompts across mul￾tiple runs (NTU RGB-AB subset). challenging activities do not inherently degrade identifica￾tion performance, as the model prioritizes biometrics fea￾tures when necessary, ensuring balanced identification. Utility a…
Figure 6
Figure 6. Figure 6: Top-2 rank retrieval comparison of our model and ABNet [3]. For a given probe image (left), the middle two im￾ages show incorrect matches retrieved by ABNet due to its over￾reliance on activities, while the right two images show correct matches retrieved by our model, …
Figure 7
Figure 7. Figure 7: Examples of generated structured textual descriptions [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Prompt structure vs. accuracy and VLM reliance (NTU [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Mutual info (MI) analysis. C.4. Risk of VLMs’ inherent bias propagation To address potential VLM bias linking appearance with identity, we use the VLM only to generate controlled at￾tributes’ descriptions within predefined, structured prompt templates, not free-form te…
Figure 10
Figure 10. Figure 10: Qualitative results. Here we present the top 4 rank re￾trieval results for a given probe (left) of our model on same-activity (top) and cross-activity (bottom) on NTU RGB-AB dataset [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 53 canonical work pages

  1. [1]

    Past, present, and future of face recognition: A review

    Insaf Adjabi, Abdeldjalil Ouahabi, Amir Benzaoui, and Ab- delmalik Taleb-Ahmed. Past, present, and future of face recognition: A review. Electronics, 9(8):1188, 2020. 1

    work page 2020

  2. [2]

    Flamingo: A visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems , pages 23716– 23736, 2022. 2

    work page 2022

  3. [3]

    Activity- biometrics: Person identification from daily activities

    Shehreen Azad and Yogesh Singh Rawat. Activity- biometrics: Person identification from daily activities. In CVPR, pages 287–296, 2024. 1, 2, 4, 5, 6, 8, 3

    work page 2024

  4. [4]

    Understanding depth and height percep- tion in large visual-language models

    Shehreen Azad, Yash Jain, Rishit Garg, Vibhav Vineet, and Yogesh Rawat. Understanding depth and height percep- tion in large visual-language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3611–3620, 2025. 2

    work page 2025

  5. [5]

    Hierarq: Task-aware hierarchical q-former for enhanced video understanding

    Shehreen Azad, Vibhav Vineet, and Yogesh Singh Rawat. Hierarq: Task-aware hierarchical q-former for enhanced video understanding. In Proceedings of the Computer Vi- sion and Pattern Recognition Conference, pages 8545–8556,

  6. [6]

    Salient-to-broad transition for video person re- identification

    Shutao Bai, Bingpeng Ma, Hong Chang, Rui Huang, and Xilin Chen. Salient-to-broad transition for video person re- identification. In CVPR, pages 7339–7348, 2022. 2, 5

    work page 2022

  7. [7]

    Pstr: End-to-end one-step person search with trans- formers

    Jiale Cao, Yanwei Pang, Rao Muhammad Anwer, Hisham Cholakkal, Jin Xie, Mubarak Shah, and Fahad Shahbaz Khan. Pstr: End-to-end one-step person search with trans- formers. In CVPR, pages 9458–9467, 2022. 2, 5

    work page 2022

  8. [8]

    Learning recurrent 3d attention for video-based person re- identification

    Guangyi Chen, Jiwen Lu, Ming Yang, and Jie Zhou. Learning recurrent 3d attention for video-based person re- identification. IEEE TIP, 29:6963–6976, 2020. 2

    work page 2020

Show all 64 references
  1. [9]

    Learning 3d shape feature for texture-insensitive person re-identification

    Jiaxing Chen, Xinyang Jiang, Fudong Wang, Jun Zhang, Feng Zheng, Xing Sun, and Wei-Shi Zheng. Learning 3d shape feature for texture-insensitive person re-identification. In CVPR, pages 8146–8155, 2021. 2

    2021

  2. [10]

    Unveiling the power of clip in unsupervised visible-infrared person re-identification

    Zhong Chen, Zhizhong Zhang, Xin Tan, Yanyun Qu, and Yuan Xie. Unveiling the power of clip in unsupervised visible-infrared person re-identification. In Proceedings of the 31st ACM International Conference on Multimedia , pages 3667–3675, 2023. 2, 5, 6

