Pith. sign in

REVIEW 4 major objections 5 minor 64 references

Cross-Modal Synergies: Unveiling the Potential of Motion-Aware Fusion Networks in Handling Dynamic and Static ReID Scenarios

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read MOTAR-FUSE claims that one network can extract motion cues from static images and use them to improve person re-identification across holistic, occluded, and video benchmarks.

desk verdict Plausible idea, but the central motion-consistency loss is never defined and the headline claim is undercut by the paper's own video results; worth a careful revise-and-resubmit. read the letter →

arxiv 2502.00665 v1 pith:746YWWAO submitted 2025-02-02 cs.CV

Fuxi Ling , Hongye Liu , Guoqiang Huang , Jing Li , Hong Wu , Zhihao Tang This is my paper · ORCID
classification cs.CV
keywords personre-identificationmotion-awaretransformermotionconsistencytaskoccludedvideovisualadapterstaticimageinference
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

MOTAR-FUSE is a person re-identification network that claims to pull motion cues out of static images and use them, together with appearance features, to recognize people across holistic, occluded, and video benchmarks. The authors argue that a motion-aware transformer trained with a motion consistency task on video clips learns to isolate dynamic human body parts, and that this learned motion knowledge transfers to single-image inference. If correct, the method unifies image-based and video-based ReID in one architecture: the same visual adapter accepts either a frame or a clip, and the motion module adds occlusion robustness without needing video at inference time. The paper reports the highest Rank-1 among the methods compared on Market-1501, Occluded-REID, Partial-REID, Partial-iLIDS, and MARS, while remaining close on LS-VID, iLiDS-VID, and PRID-2011.

What carries the argument

The load-bearing component is the motion-aware transformer (MAT) together with its motion consistency task. MAT uses learnable queries that cross-attend over visual tokens to isolate human body parts and refine them into motion tokens; the motion consistency task is the training signal intended to make those tokens encode dynamic motion even when the input is a single image. The paper's Eq. (1) writes the total loss as a weighted sum of a recognition loss and the motion consistency loss $L_{mc}$, but the exact form of $L_{mc}$ is not given. The companion mechanism is the dual-input visual adapter, which adjusts frame count and positional embeddings so the same network can ingest one image or a video clip, and the fusion encoder's hybrid class token is the final identity representation.

What would settle it

Run the paper's training recipe with the motion consistency term removed ($\lambda_{mc}=0$) while keeping video pre-training, then measure Rank-1 on Occluded-REID and Partial-REID; if the drop is as small as the 0.4--0.7 point gaps in Table 7, the motion signal is not doing the load-bearing work. A sharper check is to shuffle frame order during video pre-training and observe whether the motion tokens and downstream accuracy change, since true motion consistency should be sensitive to temporal order.

Watch

Extended reading notes

Core claim

On its own terms, the central claim is that motion information, usually available only from video or optical flow, can be recovered from a single still image and that this recovered motion improves person re-identification. The proposed MOTAR-FUSE network has four parts: a pre-trained ViT-B/16 visual encoder, a dual-input visual adapter that turns either one frame or a clip into visual tokens, a motion-aware transformer that converts visual tokens into motion tokens through learnable-query cross-attention, and a fusion encoder that merges visual and motion tokens into a hybrid class token used for retrieval. Training uses a weighted loss $L = \lambda_g L_g + \lambda_{mc} L_{mc}$, where $L_g$ combines cross-entropy and triplet loss and $L_{mc}$ is a motion consistency objective meant to teach the transformer to model human dynamics. The experiments are arranged to show the design transfers across three regimes: holistic image ReID, occluded and partial ReID, and video ReID.

Load-bearing premise

The mechanism depends on the assumption that the motion consistency task, trained on video clips, actually teaches the transformer to extract useful motion cues from single static images; if that transfer does not happen, the claimed occlusion robustness has no foundation.

