Pith. sign in

REVIEW 4 major objections 7 minor 85 references

PRISM fuses RGB, NIR and TIR with Prompt-S6 and semantic token pruning to make multi-modal object re-identification more accurate and efficient.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 21:48 UTC pith:BYUUGAM6

load-bearing objection Solid engineering stack for tri-modal ReID: PS6’s decoupled B/C prompting is a real operator-level idea, backed by thorough ablations and efficiency numbers, but the SOTA language overreaches on thin margins and one table is internally inconsistent. the 4 major comments →

arxiv 2607.23451 v1 pith:BYUUGAM6 submitted 2026-07-26 cs.CV

Multi-Modal Object Re-Identification with Prompt-S6 and Semantic-Aware Knowledge Guidance

Weixiang Zhou , Jiabei Zuo , Yuhao Wang , Cong Wang , Huchuan Lu , Zhixun Su This is my paper
classification cs.CV
keywords multi-modal object re-identificationPrompt-S6semantic-aware learningtoken pruningstate space modelsMambaRGB-NIR-TIRprogressive fusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Matching the same person or vehicle across cameras fails when RGB alone meets darkness, glare or occlusion, so multi-modal ReID combines visible, near-infrared and thermal images. Prior work either leaves background clutter in the features, fuses only two modalities at a time, or pays quadratic attention cost. This paper introduces PRISM: a Prompt-S6 operator that keeps Mamba’s linear complexity while letting external prompts (masks or other spectra) control what is written into and read from the state, plus Semantic-Driven Token Pruning that uses foundation-model masks to drop background tokens, and a Progressive Fusion Network that stages intra-, inter- and full tri-modal alignment. On four public benchmarks the resulting representations raise mAP and Rank-1 while using less memory than recent alternatives, showing that semantic-guided pruning plus efficient tri-modal state-space fusion can make multi-spectral ReID practical under real degradations.

Core claim

The authors claim that Prompt-S6—decoupling the state-space write (B) and read (C) projections so each can be conditioned on a different modality—together with Semantic-Driven Token Pruning and a three-stage Progressive Fusion Network, yields more robust tri-modal (RGB–NIR–TIR) ReID features than pairwise attention or earlier Mamba fusions, reaching state-of-the-art or better accuracy and efficiency on four multi-modal object ReID benchmarks.

What carries the argument

Prompt-S6 (PS6): a selective state-space block in which the input projection B is generated from one prompt sequence and the output projection C from another, so masks or other spectra can asymmetrically gate state injection and emission while preserving linear complexity; it powers both Target-Aware Feature Enhancement inside SDTP and the inter-/tri-modal stages of PFN.

Load-bearing premise

Offline masks taken only from the RGB image (pose keypoints for people, box-prompted SAM2 for vehicles) stay good enough semantic priors under darkness, occlusion and thermal distortion; if they miss or shatter the object, pruning discards identity cues.

What would settle it

Swap the RGB-derived masks for empty, randomly shifted, or NIR/TIR-generated masks and re-measure RGBNT201 mAP; a collapse toward the no-mask ablation (~75.8%) would confirm the claim depends on those priors, while stable high accuracy would falsify it.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Background can be suppressed in multi-modal ReID without quadratic attention by PS6-conditioned token pruning.
  • Full tri-modal alignment is reachable in a progressive linear-complexity pipeline rather than only pairwise cross-attention.
  • Offline foundation-model masks suffice for both person and vehicle ReID when tokens are kept via cross-modal top-k union.
  • The same PS6 operator improves low-level multi-modal fusion metrics, indicating reuse beyond ReID.
  • Early sequence shortening plus linear SSM cost keeps memory and FLOPs competitive for deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Decoupled B/C prompting may transfer to other multi-sensor stacks (RGB-depth, multi-spectral remote sensing) where one stream should gate another without early fusion.
  • If mask quality is the bottleneck, a lightweight jointly trained mask head could remove the offline dependency while retaining the pruning logic.
  • Cross-modal union at 0.5 retention is a reusable sparsification recipe worth testing in video ReID and multi-camera tracking.
  • Linear tri-modal interaction makes edge multi-spectral ReID more plausible than transformer-heavy alternatives.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The manuscript presents PRISM, a tri-modal (RGB–NIR–TIR) object ReID framework built on a proposed Prompt-S6 (PS6) operator — a Mamba-style selective SSM in which the input projection B and output projection C are generated from decoupled auxiliary prompt sequences (B = Linear_B(t), C = Linear_C(t_p)) rather than from the primary sequence or early-fused features. On top of PS6 the authors build Semantic-Driven Token Pruning (SDTP: Target-Aware Feature Enhancement plus a Tail Drop Module with Cross-Modal Union over top-k token indices, retention 0.5) guided by offline RGB-only masks (OpenPifPaf keypoints for persons, SAM2 with dynamic box prompts for vehicles), and a Progressive Fusion Network (PFN: intra-modal, pairwise inter-modal, and cyclic tri-modal aggregation). The paper reports state-of-the-art or competitive results on RGBNT201 (80.5 mAP / 84.0 R-1), RGBNT100, MSVR310, and WMVEID863, with ablations (Tables IX–XI, XV, XVI, XX), robustness studies under synthetic misalignment and occlusion (Tables V–VIII), efficiency comparisons (Tables IV, XVII–XIX), qualitative analyses, and a public code release.

