Pith. sign in

REVIEW 5 major objections 4 minor 40 references

Enhancing Vision-Language Tracking by Effectively Converting Textual Cues into Visual Cues

T0 review · 5 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Converting text to heatmaps gives a visual tracker state-of-the-art vision-language tracking, with up to 8.2% AUC gain on MGIT.

desk verdict A genuinely novel plug-and-play mechanism—frozen grounding model converts text into a heatmap for a visual tracker—but the paper never shows that this beats the visual-only base tracker, so the headline gains are unattributed. read the letter →

arxiv 2412.19648 v1 pith:MVD7QCVA submitted 2024-12-27 cs.CV cs.MM

X. Feng , D. Zhang , S. Hu , X. Li , M. Wu , J. Zhang , X. Chen , K. Huang This is my paper · ORCID
classification cs.CVcs.MM
keywords vision-languagetrackingtext-to-heatmapconversionmultimodalfusionvisualobjectfoundationmodelsplug-and-playtrackertargetdistributionheatmap
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

Vision-language tracking is hampered by a data imbalance: existing training sets contain far fewer text annotations than video frames, so trackers never learn to align language with image content. This paper argues that the problem can be circumvented rather than cured by outsourcing text-image alignment to a frozen grounding model. The proposed plug-and-play method, CTVLT, converts the language description into a spatial heatmap marking where the text points, then fuses that heatmap into the visual tracker's features. On the MGIT, TNL2K, and LaSOT benchmarks, this lifts a standard visual tracker to state-of-the-art accuracy, with the largest reported gain on MGIT.

What carries the argument

The central mechanism is the textual cue mapping module, which computes a dot-product attention map between the shallowest-scale image features and the text features from a frozen grounding model, then refines that map with a self-attention-like correlation operation: $f^{sa}_{x1} = f^G_{x1} (f^G_{x1})^T f^G_{x1}$, followed by a second dot product with the text features to produce a single-channel target heatmap $H_l$. The heatmap guidance module then encodes $H_l$ with a small CNN, concatenates it with the tracker's search feature map along the channel dimension, and processes the stack with a second CNN to return fused search tokens of the same dimension, so the prediction module is unchanged.

What would settle it

Compute a quantitative heatmap-localization metric, such as the distance from the heatmap's peak to the ground-truth target center or the overlap between a thresholded heatmap and the ground-truth box, over a large random sample of frames from MGIT, TNL2K, and LaSOT; if the heatmap's peak falls outside the target region on a substantial fraction of frames (for example, more than 20 percent), the core assumption that the heatmap reliably localizes the textual reference would be contradicted, and the reported gains would need another explanation.

Watch

Extended reading notes

Core claim

The paper's central claim is that a vision-language tracker does not need to learn text-image alignment from scarce tracking data. By inserting a frozen foundation grounding model between the tracker's encoder and prediction module, the textual cue is transformed into a target-distribution heatmap through a training-free dot-product attention between the grounding model's shallow-scale image features and the text features, refined by a self-attention-like correlation operation. The heatmap is then encoded by a lightweight CNN, concatenated with the tracker's search feature map, and processed by a second CNN to produce fused search tokens of the same dimension, leaving the tracker's prediction module unchanged. The authors report state-of-the-art results on three benchmarks and an ablation showing that the refined heatmap outperforms both direct textual features and the naive attention map.

Load-bearing premise

The method assumes that a dot-product attention between the grounding model's shallow-scale image features and the text features yields a heatmap that consistently pinpoints the text-described target, so that the fused heatmap guides rather than misleads the tracker.

Editorial extensions

If this is right

  • A standard visual tracker can be turned into a vision-language tracker by appending this plug-and-play module, without modifying its prediction head.
  • The performance gap between dedicated vision-language trackers and converted visual trackers closes on MGIT, TNL2K, and LaSOT, with the largest gains on MGIT (8.2% AUC).
  • Textual cues become interpretable visual guidance, which could make tracking failures easier to diagnose by inspecting the heatmap.
  • Because the grounding model is frozen and the mapping module is training-free, the method can attach to any transformer-based tracker with token-based search features.
  • The data imbalance problem in vision-language tracking is sidestepped by importing pre-trained text-image alignment rather than learning it from scarce tracking data.