Editorial extensions

If this is right

  • A single MOTAR-FUSE network can replace separate image- and video-ReID systems, because the same visual adapter accepts either one frame or a clip.
  • Occlusion handling no longer requires part annotations, pose estimators, or external segmentation at inference; the motion-aware transformer learns body-part structure from the motion consistency task.
  • Surveillance pipelines that only have still frames can still get motion-like cues, since inference needs only a single image.
  • Video pre-training is an accelerator rather than a hard requirement; the authors state the model can be trained from scratch with more epochs, at a small accuracy cost (Table 7).
  • Learnable-query length interacts with occluded data: query length 10 is the reported optimum, and longer queries hurt occluded and partial benchmarks more than holistic ones.

Reading between the lines

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

  • A reader cannot yet reconstruct the core training signal: Section 3.2 describes the motion consistency task only in words and Eq. (1) leaves $L_{mc}$ undefined, so the mechanism is underspecified until the loss is published.
  • If the motion-transfer result is real, the same visual-adapter-plus-motion-token design should transfer to other single-image tasks with implicit dynamics, such as gait recognition from one frame or pose-invariant action recognition; the paper does not test these.
  • The paper's own ablations (Tables 5 and 7) show the motion consistency task and video pre-training each move accuracy by less than one point, so a plausible alternative reading is that most of the gain comes from the transformer architecture and part-based queries rather than from motion as such.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes MOTAR-FUSE, a dual-input transformer architecture for person re-identification (ReID) that processes both static images and video frames. The system comprises a ViT visual encoder, a visual adapter, a motion-aware transformer with learnable queries, and a fusion encoder. The central novelty is a 'motion consistency task' that, the authors claim, enables the network to extract motion cues from static images and thereby improve occluded-person ReID. The method is evaluated on Market-1501, Occluded-REID, Partial-REID, Partial-iLIDS, MARS, LS-VID, iLiDS-VID, and PRID-2011. The paper reports state-of-the-art results on the occluded and holistic benchmarks and competitive results on the video benchmarks.

Significance. If the motion-consistency mechanism were fully specified and the experimental protocol were sound, the idea of a single network that handles both still-image and video ReID with improved occlusion robustness would be a useful contribution. The paper provides extensive benchmark comparisons and ablations, which is a strength. However, the core contribution is not verifiable as written: the motion consistency loss Lmc is never defined, the learnable query length is selected using the test benchmarks, and the video benchmark results in Table 1 do not support the abstract's blanket claim of 'superior performance'. These issues are load-bearing for the central claims of the paper.