Significance. If the results hold, the paper contributes a genuinely lightweight mechanism for injecting cross-modal control into SSM state dynamics (decoupled B/C prompt conditioning), which is a cleaner design point than the early-fusion or shared-state multi-modal Mamba variants it contrasts with, and the efficiency story (Table XVII linear scaling, Table XIX lowest memory at 1718 MiB) is practically relevant for deployment. Strengths that deserve explicit credit: a public code release; evaluation on four benchmarks spanning person and vehicle ReID; a reasonably complete ablation ladder (module-level, component-level, stacking depth, retention rate, mask-generator choice); robustness studies under pixel-shift misalignment and occlusion that go beyond clean-condition leaderboard numbers; and an honest acknowledgment (§IV.B.4) that MFRNet and IDEA beat PRISM in some degraded regimes. The mask-failure concern is bounded by the paper's own evidence: the w/o-mask ablation (Table IX, model C) still yields 75.8 mAP on RGBNT201, so SDTP degrades gracefully to a competitive baseline rather than collapsing. The main threats to significance are not conceptual but evidentiary: the superiority claim rests

major comments (4)
  1. [§IV.C.4, Table XIV (vs. Table I and Table XII)] Table XIV is internally inconsistent with the headline results and the stated pipeline. The table is captioned RGBNT201, and its 'Ours' row is 'SAM2 w/ Dynamic Box Prompt' at 80.5 mAP / 85.9 Rank-1. But §III.A and §IV.C.4 state that persons on RGBNT201 use OpenPifPaf keypoint-aggregated masks, and the Table I headline (80.5 / 84.0) matches Table XII's OpenPifPaf-RGB row, not Table XIV. So Table XIV's 'Ours' row either (a) describes a different mask pipeline than the headline model while being labeled 'Ours', or (b) is mislabeled/mismeasured (mAP coincidentally identical at 80.5 while Rank-1 jumps 84.0 → 85.9). Either way the configuration-to-number mapping is not self-consistent, and readers cannot tell which pipeline produced the reported SOTA number. This must be corrected and the affected tables (I, XII, XIV) reconciled.
  2. [§IV.B.1–2, Tables I–III, V, X, XVI] The SOTA-superiority claims are overstated relative to the paper's own tables. §IV.B.1 claims PRISM 'surpasses existing SOTA methods across all metrics' on RGBNT201, yet the clean-condition mAP margin over IDEA is 0.3 (80.5 vs 80.2, Table I) and MFRNet's clean mAP is 80.7 (Table V), which the text itself acknowledges. On WMVEID863 the margin over FACENet/MambaPro is 0.9 mAP (Table III). On RGBNT100 PRISM's 86.1 mAP loses to IDEA (87.2) and DeMo (86.2) (Table II), and on MSVR310 it loses mAP to PromptMA (55.2), DeMo (49.2) by wide margins — the text says 'competitive' there, which is fair, but the abstract/contribution framing of broad superiority is not. More importantly, all numbers are single-run on small test sets (MSVR310: 2,087 triplets; WMVEID863: 4,709), and sub-1% ReID margins are routinely within seed/schedule noise. The ablation ladder has the same problem (e.g., Table X: 79.3
  3. [§III.B.1, Table IV] The central architectural claim for PS6 — that decoupling B = Linear_B(t) and C = Linear_C(t_p) yields strictly more flexible cross-modal control than early-sum/concat multi-modal SSMs (§III.B.1) — is argued only by construction and supported by a single aggregate comparison (Table IV). Table IV compares PS6 against Mamba/VMamba/Vision Mamba/CrossMamba swaps, which is useful, but it does not isolate the decoupling itself: there is no ablation where B and C are both derived from the same prompt (t = t_p), or where B, C come from early-fused features while keeping everything else fixed. Since the decoupling is presented as 'the core technical contribution of PS6', a targeted ablation (shared-prompt vs. decoupled-prompt, same parameter budget) is needed to show the gain comes from the claimed mechanism rather than from, e.g., the extra prompt pathway capacity.
  4. [§III.A, §III.B.3 Eq. (6), §IV.C.4–5, Tables V–VIII, XV] The mask-reliability assumption is load-bearing for SDTP and only partially stress-tested. All masks are generated offline from RGB only (§III.A, §IV.C.4), and the w/o-mask ablation shows a 4.7 mAP drop on RGBNT201 (Table IX, D→C), so SDTP's contribution is real but mask-dependent. The robustness studies (Tables V–VIII) shift/occlude the images but do not appear to corrupt the masks correspondingly — i.e., the evaluated degradation model assumes the offline mask stays clean while the inputs degrade. In a real deployment with low-light RGB, the OpenPifPaf/SAM2 masks would degrade too. A small experiment with masks recomputed on degraded RGB (or synthetically eroded/dilated masks) would directly test the stated claim that 'the mask serves as a coarse semantic prior' and that CMU confers robustness to mask imperfection. Relatedly, Table XV fixes retention at 0.5 with the justification invok
minor comments (7)
  1. [Algorithm 1] Algorithm 1, line 6: 'Parameter A ∈ R^(Np,1)' — the state matrix A depending on sequence length Np is unusual; presumably this is a typo for the state dimension. Please check shapes throughout Algorithm 1 (e.g., Δ ∈ R^(D,D) on line 4 vs. the broadcast on line 7).
  2. [§III.B.2, Eq. (3)] Eq. (3): PS6(LN(F_m), LN(F_m), M_m) passes the modality feature as both primary input and gating prompt, with the mask only as projection prompt. A sentence of motivation for this asymmetric choice (why not mask as gating prompt, which the text's 'soft gating mechanism' language in §III.B.1 suggests) would help.
  3. [§IV.C.5, Table XV] Table XV caption and text disagree on terminology: the table compares 'Union' vs 'Separate' token selection, while the text discusses 'drop rate' and later 'retention rate'; please unify. Also state explicitly whether the 0.5 column header is retention or drop rate.
  4. [Figs. 1–2] Fig. 1(b) is referenced as showing efficiency/parameter comparisons but its printed content (model sizes 2M/5M, 'TAFE: 3.94M, PS6: 0.79M') is hard to parse as typeset; consider redrawing with axis labels. Fig. 2 is dense; the TAFE/TDM/PFN signal flow would benefit from annotated tensor shapes.
  5. [§IV.D.1, Table XXI] Table XXI: the separation metric for IDEA (2.9998) is lower than MambaPro (3.5274) despite IDEA's higher mAP; a brief comment on why separation/overlap do not track mAP monotonically across methods would preempt confusion.
  6. [§IV.A.2] §IV.A.2: batch size and epochs for WMVEID863 are not given (only RGBNT201/MSVR310 and RGBNT100 are specified). Please complete the training-protocol details for reproducibility, and report the number of training runs per table.
  7. [References] Several recently cited works appear as arXiv preprints ([47], [71], [80]) or with future-dated venues ([11], [18]–[20]); please update to published versions where available.