Reading between the lines

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

  • Beyond the paper's claims, the same text-to-heatmap conversion could be applied to other open-vocabulary tracking settings, such as descriptions that change mid-sequence or user-provided referential expressions.
  • The paper does not quantify heatmap quality; a systematic study of how often the heatmap peak falls on the target could separate the method's true contribution from the tracker's inherent visual strength.
  • The conversion strategy may transfer to related tasks like visual grounding and referring segmentation, where a frozen grounding model could supply a spatial prior to a lightweight head.
  • Since the grounding model runs on every frame, an asynchronous inference or distillation of the heatmap into the tracker could recover speed while preserving the reported gains.
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

5 major / 4 minor

Summary. The paper proposes CTVLT, a plug-and-play module that converts language descriptions into target-distribution heatmaps using a frozen Grounding DINO encoder, then fuses these heatmaps into the search features of a visual tracker (AQATrack) through two small CNN modules. The textual cue mapping module is training-free, while the heatmap guidance module is trainable. The method is evaluated on MGIT, TNL2K, and LaSOT, where it reports state-of-the-art AUC and precision scores, and an ablation compares direct textual feature fusion, a naive attention map, and the proposed refined heatmap.

Significance. If the reported gains are reproducible and correctly attributed, the text-to-heatmap conversion idea is a practically useful way to inject language guidance into existing visual trackers without retraining their encoders. The training-free mapping module and the frozen grounding encoder make the approach relatively lightweight to adapt, and the released code is a concrete asset for the community. However, the current experimental protocol does not yet establish that the gains come from the proposed conversion mechanism: there is no visual-only baseline, the training data are unspecified, the scale choice is justified only qualitatively, and the ablation lacks a no-text control. The central claim is therefore plausible but not yet fully supported.

major comments (5)
  1. [Section III-A and Table I] The comparison cannot attribute the reported gains to the proposed text-to-heatmap conversion because the visual-only AQATrack baseline is never reported. The paper only compares against other vision-language trackers; there is no row showing AQATrack without the text branch on MGIT, TNL2K, or LaSOT. Without this control, the +8.2% AUC gain over JointNLT on MGIT could come from the stronger visual architecture of AQATrack rather than from the textual cue mapping mechanism. Please add the visual-only baseline and report the delta attributable to the text branch.
  2. [Section III-A] The training protocol is under-specified: the text says '30,000 samples randomly selected in each epoch' but does not state which datasets, which splits, or whether the evaluation benchmark videos are included in training. If the 30k samples are drawn from the training splits of MGIT, TNL2K, or LaSOT, the comparison against methods trained on different or smaller data is biased. Specify the exact training corpora and splits, and if possible evaluate under a protocol matched to prior methods.
  3. [Section II-B, Eqs. (2)-(4), and Fig. 3] The choice of the shallow attention map A'_1 is justified only by qualitative inspection of a few examples and the statement that the pattern is consistent across many samples. Since the entire method depends on the heatmap localizing the text-described target, provide a quantitative heatmap-quality evaluation (e.g., localization error or IoU of the heatmap peak against the target box) and an ablation over the four scales k=1..4. Also discuss failure cases where the heatmap is wrong and how the fusion module behaves in those cases.
  4. [Section III-C, Table II] The ablation study has no no-text control. Rows #1, #2, and #3 all inject some form of textual information into the tracker, so the table only compares different ways of using text. It does not demonstrate that converting text into heatmaps is better than using no textual guidance at all. Add a visual-only AQATrack row and, ideally, a row with an uninformative or random heatmap to test whether the benefit comes from the heatmap content or merely from the additional trainable parameters.
  5. [Section III-A and Table I] No error bars or multiple-run statistics are reported for the main results. Because each epoch randomly subsamples 30,000 samples and training runs for 80 epochs, single-run differences of about one AUC point (for example, 62.2 vs. 59.3 on TNL2K) may be within run-to-run variance. Report the mean and standard deviation over at least three random seeds, and state how many runs the reported numbers correspond to.
minor comments (4)
  1. [Section I] The text mentions 'VLT TT' where the reference list and Table I use 'VLTTT' [33]; please correct the typo for consistency.
  2. [Section II-B] The notation for the refined map is inconsistent: the paper uses Asa_1, Asa_1', and H_l without clearly defining the reshaping and normalization steps. Please define all dimensions and the exact relationship between these symbols.
  3. [Fig. 3] The attention maps in Fig. 3 would be much easier to evaluate if a colorbar or explicit normalization description were provided; the claim that deeper scales have less correspondence with the target is hard to verify from the current grayscale panels.
  4. [Table I caption] The caption says the best two results are highlighted in red and blue, but it does not explain which color corresponds to the best and which to the second best; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CTVLT text-to-heatmap conversion is an explicit construction from an external frozen grounding model, benchmark evaluation is external, and the self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained. The textual-cue-to-heatmap mapping is defined by explicit equations (Eqs. 1-4) operating on features from Grounding DINO, an externally trained foundation model, and the textual cue mapping module is stated to contain no trainable parameters, so no fitted constant is later renamed as a prediction. The learned components (CNN1/CNN2 in the heatmap guidance module) are evaluated against non-overlapping prior methods on external benchmarks (TNL2K, LaSOT) as well as the authors' MGIT, and the reported gains are not forced by the construction because the training objective does not directly optimize the metrics being reported. Author self-citations appear (MGIT, MemVLT, Ref. [26]) but the central alignment capability is imported from an outside model, not from those citations, so none is load-bearing. The missing visual-only baseline in Table II and the unspecified composition of the 30,000 training samples are real attribution and reproducibility concerns, but they concern experimental isolation rather than a reduction of an output to an input by definition; they are therefore outside the circularity patterns defined here and do not raise the circularity score.