major comments (4)
  1. [Section 3.4, Eq. (1)] The total loss is defined as L = λg Lg + λmc Lmc, but Lmc is never defined anywhere in the manuscript. Section 3.2 states that a 'motion consistency task with a custom loss function' is introduced, and Section 3.1 says training involves 'pairing images within each batch to compute motion consistency loss', yet no equation, algorithm, pair-construction rule, or supervision source for Lmc is provided. Because this loss is the paper's central mechanism for extracting motion cues from static images, the method is not implementable or falsifiable as presented. The ablation in Table 5 (Occluded-REID 86.7 vs 83.9; Partial-REID 89.2 vs 83.6) cannot be attributed to a specific component without this definition.
  2. [Section 4.5.1 and Table 6] The learnable query length is tuned directly on the test benchmarks. Table 6 reports Rank-1 on Market-1501, Occluded-REID, Partial-REID, and MARS for query lengths 2, 4, 6, 8, 10, 16, and 32, and the text concludes that length 10 is optimal because it yields the highest Rank-1 on these same datasets. No validation split or model-selection protocol is described. This is a form of test-set overfitting: the final configuration is selected to maximize the reported metric on the data used for the final numbers, so the reported results are optimistic and do not provide a valid estimate of generalization.
  3. [Section 4.3 and Table 1] The abstract claims that MOTAR-FUSE 'achieves superior performance compared to existing approaches', but on LS-VID the method trails CA ViT (Rank-1 87.6 vs 89.2) and on iLiDS-VID it trails CA ViT (Rank-1 92.4 vs 93.3). The text in Section 4.3 says the method shows 'superior performance, particularly in terms of Rank-1 accuracy and mAP' on MARS and LS-VID; on LS-VID the Rank-1 is lower than CA ViT and the mAP gain is only 0.7 points (79.9 vs 79.2). The stated performance claim is overstated relative to the evidence in the paper's own tables.
  4. [Section 4.6.1 and Table 7] The video pre-training ablation shows only small differences (0.4–0.7 percentage points on Occluded-REID and Partial-REID). This weakens the claim that motion knowledge from video pre-training is a decisive factor for occlusion robustness, especially when combined with the undefined motion consistency loss. The paper's own Section 5.2 lists the reliance on extensive pre-training as a limitation, but Table 7 suggests the model performs nearly as well without it; this tension is not resolved in the manuscript.
minor comments (5)
  1. [Figure 1] The caption reads 'Pervious SOTA' and should be 'Previous SOTA'. The numeric labels in the figure are not explained and are not directly tied to any table, which makes the figure difficult to interpret.
  2. [Section 4.4] The ablation section lists only two bullets without presenting the corresponding results; Tables 5 and 7 appear later in Section 4.5 and 4.6, which makes the narrative hard to follow. The reader should be pointed to the relevant tables at the point of discussion.
  3. [Section 3.2 and Figure 3] The module is called 'Motion-Aware Transformer' in Section 3.2, but Figure 3 uses the label 'MFA', which is not defined. Also, Figure 3's caption mentions 'Our MFA' while the rest of the paper uses 'MOTAR-FUSE'; please use consistent terminology.
  4. [Section 4.2 and 3.4] Implementation details are incomplete: for image data, the batch size and number of identities are not specified; the 'stabilization phase using video datasets' in Section 3.4 does not state which video datasets, how many epochs, or the values of the loss weights λg and λmc. These details are needed for reproducibility.
  5. [Equation (2)] The notation 'hm cls' in Eq. (2) is not defined; Figure 2 uses 'h_m_cls' and the text refers to 'hybrid class token'. Please define the symbol consistently at first use.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported benchmark results are partly forced by test-set selection of the learnable query length; the motion-consistency loss that anchors the central claim is never defined, so the core mechanism remains unverifiable.

  1. fitted input called prediction [Section 4.5 (Detailed Analysis of Learnable Query Length), Table 6, Section 4.5.1]
    "We systematically varied the length of the learnable queries in the transformer and assessed the model's performance on datasets like Market-1501, Occluded-REID, Partial-REID, and MARS. ... The optimal performance, as evidenced by the highest Rank-1 accuracy, was consistently observed at a query length of 10."

    Table 6 reports Rank-1 for query lengths 2, 4, 6, 8, 10, 16, 32 on the same benchmarks used for the final evaluation. The text declares length 10 optimal because it gives the highest Rank-1 on those datasets, and no validation split is reported. The final 'Ours' rows in Tables 1-4 reproduce the length-10 row of Table 6 exactly (e.g., Market-1501 95.7, Occluded-REID 86.7, Partial-REID 89.2, MARS 91.1). Thus the headline benchmark numbers are, by construction, the maximum of the hyperparameter grid selected on the test data; the 'prediction' of superior performance is not an independent evaluation but the argmax of the search criterion.

full rationale

The paper's central methodological claim is that a motion consistency task lets a transformer extract motion cues from static images and thereby improve ReID. That claim cannot be checked as written: Eq. (1) includes Lmc, but no definition, algorithm, or supervision source for Lmc is provided anywhere, and Eq. (2) expands only Lg. This is a severe omitted-definition and reproducibility problem, but it is not itself a circular reduction, because no equation lets one exhibit Lmc as equivalent to an input. The one concrete circular step is the query-length selection in Section 4.5: the configuration labeled optimal is chosen as the argmax of Rank-1 on the same test benchmarks used for the final comparison, so the reported 'Ours' results are, by construction, the best row of the searched grid rather than an independent prediction. The self-citation [24] (an author of the present paper) appears in Related Work in a sentence about TransReID; it is a citation error and is not load-bearing for the central argument. The paper also acknowledges reliance on video pre-training in Section 5.2, but that is a stated limitation, not circularity. Score 6 reflects partial circularity in the headline empirical results: one key reported quantity is forced by the selection procedure, while the motion mechanism itself is unverifiable but not definitionally circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central method rests on an unspecified motion-consistency loss, the transfer of video-learned motion to still images, and a query length chosen on the test benchmarks. Those are the load-bearing inputs not independently established. The pretrained ViT and published baseline results are standard inherited inputs. No new physical or external entities are introduced: motion tokens and the hybrid class token are internal learned representations.