Circularity Check

0 steps flagged

No derivation circularity: empirical ReID architecture paper evaluated on held-out public benchmarks.

full rationale

PRISM is an engineering/methods paper. Its central claims are architectural (PS6 as a prompt-decoupled SSM operator; SDTP with TAFE+TDM; three-stage PFN) and empirical (mAP/CMC on RGBNT201, RGBNT100, MSVR310, WMVEID863). There is no first-principles derivation that reduces a claimed prediction to a fitted constant or to a self-defined quantity. Losses are standard label-smoothed CE + triplet (Eqs. 18–19). Masks come from external frozen segmentors (OpenPifPaf, SAM2) used offline as priors, not as quantities fitted to the ReID labels being predicted. Hyperparameters (retention 0.5, TAFE/PFN stack depths) are chosen by ablation on the same domains—an ordinary ML practice, not a fitted-input-called-prediction loop. Overlapping-author citations (TOP-ReID, MambaPro, IDEA, DeMo) appear as baselines to beat, not as uniqueness theorems that force the design. No equation equates a reported metric to its own input by construction. Table inconsistencies and sub-1% margins noted by the skeptic are evidence-quality issues, not circularity. Score 0; steps empty.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Load-bearing content is architectural and empirical. The claim rests on standard SSM/Mamba math, CLIP features, offline foundation-model masks, and several hand-chosen depths/rates tuned on the target benchmarks. No new physical entities; invented items are neural modules.