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

The paper introduces no new physical entities. The method's central claim rests on the trustworthiness of Grounding DINO's cross-modal features, a hand-selected feature scale, and an unspecified training setup. The learned CNN weights are standard. The main free parameters are design choices made without ablations.

free parameters (3)
  • attention scale index k = 1
    Chosen based on qualitative analysis of attention maps across scales (Sec. II-B, Fig. 3); no quantitative criterion is reported.
  • heatmap embedding channels = 64
    Design choice for CNN1 output dimension, set without ablation study (Sec. II-C).
  • number of CNN layers in CNN1/CNN2 = 3 each
    Architecture choice stated in Sec. III-A without supporting ablation.
assumptions (4)
  • domain assumption Grounding DINO's aligned text-image features provide reliable spatial localization of the text-described target in search images.
    Central premise of the textual cue mapping module (Sec. II-B); no quantitative evaluation of heatmap quality is provided.
  • domain assumption Shallow image features from Grounding DINO preserve spatial detail better than deeper features for this task.
    Stated from qualitative analysis in Sec. II-B and Fig. 3, not statistically validated.
  • domain assumption The unspecified training dataset(s) provide sufficient supervision to train CNN1 and CNN2 without degrading the frozen tracker.
    Training data is not specified (Sec. III-A), so this assumption is unverifiable from the paper.
  • domain assumption AQATrack is a suitable base tracker whose prediction module can consume the modified search features without architectural changes.
    The plug-and-play claim depends on this compatibility (Sec. II-C); no experiment isolates the base tracker's own performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Vision-Language Tracking by Effectively Converting Textual Cues into Visual Cues." pith.science (2026). https://pith.science/paper/MVD7QCVA

@misc{pith2026241219648,
  author       = {Pith},
  title        = {Pith review of: Enhancing Vision-Language Tracking by Effectively Converting Textual Cues into Visual Cues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MVD7QCVA}},
  note         = {Machine review of arXiv:2412.19648}
}
read the original abstract

Vision-Language Tracking (VLT) aims to localize a target in video sequences using a visual template and language description. While textual cues enhance tracking potential, current datasets typically contain much more image data than text, limiting the ability of VLT methods to align the two modalities effectively. To address this imbalance, we propose a novel plug-and-play method named CTVLT that leverages the strong text-image alignment capabilities of foundation grounding models. CTVLT converts textual cues into interpretable visual heatmaps, which are easier for trackers to process. Specifically, we design a textual cue mapping module that transforms textual cues into target distribution heatmaps, visually representing the location described by the text. Additionally, the heatmap guidance module fuses these heatmaps with the search image to guide tracking more effectively. Extensive experiments on mainstream benchmarks demonstrate the effectiveness of our approach, achieving state-of-the-art performance and validating the utility of our method for enhanced VLT.