free parameters (3)
  • learnable query length in motion-aware transformer = 10
    Table 6 sweeps lengths 2, 4, 6, 8, 10, 16, and 32 on the benchmark test sets, and 10 is retained because it gives the highest Rank-1 on the same test sets used for the final reported results.
  • loss scaling weights lambda_g and lambda_mc = not reported
    Eq. (1) defines the hybrid loss with scaling factors, but their values are never given. They are hand-set or tuned without disclosure.
  • video pre-training configuration = not reported
    Section 3.4 and Table 7 depend on a video stabilization pre-training phase, but the pre-training dataset, duration, frames per clip, and schedule are not specified.
assumptions (4)
  • domain assumption A motion consistency loss can supervise useful motion extraction from static images.
    Eq. (1) includes Lmc, but Section 3.2 never defines the loss or its supervision source. The whole method depends on this signal.
  • domain assumption Motion cues learned on video transfer to single-image inference.
    Section 3.4 trains on video clips first, then applies the model to still images. Table 7 shows only 0.4 to 0.7 point gains, so the transfer is weak but load-bearing.
  • domain assumption Benchmark numbers for compared methods are accurate and evaluation protocols are aligned.
    The comparison tables rely entirely on published results without re-running baselines. Any protocol mismatch changes the ranking.
  • domain assumption A pretrained ViT-B/16 backbone provides suitable visual features.
    Section 3.1 adopts ImageNet-pretrained ViT-B/16 without questioning its suitability for surveillance imagery with occlusions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Modal Synergies: Unveiling the Potential of Motion-Aware Fusion Networks in Handling Dynamic and Static ReID Scenarios." pith.science (2026). https://pith.science/paper/746YWWAO

@misc{pith2026250200665,
  author       = {Pith},
  title        = {Pith review of: Cross-Modal Synergies: Unveiling the Potential of Motion-Aware Fusion Networks in Handling Dynamic and Static ReID Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/746YWWAO}},
  note         = {Machine review of arXiv:2502.00665}
}
read the original abstract

Navigating the complexities of person re-identification (ReID) in varied surveillance scenarios, particularly when occlusions occur, poses significant challenges. We introduce an innovative Motion-Aware Fusion (MOTAR-FUSE) network that utilizes motion cues derived from static imagery to significantly enhance ReID capabilities. This network incorporates a dual-input visual adapter capable of processing both images and videos, thereby facilitating more effective feature extraction. A unique aspect of our approach is the integration of a motion consistency task, which empowers the motion-aware transformer to adeptly capture the dynamics of human motion. This technique substantially improves the recognition of features in scenarios where occlusions are prevalent, thereby advancing the ReID process. Our comprehensive evaluations across multiple ReID benchmarks, including holistic, occluded, and video-based scenarios, demonstrate that our MOTAR-FUSE network achieves superior performance compared to existing approaches.

Figures

Figures reproduced from arXiv: 2502.00665 by the authors.

Figure 1
Figure 1. Graphical depiction of MOTAR-FUSE’s comparative [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed MOTAR-FUSE module. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of human part segmentation under different conditions: (a) shows enhanced focus on the human body in a cluttered [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 51 canonical work pages

  1. [1]

    Vivit: A video vi- sion transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vi- sion transformer. In ICCV, 2021. 1

    work page 2021

  2. [2]

    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. 4

    work page

  3. [3]

    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, 2022. 1, 2

    work page 2022

  4. [4]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3

    arXiv 2010

  5. [5]

    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, 2021. 4

    work page 2021

  6. [6]

    Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding

    Hao Feng, Qi Liu, Hao Liu, Jingqun Tang, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding. Science China Infor- mation Sciences, 67(12):1–14, 2024. 2