free parameters (5)
  • TDM retention / drop rate = 0.5
    Token retention rate swept and set to 0.5 via CMU union on RGBNT201 (Table XV); directly affects which patches enter PFN.
  • TAFE stack depth k = 2
    Number of TAFE/PS6 layers chosen by ablation; k=2 best (Table X).
  • PFN IT/TA stack configuration = Intra=1, Inter=2, Tri=2
    Intra/Inter/Tri module counts ablated; (1,2,2) selected (Table XVI).
  • Optimizer learning rates and schedule = 3.5e-4 / 5e-6
    Adam 3.5e-4 overall, 5e-6 for visual encoder; epochs 50–60 and batch sizes dataset-specific (§IV.A.2).
  • Top-k / union token selection thresholding inside scorers Θ_m
    Per-modality scorers and cross-modal union define kept index set; coupled to retention rate and not uniquely determined by theory.
axioms (5)
  • domain assumption Selective SSM/Mamba recurrence with input-dependent Δ, B, C correctly models long token sequences at linear cost (Gu & Dao; Algorithm 1).
    PS6 inherits Mamba’s SSM kernel; correctness of sequence mixing rests on this background theory (§III.B.1).
  • domain assumption CLIP vision encoder features are a suitable shared backbone for RGB/NIR/TIR ReID patches and class tokens.
    All main results use frozen-or-finetuned CLIP visual encoder F_visual (§III.A, §IV.A).
  • domain assumption RGB-derived segmentation masks (OpenPifPaf keypoints for persons; SAM2 box prompts for vehicles) supply spatially useful foreground priors even under NIR/TIR degradation and mild misalignment.
    SDTP is defined around M_m from F_semantic(I_R); ablations show large dependence on this prior (§III.A–B, Tables XII–XIV).
  • domain assumption Label-smoothing cross-entropy plus triplet loss on class tokens and fused embeddings is an adequate training objective for multi-modal ReID.
    Objective copied from DeMo-style setup (§III.D, §IV.A.2).
  • ad hoc to paper Decoupling Linear_B(t) and Linear_C(t_p) yields strictly more flexible cross-modal control than early-sum/concat multi-modal SSMs.
    Core design thesis of PS6 (§III.B.1); supported empirically vs Mamba variants (Table IV) but not proved.
invented entities (3)
  • Prompt-S6 (PS6) no independent evidence
    purpose: Lightweight cross-modal SSM block where gating prompt supplies B and projection prompt supplies C for asymmetric state control.
    Central new operator; enables TAFE and PFN. Independent evidence limited to ablations and a small image-fusion transfer (Table XXII), not external theory.
  • Semantic-Driven Token Pruning (SDTP: TAFE + TDM/CMU) no independent evidence
    purpose: Mask-conditioned enhancement then cross-modal consensus tail-drop of background tokens.
    System-level module defined in §III.B; gains tied to paper’s masks and scorers.
  • Progressive Fusion Network (PFN: IM + IT + TA) no independent evidence
    purpose: Three-stage path from intra-modal MHSA to pairwise PS6 to cyclic tri-modal PS6 aggregation.
    Architectural pipeline in §III.C; not a physical entity, no external falsifiable handle beyond ReID metrics.

pith-pipeline@v1.2.0-grok45-kimik3 · 29243 in / 4122 out tokens · 86505 ms · 2026-07-30T21:48:27.462137+00:00 · methodology

0 comments
read the original abstract

Multi-modal object Re-Identification (ReID) aims to retrieve specific objects by integrating complementary information from multiple modalities. However, existing multi-modal ReID methods do not effectively address background interference suppression or achieve tri-modal alignment, instead focusing on pairwise feature fusion. Moreover, many current aggregation approaches suffer from high computational complexity. To address these limitations, we propose PRISM, a novel multi-modal ReID framework built upon Prompt-S6 (PS6) and semantic-aware knowledge guidance. PS6 maintains the linear complexity and strong sequence modeling capability of Mamba while enabling efficient cross-modal interaction. Leveraging these advantages, we design two key components: Semantic-Driven Token Pruning (SDTP) and Progressive Fusion Network (PFN). Parsing semantic priors from the segmentation foundation models, the SDTP then leverages these priors and applies dynamic token pruning to suppress background noise and refine feature representations. The PFN progressively aggregates multi-modal features to achieve tri-modal alignment and fully exploit modality complementarity. With the proposed modules, PRISM generates more robust multi-modal representations under complex scenarios. Extensive experiments on four multi-modal object ReID benchmarks demonstrate the effectiveness and efficiency of our approach. The source code is available at https://github.com/zw-absin/PRISM.

Figures

Figures reproduced from arXiv: 2607.23451 by Cong Wang, Huchuan Lu, Jiabei Zuo, Weixiang Zhou, Yuhao Wang, Zhixun Su.

Figure 1
Figure 1. Figure 1: (a) Our method addresses two key limitations: ignoring background information and struggling with tri-modal alignment. These are addressed through semantic-driven token pruning and a progressive fusion network. (b) Enabled by PS6, our framework achieves stronger cross-modal interactions and improved performance. (c) Token selection visualization demonstrates that our method retains more informative regions… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed PRISM framework. (a) Semantic masks and multi-spectral features are extracted using a segmentation network and vision [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Implementation details of the PFN, illustrated with an RGB modality input [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance evaluation under challenging conditions on the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of semantic masks generated on the RGBNT201 and RGBNT100 datasets. Our mask generation approach produces more [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cosine similarity distribution visualization. Incorporating IT and TA improves separation and reduces overlap, indicating enhanced discriminative [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of feature distributions using t-SNE [ [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Rank list comparison of our model with varying components on RGBNT201. The progressive performance improvement with the addition of each [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Rank list comparison of our model with varying components on RGBNT100. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of channel activation maps. Interacting with semantic [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of channel activation maps for vehicle ReID. The inter [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