    2023

  3. [11]

    Instructblip: Towards general- purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general- purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 5, 8

    2023 arXiv

  4. [12]

    Mevid: Multi-view extended videos with identities for video person re-identification

    Daniel Davila, Dawei Du, Bryon Lewis, Christopher Funk, Joseph Van Pelt, Roderic Collins, Kellie Corona, Matt Brown, Scott McCloskey, Anthony Hoogs, et al. Mevid: Multi-view extended videos with identities for video person re-identification. In WACV, pages 1634–1643, 2023. 5

    2023

  5. [13]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5

    2018 arXiv

  6. [14]

    Learning disentangled rep- resentation for robust person re-identification

    Chanho Eom and Bumsub Ham. Learning disentangled rep- resentation for robust person re-identification. Adv. Neural Inform. Process. Syst., 32, 2019. 6

    2019

  7. [15]

    Video-based person re-identification with spatial and tempo- ral memory networks

    Chanho Eom, Geon Lee, Junghyup Lee, and Bumsub Ham. Video-based person re-identification with spatial and tempo- ral memory networks. In ICCV, pages 12036–12045, 2021. 5, 2

    2021

  8. [16]

    Gaitpart: Temporal part-based model for gait recognition

    Chao Fan, Yunjie Peng, Chunshui Cao, Xu Liu, Saihui Hou, Jiannan Chi, Yongzhen Huang, Qing Li, and Zhiqiang He. Gaitpart: Temporal part-based model for gait recognition. In CVPR, pages 14225–14233, 2020. 1

    2020

  9. [17]

    Opengait: Revisiting gait recognition towards better practicality

    Chao Fan, Junhao Liang, Chuanfu Shen, Saihui Hou, Yongzhen Huang, and Shiqi Yu. Opengait: Revisiting gait recognition towards better practicality. In CVPR, pages 9707–9716, 2023. 1

    2023

  10. [18]

    Appearance-preserving 3d convolution for video-based person re-identification

    Xinqian Gu, Hong Chang, Bingpeng Ma, Hongkai Zhang, and Xilin Chen. Appearance-preserving 3d convolution for video-based person re-identification. In ECCV, pages 228–

  11. [19]

    Clothes-changing person re-identification with rgb modality only

    Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. Clothes-changing person re-identification with rgb modality only. In CVPR, pages 1060–1069, 2022. 1, 2, 4, 5, 6

    2022

  12. [20]

    Semantic-aware consistency network for cloth- changing person re-identification

    Peini Guo, Hong Liu, Jianbing Wu, Guoquan Wang, and Tao Wang. Semantic-aware consistency network for cloth- changing person re-identification. In ACM MM, 2023. 2, 5

    2023

  13. [21]

    Dense interaction learning for video-based person re-identification

    Tianyu He, Xin Jin, Xu Shen, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Dense interaction learning for video-based person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 1490–1501, 2021. 2

    2021

  14. [22]

    Instruct-reid: A multi-purpose per- son re-identification task with instructions

    Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, et al. Instruct-reid: A multi-purpose per- son re-identification task with instructions. In CVPR, pages 17521–17531, 2024. 1, 2, 5, 6

    2024

  15. [23]

    Fine-grained shape-appearance mutual learning for cloth-changing person re-identification

    Peixian Hong, Tao Wu, Ancong Wu, Xintong Han, and Wei- Shi Zheng. Fine-grained shape-appearance mutual learning for cloth-changing person re-identification. In CVPR, pages 10513–10522, 2021. 2

    2021

  16. [24]

    Temporal complementary learning for video person re-identification

    Ruibing Hou, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Temporal complementary learning for video person re-identification. In ECCV, pages 388–405. Springer,

  17. [25]

    Bicnet-tks: Learning efficient spatial- temporal representation for video person re-identification

    Ruibing Hou, Hong Chang, Bingpeng Ma, Rui Huang, and Shiguang Shan. Bicnet-tks: Learning efficient spatial- temporal representation for video person re-identification. In CVPR, pages 2014–2023, 2021. 2, 5, 6 9

    2014

  18. [26]

    Empowering visible- infrared person re-identification with large foundation mod- els