    2024

  7. [7]

    Unidoc: A uni- versal large multimodal model for simultaneous text detec- tion, recognition, spotting and understanding

    Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wen- gang Zhou, Houqiang Li, and Can Huang. Unidoc: A uni- versal large multimodal model for simultaneous text detec- tion, recognition, spotting and understanding. arXiv preprint arXiv:2308.11592, 2023. 2

    arXiv 2023

  8. [8]

    Pose- guided visible part matching for occluded person reid

    Shang Gao, Jingya Wang, Huchuan Lu, and Zimo Liu. Pose- guided visible part matching for occluded person reid. In CVPR, 2020. 2, 5

    work page 2020

Show all 64 references
  1. [9]

    Fd-gan: Pose-guided feature distilling gan for robust person re-identification

    Yixiao Ge, Zhuowan Li, Haiyu Zhao, Guojun Yin, Shuai Yi, Xiaogang Wang, and Hongsheng Li. Fd-gan: Pose-guided feature distilling gan for robust person re-identification. arXiv preprint arXiv:1810.02936, 2018. 1

    2018 arXiv

  2. [10]

    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, 2020. 1, 4

    2020

  3. [11]

    Recognizing partial biometric patterns

    Lingxiao He, Zhenan Sun, Yuhao Zhu, and Yunbo Wang. Recognizing partial biometric patterns. arXiv preprint arXiv:1810.07399, 2018. 5

    2018 arXiv

  4. [12]

    Foreground-aware pyramid reconstruc- tion for alignment-free occluded person re-identification

    Lingxiao He, Yinggang Wang, Wu Liu, He Zhao, Zhenan Sun, and Jiashi Feng. Foreground-aware pyramid reconstruc- tion for alignment-free occluded person re-identification. In ICCV, 2019. 5

    2019

  5. [13]

    Transreid: Transformer-based object re- identification

    Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re- identification. In CVPR, 2021. 2, 5

    2021

  6. [14]

    Person re-identification by descriptive and discrim- inative classification

    Martin Hirzer, Csaba Beleznai, Peter M Roth, and Horst Bischof. Person re-identification by descriptive and discrim- inative classification. In SCIA, 2011. 4

    2011

  7. [15]

    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, 2021. 2

    2021

  8. [16]

    Interaction-and-aggregation network for person re-identification

    Ruibing Hou, Bingpeng Ma, Hong Chang, Xinqian Gu, Shiguang Shan, and Xilin Chen. Interaction-and-aggregation network for person re-identification. In CVPR, pages 9317– 9326, 2019. 1

    2019

  9. [17]

    Unsupervised learning of object landmarks through conditional image generation

    Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. NIPS, 2018. 2

    2018

  10. [18]

    Flow guided mutual at- tention for person re-identification

    Madhu Kiran, Amran Bhuiyan, Louis-Antoine Blais-Morin, Ismail Ben Ayed, Eric Granger, et al. Flow guided mutual at- tention for person re-identification. Image and Vision Com- puting, 2021. 2

    2021

  11. [19]

    Global-local temporal representations for video per- son re-identification

    Jianing Li, Jingdong Wang, Qi Tian, Wen Gao, and Shiliang Zhang. Global-local temporal representations for video per- son re-identification. In ICCV, 2019. 4

    2019

  12. [20]

    Global-local temporal representations for video per- son re-identification

    Jianing Li, Jingdong Wang, Qi Tian, Wen Gao, and Shiliang Zhang. Global-local temporal representations for video per- son re-identification. In CVPR, 2019. 4

    2019

  13. [21]

    Multi- scale 3d convolution network for video based person re- identification

    Jianing Li, Shiliang Zhang, and Tiejun Huang. Multi- scale 3d convolution network for video based person re- identification. In AAAI, 2019. 4

    2019

  14. [22]

    Diverse part discovery: Occluded per- son re-identification with part-aware transformer

    Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded per- son re-identification with part-aware transformer. In CVPR,

  15. [23]

    Per- son re-identification by local maximal occurrence represen- tation and metric learning

    Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z Li. Per- son re-identification by local maximal occurrence represen- tation and metric learning. In CVPR, 2015. 1

    2015

  16. [24]

    Rethink motion information for occluded person re-identification

    Hongye Liu and Xiai Chen. Rethink motion information for occluded person re-identification. Applied Sciences , 14(6):2558, 2024. 2

    2024

  17. [25]

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

    Xuehu Liu, Pingping Zhang, Chenyang Yu, Huchuan Lu, and Xiaoyun Yang. Watching you: Global-guided reciprocal learning for video-based person re-identification. In CVPR,

  18. [26]

    Spts v2: single-point scene text spotting

    Yuliang Liu, Jiaxin Zhang, Dezhi Peng, Mingxin Huang, Xinyu Wang, Jingqun Tang, Can Huang, Dahua Lin, Chun- hua Shen, Xiang Bai, et al. Spts v2: single-point scene text spotting. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 2023. 1

    2023

  19. [27]

    A bounding box is worth one token: Interleav- ing layout and text in a large language model for document understanding

    Jinghui Lu, Haiyang Yu, Yanjie Wang, Yongjie Ye, Jingqun Tang, Ziwei Yang, Binghong Wu, Qi Liu, Hao Feng, Han Wang, et al. A bounding box is worth one token: Interleav- ing layout and text in a large language model for document understanding. arXiv preprint arXiv:2407.01976, 2024. 2

    2024 arXiv

  20. [28]

    Pose-guided feature alignment for occluded person re-identification

    Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In ICCV, 2019. 5

    2019

  21. [29]

    Mct- bench: Multimodal cognition towards text-rich visual scenes benchmark

    Bin Shan, Xiang Fei, Wei Shi, An-Lan Wang, Guozhi Tang, Lei Liao, Jingqun Tang, Xiang Bai, and Can Huang. Mct- bench: Multimodal cognition towards text-rich visual scenes benchmark. arXiv preprint arXiv:2410.11538, 2024. 2

    2024 arXiv

  22. [30]

    Per- son re-identification with a locally aware transformer

    Charu Sharma, Siddhant R Kapil, and David Chapman. Per- son re-identification with a locally aware transformer. arXiv preprint arXiv:2106.03720, 2021. 1

    2021 arXiv

  23. [31]

    First order motion model for image animation

    Aliaksandr Siarohin, St ´ephane Lathuili`ere, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. NIPS, 2019. 2 9

    2019

  24. [32]

    Motion- supervised co-part segmentation

    Aliaksandr Siarohin, Subhankar Roy, St ´ephane Lathuili `ere, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Motion- supervised co-part segmentation. In ICPR, 2021. 2

    2021

  25. [33]

    Attentive eraser: Unleashing diffusion model’s object re- moval potential via self-attention redirection guidance.arXiv preprint arXiv:2412.12974, 2024

    Wenhao Sun, Benlei Cui, Jingqun Tang, and Xue-Mei Dong. Attentive eraser: Unleashing diffusion model’s object re- moval potential via self-attention redirection guidance.arXiv preprint arXiv:2412.12974, 2024. 2

    2024

  26. [34]

    Perceive where to focus: Learn- ing visibility-aware part-level features for partial person re- identification

    Yifan Sun, Qin Xu, Yali Li, Chi Zhang, Yikang Li, Shengjin Wang, and Jian Sun. Perceive where to focus: Learn- ing visibility-aware part-level features for partial person re- identification. In CVPR, 2019. 5

    2019

  27. [35]

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

    Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). InECCV,

  28. [36]

    Dy- namic prototype mask for occluded person re-identification

    Lei Tan, Pingyang Dai, Rongrong Ji, and Yongjian Wu. Dy- namic prototype mask for occluded person re-identification. In ACM MM, 2022. 2, 5

    2022

  29. [37]

    Character recognition competition for street view shop signs

    Jingqun Tang, Weidong Du, Bin Wang, Wenyang Zhou, Shuqi Mei, Tao Xue, Xing Xu, and Hai Zhang. Character recognition competition for street view shop signs. National Science Review, 10(6):nwad141, 2023. 1