85 extracted references · 16 linked inside Pith

  1. [1]

    Receptive multi- granularity representation for person re-identification,

    G. Wang, Y . Yuan, J. Li, S. Ge, and X. Zhou, “Receptive multi- granularity representation for person re-identification,”TIP, 2020

    2020

  2. [2]

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

    X. Liu, C. Yu, P. Zhang, and H. Lu, “Deeply coupled convolution– transformer with spatial–temporal complementary learning for video- based person re-identification,”TNNLS, 2023

    2023

  3. [3]

    Other tokens matter: Exploring global and local features of vision transformers for object re-identification,

    Y . Wang, P. Zhang, D. Wang, and H. Lu, “Other tokens matter: Exploring global and local features of vision transformers for object re-identification,”CVIU, 2024

    2024

  4. [4]

    A video is worth three views: Trigeminal transformers for video-based person re- identification,

    X. Liu, P. Zhang, C. Yu, X. Qian, X. Yang, and H. Lu, “A video is worth three views: Trigeminal transformers for video-based person re- identification,”TITS, 2024. 15

    2024

  5. [5]

    Watching you: Global- guided reciprocal learning for video-based person re-identification,

    X. Liu, P. Zhang, C. Yu, H. Lu, and X. Yang, “Watching you: Global- guided reciprocal learning for video-based person re-identification,” in CVPR, 2021

    2021

  6. [6]

    Hat: Hierarchical aggregation transformers for person re-identification,

    G. Zhang, P. Zhang, J. Qi, and H. Lu, “Hat: Hierarchical aggregation transformers for person re-identification,” inACM MM, 2021

    2021

  7. [7]

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

    C. Yu, X. Liu, Y . Wang, P. Zhang, and H. Lu, “Tf-clip: Learning text-free clip for video-based person re-identification,” inAAAI, 2024

    2024

  8. [8]

    Joint self-attention and scale- aggregation for self-calibrated deraining network,

    C. Wang, Y . Wu, Z. Su, and J. Chen, “Joint self-attention and scale- aggregation for self-calibrated deraining network,” inACM MM, 2020

    2020

  9. [9]

    Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution,

    L. Wang, C. Wang, J. Pan, X. Liu, W. Zhou, X. Sun, W. Wang, and Z. Su, “Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution,”TCSVT, 2025

    2025

  10. [10]

    Adapting large vlms with iterative and manual instructions for generative low-light enhancement,

    X. Sun, L. Wang, C. Wang, Y . Jin, K.-m. Lam, Z. Su, Y . Yang, and J. Pan, “Adapting large vlms with iterative and manual instructions for generative low-light enhancement,”arXiv:2507.18064, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Neural discrimination-prompted transformers for efficient uhd image restoration and enhancement: C. wang et al

    C. Wang, J. Pan, L. Wang, W. Wang, and Y . Yang, “Neural discrimination-prompted transformers for efficient uhd image restoration and enhancement: C. wang et al.”IJCV, 2026

    2026

  12. [12]

    Deep learning-driven ultra-high-definition image restoration: A survey,

    L. Wang, W. Zhou, C. Wang, K.-M. Lam, Z. Su, and J. Pan, “Deep learning-driven ultra-high-definition image restoration: A survey,” arXiv:2505.16161, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Top-reid: Multi- spectral object re-identification with token permutation,

    Y . Wang, X. Liu, P. Zhang, H. Lu, Z. Tu, and H. Lu, “Top-reid: Multi- spectral object re-identification with token permutation,” inAAAI, 2024

    2024

  14. [14]

    Learning commonality, divergence and variety for unsupervised visible-infrared person re- identification,

    J. Shi, X. Yin, Y . Zhang, Y . Xie, Y . Quet al., “Learning commonality, divergence and variety for unsupervised visible-infrared person re- identification,”NeurIPS, 2024

    2024

  15. [15]

    Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification,

    B. Yang, J. Chen, and M. Ye, “Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification,” inCVPR, 2024

    2024

  16. [16]

    Towards grand unified representation learning for unsupervised visible-infrared person re-identification,

    ——, “Towards grand unified representation learning for unsupervised visible-infrared person re-identification,” inICCV, 2023

    2023

  17. [17]

    Video-level language- driven video-based visible-infrared person re-identification,

    S. Li, J. Leng, C. Kuang, M. Tan, and X. Gao, “Video-level language- driven video-based visible-infrared person re-identification,”TIFS, 2025

    2025

  18. [18]

    Spatial-temporal high-frequency learning for video-based visible-infrared person re- identification,

    S. Tao, S. Li, J. Ye, N. Dong, F. Li, and H. Li, “Spatial-temporal high-frequency learning for video-based visible-infrared person re- identification,”TCSVT, 2026

    2026

  19. [19]

    Hierarchical prompt learning for image-and text-based person re-identification,

    L. Zhou, S. Li, N. Dong, Y . Tai, Y . Zhang, and H. Li, “Hierarchical prompt learning for image-and text-based person re-identification,” in AAAI, 2026