    Zhangyi Hu, Bin Yang, and Mang Ye. Empowering visible- infrared person re-identification with large foundation mod- els. In NeurIPS, 2024. 2, 5, 6

    2024

  19. [27]

    Celebrities-reid: A benchmark for clothes variation in long- term person re-identification

    Yan Huang, Qiang Wu, Jingsong Xu, and Yi Zhong. Celebrities-reid: A benchmark for clothes variation in long- term person re-identification. In IJCNN, pages 1–8. IEEE,

  20. [28]

    Rethink- ing temporal fusion for video-based person re-identification on semantic and time aspect

    Xinyang Jiang, Yifei Gong, Xiaowei Guo, Qize Yang, Feiyue Huang, Wei-Shi Zheng, Feng Zheng, and Xing Sun. Rethink- ing temporal fusion for video-based person re-identification on semantic and time aspect. In AAAI, pages 11133–11140,

  21. [29]

    Cloth-changing person re- identification from a single image with gait prediction and regularization

    Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re- identification from a single image with gait prediction and regularization. In CVPR, pages 14278–14287, 2022. 2

    2022

  22. [30]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2

    2024 arXiv

  23. [31]

    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 ICML, pages 19730–19742. PMLR, 2023. 2, 3

    2023

  24. [32]

    Clip-reid: exploiting vision-language model for image re-identification without concrete text labels

    Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 1, 2, 5, 6

    2023

  25. [33]

    Gaitedge: Beyond plain end-to-end gait recognition for better practicality

    Junhao Liang, Chao Fan, Saihui Hou, Chuanfu Shen, Yongzhen Huang, and Shiqi Yu. Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In ECCV, pages 375–390. Springer, 2022. 1

    2022

  26. [34]

    Differ: Disentangling iden- tity features via semantic cues for clothes-changing person re-id

    Xin Liang and Yogesh S Rawat. Differ: Disentangling iden- tity features via semantic cues for clothes-changing person re-id. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13980–13989, 2025. 2

    2025

  27. [35]

    Gait recognition via effective global-local feature representation and local tempo- ral aggregation

    Beibei Lin, Shunli Zhang, and Xin Yu. Gait recognition via effective global-local feature representation and local tempo- ral aggregation. In ICCV, pages 14648–14656, 2021. 1

    2021

  28. [36]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892–34916, 2023. 2, 5, 8

    2023

  29. [37]

    Deeply coupled convolution–transformer with spatial– temporal complementary learning for video-based person re- identification

    Xuehu Liu, Chenyang Yu, Pingping Zhang, and Huchuan Lu. Deeply coupled convolution–transformer with spatial– temporal complementary learning for video-based person re- identification. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2

    2023

  30. [38]

    Magface: A universal representation for face recognition and quality assessment

    Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In CVPR, pages 14225–14234, 2021. 1

    2021

  31. [39]

    Gpt-4 technical report

    OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2024. 8

    2024 arXiv

  32. [40]

    Fine-grained re-identification

    Priyank Pathak. Fine-grained re-identification. arXiv preprint arXiv:2011.13475, 2020. 2

    2011 arXiv

  33. [41]

    Colors see colors ig- nore: Clothes changing reid with color disentanglement

    Priyank Pathak and Yogesh S Rawat. Colors see colors ig- nore: Clothes changing reid with color disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 2

    2025

  34. [42]

    Coarse attribute pre- diction with task agnostic distillation for real world clothes changing reid

    Priyank Pathak and Yogesh S Rawat. Coarse attribute pre- diction with task agnostic distillation for real world clothes changing reid. arXiv preprint arXiv:2505.12580, 2025. 2

    2025

  35. [43]

    Video person re-id: Fantastic techniques and where to find them (student abstract)

    Priyank Pathak, Amir Erfan Eshratifar, and Michael Gormish. Video person re-id: Fantastic techniques and where to find them (student abstract). InAAAI, pages 13893– 13894, 2020. 1, 2, 6

    2020

  36. [44]

    Robust re-identification by multiple views knowledge distil- lation

    Angelo Porrello, Luca Bergamini, and Simone Calderara. Robust re-identification by multiple views knowledge distil- lation. In ECCV, pages 93–110. Springer, 2020. 2, 5

    2020

  37. [45]

    Long-term cloth-changing person re-identification

    Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yanwei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. Long-term cloth-changing person re-identification. In ACCV, 2020. 2

    2020

  38. [46]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2

    2021

  39. [47]

    Probing conceptual understanding of large visual-language models

    Madeline Schiappa, Raiyaan Abdullah, Shehreen Azad, Jared Claypoole, Michael Cogswell, Ajay Divakaran, and Yogesh Rawat. Probing conceptual understanding of large visual-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pag...