    2023

  30. [38]

    Textsquare: Scaling up text-centric visual instruction tuning

    Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, et al. Textsquare: Scaling up text-centric visual instruction tuning. arXiv preprint arXiv:2404.12803, 2024. 2

    2024 arXiv

  31. [39]

    Mtvqa: Benchmarking multilingual text-centric visual question answering

    Jingqun Tang, Qi Liu, Yongjie Ye, Jinghui Lu, Shu Wei, Chunhui Lin, Wanqing Li, Mohamad Fitri Faiz Bin Mah- mood, Hao Feng, Zhen Zhao, et al. Mtvqa: Benchmarking multilingual text-centric visual question answering. arXiv preprint arXiv:2405.11985, 2024. 2

    2024 arXiv

  32. [40]

    Optimal boxes: boosting end-to- end scene text recognition by adjusting annotated bounding boxes via reinforcement learning

    Jingqun Tang, Wenming Qian, Luchuan Song, Xiena Dong, Lan Li, and Xiang Bai. Optimal boxes: boosting end-to- end scene text recognition by adjusting annotated bounding boxes via reinforcement learning. In European Conference on Computer Vision, pages 233–248. Springer, 2022. 2

    2022

  33. [41]

    You can even annotate text with voice: Transcription-only-supervised text spotting

    Jingqun Tang, Su Qiao, Benlei Cui, Yuhang Ma, Sheng Zhang, and Dimitrios Kanoulas. You can even annotate text with voice: Transcription-only-supervised text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, MM ’22, page 4154–4163, New York, NY , ...

    2022

  34. [42]

    Few could be better than all: Feature sampling and grouping for scene text detection

    Jingqun Tang, Wenqing Zhang, Hongye Liu, MingKun Yang, Bo Jiang, Guanglong Hu, and Xiang Bai. Few could be better than all: Feature sampling and grouping for scene text detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4563...

    2022

  35. [43]

    Pargo: Bridging vision-language with partial and global views

    An-Lan Wang, Bin Shan, Wei Shi, Kun-Yu Lin, Xiang Fei, Guozhi Tang, Lei Liao, Jingqun Tang, Can Huang, and Wei- Shi Zheng. Pargo: Bridging vision-language with partial and global views. arXiv preprint arXiv:2408.12928, 2024. 2

    2024 arXiv

  36. [44]

    High-order information matters: Learning relation and topology for occluded person re-identification

    Guan’an Wang, Shuo Yang, Huanyu Liu, Zhicheng Wang, Yang Yang, Shuliang Wang, Gang Yu, Erjin Zhou, and Jian Sun. High-order information matters: Learning relation and topology for occluded person re-identification. In CVPR,

  37. [45]

    Person re-identification by video ranking

    Taiqing Wang, Shaogang Gong, Xiatian Zhu, and Shengjin Wang. Person re-identification by video ranking. In ECCV,

  38. [46]

    Feature erasing and diffusion net- work for occluded person re-identification

    Zhikang Wang, Feng Zhu, Shixiang Tang, Rui Zhao, Lihuo He, and Jiangning Song. Feature erasing and diffusion net- work for occluded person re-identification. In CVPR, 2022. 1, 2, 5

    2022

  39. [47]

    Cavit: Contextual alignment vi- sion transformer for video object re-identification

    Jinlin Wu, Lingxiao He, Wu Liu, Yang Yang, Zhen Lei, Tao Mei, and Stan Z Li. Cavit: Contextual alignment vi- sion transformer for video object re-identification. In ECCV,

  40. [48]

    Learning multi-granular hypergraphs for video-based person re-identification

    Yichao Yan, Jie Qin, Jiaxin Chen, Li Liu, Fan Zhu, Ying Tai, and Ling Shao. Learning multi-granular hypergraphs for video-based person re-identification. In CVPR, 2020. 4

    2020

  41. [49]

    Salient color names for person re- identification

    Yang Yang, Jimei Yang, Junjie Yan, Shengcai Liao, Dong Yi, and Stan Z Li. Salient color names for person re- identification. In ECCV, 2014. 1