    2026

  20. [20]

    Dynamic-static collaboration for unsupervised domain adaptive video-based visible-infrared person re-identification,

    J. Leng, Z. Wang, S. Li, and X. Gao, “Dynamic-static collaboration for unsupervised domain adaptive video-based visible-infrared person re-identification,” inAAAI, 2026

    2026

  21. [21]

    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,”TPAMI, 2021

    2021

  22. [22]

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

    P. Zhang, Y . Wang, Y . Liu, Z. Tu, and H. Lu, “Magic tokens: Select diverse tokens for multi-modal object re-identification,” inCVPR, 2024

    2024

  23. [23]

    Mask-guided contrastive attention model for person re-identification,

    C. Song, Y . Huang, W. Ouyang, and L. Wang, “Mask-guided contrastive attention model for person re-identification,” inCVPR, 2018

    2018

  24. [24]

    A mask based deep ranking neural network for person retrieval,

    L. Qi, J. Huo, L. Wang, Y . Shi, and Y . Gao, “A mask based deep ranking neural network for person retrieval,” inICME, 2019

    2019

  25. [25]

    Multi-biometric unified network for cloth-changing person re-identification,

    G. Zhang, J. Liu, Y . Chen, Y . Zheng, and H. Zhang, “Multi-biometric unified network for cloth-changing person re-identification,”TIP, 2023

    2023

  26. [26]

    Profd: Prompt-guided feature disentangling for occluded person re- identification,

    C. Cui, S. Huang, W. Song, P. Ding, M. Zhang, and D. Wang, “Profd: Prompt-guided feature disentangling for occluded person re- identification,” inACM MM, 2024

    2024

  27. [27]

    Shape-centered repre- sentation learning for visible–infrared person re-identification,

    S. Li, J. Leng, J. Gan, M. Mo, and X. Gao, “Shape-centered repre- sentation learning for visible–infrared person re-identification,”Pattern Recognition, 2025

    2025

  28. [28]

    Learning transferable 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 transferable visual models from natural language supervision,” inICML, 2021

    2021

  29. [29]

    Multiple information prompt learning for cloth-changing person re-identification,

    S. Wei, Z. Gao, C. Ma, Y . Zhao, W. Guan, and S. Chen, “Multiple information prompt learning for cloth-changing person re-identification,” TIP, 2025

    2025

  30. [30]

    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,”TIP, 2023

    2023

  31. [31]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  32. [32]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,”arXiv:2111.00396, 2021

    Pith/arXiv arXiv 2021

  33. [33]

    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, 2017

    2017

  34. [34]

    Omni-scale feature learning for person re-identification,

    K. Zhou, Y . Yang, A. Cavallaro, and T. Xiang, “Omni-scale feature learning for person re-identification,” inICCV, 2019

    2019

  35. [35]

    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

    2021

  36. [36]

    Identity-guided human semantic parsing for person re-identification,

    K. Zhu, H. Guo, Z. Liu, M. Tang, and J. Wang, “Identity-guided human semantic parsing for person re-identification,” inECCV, 2020

    2020

  37. [37]

    Keypoint promptable re-identification,

    V . Somers, A. Alahi, and C. D. Vleeschouwer, “Keypoint promptable re-identification,” inECCV, 2024

    2024

  38. [38]

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

    K. Niu, Y . Huang, W. Ouyang, and L. Wang, “Improving description- based person re-identification by multi-granularity image-text align- ments,”TIP, 2020

    2020

  39. [39]

    Robust multi-modality person re-identification,

    A. Zheng, Z. Wang, Z. Chen, C. Li, and J. Tang, “Robust multi-modality person re-identification,” inAAAI, no. 4, 2021

    2021

  40. [40]

    Interact, embed, and enlarge: Boosting modality-specific representations for multi-modal person re-identification,

    Z. Wang, C. Li, A. Zheng, R. He, and J. Tang, “Interact, embed, and enlarge: Boosting modality-specific representations for multi-modal person re-identification,” inAAAI, 2022

    2022

  41. [41]

    Multi-spectral vehicle re-identification: A challenge,

    H. Li, C. Li, X. Zhu, A. Zheng, and B. Luo, “Multi-spectral vehicle re-identification: A challenge,” inAAAI, 2020

    2020

  42. [42]

    Graph-based progressive fusion network for multi-modality vehicle re-identification,

    Q. He, Z. Lu, Z. Wang, and H. Hu, “Graph-based progressive fusion network for multi-modality vehicle re-identification,”TITS, 2023

    2023

  43. [43]

    Progressively hybrid transformer for multi-modal vehicle re-identification,

    W. Pan, L. Huang, J. Liang, L. Hong, and J. Zhu, “Progressively hybrid transformer for multi-modal vehicle re-identification,”Sensors, 2023

    2023

  44. [44]