    2024

  40. [48]

    Robustness analysis on foundational segmenta- tion models

    Madeline Chantry Schiappa, Shehreen Azad, Sachidanand Vs, Yunhao Ge, Ondrej Miksik, Yogesh S Rawat, and Vib- hav Vineet. Robustness analysis on foundational segmenta- tion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1786...

    2024

  41. [49]

    Eva-clip: Improved training techniques for clip at scale

    Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 5, 6, 8, 2

    2023 arXiv

  42. [50]

    When large vision-language models meet person re-identification

    Qizao Wang, Bin Li, and Xiangyang Xue. When large vision-language models meet person re-identification. arXiv preprint arXiv:2411.18111, 2024. 2

    2024 arXiv

  43. [51]

    Pyramid spatial-temporal aggrega- tion for video-based person re-identification

    Yingquan Wang, Pingping Zhang, Shang Gao, Xia Geng, Hu Lu, and Dong Wang. Pyramid spatial-temporal aggrega- tion for video-based person re-identification. In ICCV, pages 12026–12035, 2021. 2, 5, 6

    2021

  44. [52]

    Internvideo2: Scaling foundation models for multimodal video understanding

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yan- song Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In ECCV, pages 396–416. Springer, 2024. 8

    2024

  45. [53]

    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 , 32:6032–6046,

  46. [54]

    Person re- identification by contour sketch under moderate clothing change

    Qize Yang, Ancong Wu, and Wei-Shi Zheng. Person re- identification by contour sketch under moderate clothing change. IEEE TPAMI, 43(6):2029–2046, 2019. 2

    2019

  47. [55]

    Good is bad: Causality inspired cloth-debiasing for 10 cloth-changing person re-identification

    Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. Good is bad: Causality inspired cloth-debiasing for 10 cloth-changing person re-identification. In CVPR, pages 1472–1481, 2023. 2, 5

    2023

  48. [56]

    A pedestrian is worth one prompt: Towards language guidance person re-identification

    Zexian Yang, Dayan Wu, Chenming Wu, Zheng Lin, Jingzi Gu, and Weiping Wang. A pedestrian is worth one prompt: Towards language guidance person re-identification. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17343–17353, 2024. 2

    2024

  49. [57]

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

    Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re- identification: A survey and outlook. IEEE TPAMI, 44(6): 2872–2893, 2021. 1

    2021

  50. [58]

    Tf-clip: Learning text-free clip for video- based person re-identification

    Chenyang Yu, Xuehu Liu, Yingquan Wang, Pingping Zhang, and Huchuan Lu. Tf-clip: Learning text-free clip for video- based person re-identification. In AAAI, pages 6764–6772,

  51. [59]

    Multidirection and multiscale pyramid in transformer for video-based pedestrian retrieval

    Xianghao Zang, Ge Li, and Wei Gao. Multidirection and multiscale pyramid in transformer for video-based pedestrian retrieval. IEEE Transactions on Industrial Informatics , 18 (12):8776–8785, 2022. 6

    2022

  52. [60]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023. 8

    2023

  53. [61]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2

    2023 arXiv

  54. [62]

    Sharc: Shape and appearance recognition for per- son identification in-the-wild

    Haidong Zhu, Wanrong Zheng, Zhaoheng Zheng, and Ram Nevatia. Sharc: Shape and appearance recognition for per- son identification in-the-wild. In WACV, pages 6290–6300,

  55. [243]

    Springer, 2020. 5, 6

    2020

  56. [2024]

    a [body shape] person

    6 11 DisenQ: Disentangling Q-Former for Activity-Biometrics Supplementary Material In this supplementary material, we provide details of prompt generation in Section A, along with structured ex- amples of generated descriptions. Then we present addi- tional quantitative result...

Pith tools

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