    2014

  42. [50]

    Fine-grained person re-identification

    Jiahang Yin, Ancong Wu, and Wei-Shi Zheng. Fine-grained person re-identification. IJCV, 2020. 2

    2020

  43. [51]

    Learning multi-task correlation particle filters for visual tracking

    Tianzhu Zhang, Changsheng Xu, and Ming-Hsuan Yang. Learning multi-task correlation particle filters for visual tracking. PAMI, 2018. 1

    2018

  44. [52]

    Robust structural sparse tracking

    Tianzhu Zhang, Changsheng Xu, and Ming-Hsuan Yang. Robust structural sparse tracking. PAMI, 2018. 1

    2018

  45. [53]

    Multi-granularity reference-aided attentive feature ag- gregation for video-based person re-identification

    Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Multi-granularity reference-aided attentive feature ag- gregation for video-based person re-identification. In CVPR,

  46. [54]

    Tabpedia: Towards comprehensive visual ta- ble understanding with concept synergy

    Weichao Zhao, Hao Feng, Qi Liu, Jingqun Tang, Shu Wei, Binghong Wu, Lei Liao, Yongjie Ye, Hao Liu, Wengang Zhou, et al. Tabpedia: Towards comprehensive visual ta- ble understanding with concept synergy. arXiv preprint arXiv:2406.01326, 2024. 2

    2024 arXiv

  47. [55]

    Multi-modal in-context learning makes an ego-evolving scene text recognizer

    Zhen Zhao, Jingqun Tang, Chunhui Lin, Binghong Wu, Can Huang, Hao Liu, Xin Tan, Zhizhong Zhang, and Yuan Xie. Multi-modal in-context learning makes an ego-evolving scene text recognizer. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pa...

    2024

  48. [56]

    Harmonizing visual text comprehension and gen- eration

    Zhen Zhao, Jingqun Tang, Binghong Wu, Chunhui Lin, Shu Wei, Hao Liu, Xin Tan, Zhizhong Zhang, Can Huang, and Yuan Xie. Harmonizing visual text comprehension and gen- eration. arXiv preprint arXiv:2407.16364, 2024. 1

    2024 arXiv

  49. [57]

    Mars: A video benchmark for large-scale person re-identification

    Liang Zheng, Zhi Bie, Yifan Sun, Jingdong Wang, Chi Su, Shengjin Wang, and Qi Tian. Mars: A video benchmark for large-scale person re-identification. In ECCV, 2016. 4

    2016

  50. [58]

    Pose- invariant embedding for deep person re-identification

    Liang Zheng, Yujia Huang, Huchuan Lu, and Yi Yang. Pose- invariant embedding for deep person re-identification. TIP,

  51. [59]

    Scalable person re-identification: A benchmark

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jing- dong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, 2015. 3

    2015

  52. [60]

    Person re- identification by probabilistic relative distance comparison

    Wei-Shi Zheng, Shaogang Gong, and Tao Xiang. Person re- identification by probabilistic relative distance comparison. In CVPR, 2011. 3

    2011

  53. [61]

    Reidentifi- cation by relative distance comparison

    Wei-Shi Zheng, Shaogang Gong, and Tao Xiang. Reidentifi- cation by relative distance comparison. PAMI, 2012. 1 10

    2012

  54. [62]

    Partial person re- identification

    Wei-Shi Zheng, Xiang Li, Tao Xiang, Shengcai Liao, Jianhuang Lai, and Shaogang Gong. Partial person re- identification. In ICCV, 2015. 3, 5

    2015

  55. [63]

    Identity-guided human semantic parsing for person re-identification

    Kuan Zhu, Haiyun Guo, Zhiwei Liu, Ming Tang, and Jinqiao Wang. Identity-guided human semantic parsing for person re-identification. In ECCV, 2020. 5

    2020

  56. [64]

    Occluded person re-identification

    Jiaxuan Zhuo, Zeyu Chen, Jianhuang Lai, and Guangcong Wang. Occluded person re-identification. In ICME, 2018. 4 11

    2018

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.