    Escaping modal interactions: An efficient desanet for multi-modal object re- identification,

    W. Dong, X. Yang, D. Cheng, N. Wang, and X. Gao, “Escaping modal interactions: An efficient desanet for multi-modal object re- identification,”TIP, 2025

    2025

  45. [45]

    Visible-infrared person re-identification with modality-specific memory network,

    Y . Li, T. Zhang, X. Liu, Q. Tian, Y . Zhang, and F. Wu, “Visible-infrared person re-identification with modality-specific memory network,”TIP, 2022

    2022

  46. [46]

    Visible-infrared person re-identification via partially interactive collaboration,

    X. Zheng, X. Chen, and X. Lu, “Visible-infrared person re-identification via partially interactive collaboration,”TIP, 2022

    2022

  47. [47]

    Unicat: Crafting a stronger fusion baseline for multimodal re-identification,

    J. Crawford, H. Yin, L. McDermott, and D. Cummings, “Unicat: Crafting a stronger fusion baseline for multimodal re-identification,” arXiv:2310.18812, 2023

    Pith/arXiv arXiv 2023

  48. [48]

    Heterogeneous test-time training for multi-modal person re-identification,

    Z. Wang, H. Huang, A. Zheng, and R. He, “Heterogeneous test-time training for multi-modal person re-identification,” inAAAI, 2024

    2024

  49. [49]

    Mam- bapro: Multi-modal object re-identification with mamba aggregation and synergistic prompt,

    Y . Wang, X. Liu, T. Yan, Y . Liu, A. Zheng, P. Zhang, and H. Lu, “Mam- bapro: Multi-modal object re-identification with mamba aggregation and synergistic prompt,” inAAAI, 2025

    2025

  50. [50]

    Prompt-based modality alignment for effective multi-modal object re- identification,

    S. Zhang, W. Luo, D. Cheng, Y . Xing, G. Liang, P. Wang, and Y . Zhang, “Prompt-based modality alignment for effective multi-modal object re- identification,”TIP, 2025

    2025

  51. [51]

    Multi-modal object re- identification via sparse mixture-of-experts,

    Y . Feng, J. Li, C. Xie, L. Tan, and J. Ji, “Multi-modal object re- identification via sparse mixture-of-experts,” inICML, 2025

    2025

  52. [52]

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

    Y . Wang, Y . Liu, A. Zheng, and P. Zhang, “Decoupled feature-based mixture of experts for multi-modal object re-identification,” inAAAI, 2025

    2025

  53. [53]

    Combining recurrent, convolutional, and continuous-time models with linear state space layers,

    A. Gu, I. Johnson, K. Goel, K. Saab, T. Dao, A. Rudra, and C. R ´e, “Combining recurrent, convolutional, and continuous-time models with linear state space layers,”NeurIPS, 2021

    2021

  54. [54]

    Simplified state space layers for sequence modeling,

    J. T. Smith, A. Warrington, and S. W. Linderman, “Simplified state space layers for sequence modeling,”arXiv:2208.04933, 2022

    Pith/arXiv arXiv 2022

  55. [55]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv:2312.00752, 2023

    Pith/arXiv arXiv 2023

  56. [56]

    Sigma: Siamese mamba network for multi-modal semantic segmentation,

    Z. Wan, P. Zhang, Y . Wang, S. Yong, S. Stepputtis, K. Sycara, and Y . Xie, “Sigma: Siamese mamba network for multi-modal semantic segmentation,”arXiv:2404.04256, 2024

    Pith/arXiv arXiv 2024

  57. [57]

    Polyp-mamba: Polyp segmentation with visual mamba,

    Z. Xu, F. Tang, Z. Chen, Z. Zhou, W. Wu, Y . Yang, Y . Liang, J. Jiang, X. Cai, and J. Su, “Polyp-mamba: Polyp segmentation with visual mamba,” inMICCAI, 2024

    2024

  58. [58]

    Mamba in vision: A comprehensive survey of techniques and applications,

    M. M. Rahman, A. A. Tutul, A. Nath, L. Laishram, S. K. Jung, and T. Hammond, “Mamba in vision: A comprehensive survey of techniques and applications,”arXiv:2410.03105, 2024

    Pith/arXiv arXiv 2024

  59. [59]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,”arXiv:2401.09417, 2024

    Pith/arXiv arXiv 2024

  60. [60]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, and Y . Liu, “Vmamba: Visual state space model,”arXiv:2401.10166, 2024

    Pith/arXiv arXiv 2024

  61. [61]

    Pan-mamba: Effective pan-sharpening with state space model,

    X. He, K. Cao, J. Zhang, K. Yan, Y . Wang, R. Li, C. Xie, D. Hong, and M. Zhou, “Pan-mamba: Effective pan-sharpening with state space model,”Information Fusion, 2025

    2025

  62. [62]

    Re- mamba: a hybrid cnn-mamba aggregation network for visible-infrared person re-identification,

    H. Geng, J. Peng, W. Yang, D. Chen, H. Lv, G. Li, and Y . Shao, “Re- mamba: a hybrid cnn-mamba aggregation network for visible-infrared person re-identification,”Scientific Reports, 2024

    2024

  63. [63]

    Climb-reid: A hybrid clip-mamba framework for person re-identification,

    C. Yu, X. Liu, J. Zhu, Y . Wang, P. Zhang, and H. Lu, “Climb-reid: A hybrid clip-mamba framework for person re-identification,” inAAAI, 2025

    2025

  64. [64]

    Mambareid: Exploiting vision mamba for multi-modal object re-identification,

    R. Zhang, L. Xu, S. Yang, and L. Wang, “Mambareid: Exploiting vision mamba for multi-modal object re-identification,”Sensors, 2024

    2024

  65. [65]

    Openpifpaf: Composite fields for semantic keypoint detection and spatio-temporal association,

    S. Kreiss, L. Bertoni, and A. Alahi, “Openpifpaf: Composite fields for semantic keypoint detection and spatio-temporal association,”TITS, 2021. 16

    2021

  66. [66]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll´ar, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,”arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  67. [67]

    S 2 crossmamba: Spatial–spectral cross-mamba for multimodal remote sensing image classification,

    G. Zhang, Z. Zhang, J. Deng, L. Bian, and C. Yang, “S 2 crossmamba: Spatial–spectral cross-mamba for multimodal remote sensing image classification,”IEEE Geosci. Remote Sens. Lett., 2024

    2024

  68. [68]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv:1607.06450, 2016

    Pith/arXiv arXiv 2016

  69. [69]

    Counterfactual attention learning for fine-grained visual categorization and re-identification,

    Y . Rao, G. Chen, J. Lu, and J. Zhou, “Counterfactual attention learning for fine-grained visual categorization and re-identification,” inICCV, 2021

    2021

  70. [70]

    Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline),

    Y . Sun, L. Zheng, Y . Yang, Q. Tian, and S. Wang, “Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline),” inECCV, 2018

    2018

  71. [71]

    Dynamic en- hancement network for partial multi-modality person re-identification,

    A. Zheng, Z. He, Z. Wang, C. Li, and J. Tang, “Dynamic en- hancement network for partial multi-modality person re-identification,” arXiv:2305.15762, 2023

    Pith/arXiv arXiv 2023

  72. [72]

    Lrmm: Low rank multi-scale multi-modal fusion for person re-identification based on rgb-ni-ti,

    D. Wu, Z. Liu, Z. Chen, S. Gan, K. Tan, Q. Wan, and Y . Wang, “Lrmm: Low rank multi-scale multi-modal fusion for person re-identification based on rgb-ni-ti,”ESWA, 2025

    2025

  73. [73]

    Representation selective coupling via token sparsification for multi- spectral object re-identification,

    Z. Yu, Z. Huang, M. Hou, J. Pei, Y . Yan, Y . Liu, and D. Sun, “Representation selective coupling via token sparsification for multi- spectral object re-identification,”TCSVT, 2024

    2024

  74. [74]

    Wtsf-reid: Depth-driven window-oriented token selection and fusion for multi-modality vehicle re-identification with knowledge consistency constraint,

    Z. Yu, Z. Huang, M. Hou, Y . Yan, and Y . Liu, “Wtsf-reid: Depth-driven window-oriented token selection and fusion for multi-modality vehicle re-identification with knowledge consistency constraint,”ESWA, 2025

    2025

  75. [75]

    Idea: Inverted text with cooper- ative deformable aggregation for multi-modal object re-identification,

    Y . Wang, Y . Lv, P. Zhang, and H. Lu, “Idea: Inverted text with cooper- ative deformable aggregation for multi-modal object re-identification,” inCVPR, 2025

    2025

  76. [76]

    Rethinking the inception architecture for computer vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” inCVPR, 2016

    2016

  77. [77]

    In defense of the triplet loss for person re-identification,

    A. Hermans, L. Beyer, and B. Leibe, “In defense of the triplet loss for person re-identification,”arXiv:1703.07737, 2017

    Pith/arXiv arXiv 2017

  78. [78]

    Generative and attentive fusion for multi-spectral vehicle re-identification,

    J. Guo, X. Zhang, Z. Liu, and Y . Wang, “Generative and attentive fusion for multi-spectral vehicle re-identification,” inICSP, 2022

    2022

  79. [79]

    Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark,

    A. Zheng, X. Zhu, Z. Ma, C. Li, J. Tang, and J. Ma, “Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark,”Information Fusion, 2023

    2023

  80. [80]

    Graft: Gradual fusion transformer for multimodal re-identification,

    H. Yin, J. Li, E. Schiller, L. McDermott, and D. Cummings, “Graft: Gradual fusion transformer for multimodal re-identification,” arXiv:2310.16856, 2023

    Pith/arXiv arXiv 2023

Showing first 80 references.