Figures

Figures reproduced from arXiv: 2412.19648 by the authors.

Figure 1
Figure 1. Schematic diagram of motivation and method paradigm inno￾vation. (a): Comparison of training environments between vision-language trackers and foundation grounding models. (b): The severe scarcity of textual data limits the tracker’s ability to understand text, making direct use of textual cues for guidance challenging. (c): Our core insight is to leverage the strong text-image alignment capabilities of foundation g… view at source ↗
Figure 2
Figure 2. Framework of our vision-language tracker (CTVLT). Our pro￾posed textual cue utilization method consists of the encoder module from a foundation grounding model, along with our designed textual cue mapping module and heatmap guidance module. As a plug-and-play method, it can be seamlessly integrated between the encoder and prediction modules of a visual tracker, transforming it into a vision-language tracker. BERT [1… view at source ↗
Figure 3
Figure 3. Visualization of different attention maps for search features with respect to textual features. (a): Given search image and textual cue. (b)- (e): Attention maps at different scales obtained through a naive process. (f): Refined result after applying our proposed method. our core insight is to convert textual cues into visual cues that the tracker can easily understand, in the form of a target spatial distribution h… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 31 canonical work pages

  1. [1]

    Online object tracking: A benchmark,

    Y . Wu, J. Lim, and M.-H. Yang, “Online object tracking: A benchmark,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2013, pp. 2411–2418

    work page 2013

  2. [2]

    Sotverse: A user-defined task space of single object tracking,

    S. Hu, X. Zhao, and K. Huang, “Sotverse: A user-defined task space of single object tracking,” International Journal of Computer Vision , vol. 132, p. 872–930, 2024

    work page 2024

  3. [3]

    Global instance tracking: Locating target more like humans,

    S. Hu, X. Zhao, L. Huang, and K. Huang, “Global instance tracking: Locating target more like humans,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 1, pp. 576–592, 2022

    work page 2022

  4. [4]

    Biodrone: A bionic drone-based single object tracking benchmark for robust vision,

    X. Zhao, S. Hu, Y . Wang, J. Zhang, Y . Hu, R. Liu, H. Ling, Y . Li, R. Li, K. Liu, and J. Li, “Biodrone: A bionic drone-based single object tracking benchmark for robust vision,” International Journal of Computer Vision, vol. 132, p. 1659–1684, 2024

    work page 2024

  5. [5]

    Tracking by natural language specification,

    Z. Li, R. Tao, E. Gavves, C. G. Snoek, and A. W. Smeulders, “Tracking by natural language specification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2017, pp. 6495–6503

    work page 2017

  6. [6]

    A multi-modal global instance tracking benchmark (mgit): Better locating target in complex spatio-temporal and causal relationship,

    S. Hu, D. Zhang, M. Wu, X. Feng, X. Li, X. Zhao, and K. Huang, “A multi-modal global instance tracking benchmark (mgit): Better locating target in complex spatio-temporal and causal relationship,” in Thirty- seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023

    work page 2023

  7. [7]

    Dtllm-vlt: Diverse text generation for visual language tracking based on llm,

    X. Li, X. Feng, S. Hu, M. Wu, D. Zhang, J. Zhang, and K. Huang, “Dtllm-vlt: Diverse text generation for visual language tracking based on llm,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 7283–7292

    2024

  8. [8]

    How texts help? a fine-grained evaluation to reveal the role of language in vision-language tracking,

    X. Li, S. Hu, X. Feng, D. Zhang, M. Wu, J. Zhang, and K. Huang, “How texts help? a fine-grained evaluation to reveal the role of language in vision-language tracking,” arXiv preprint arXiv:2411.15600 , 2024

    arXiv 2024

Show all 40 references
  1. [9]

    Memvlt: Vision-language tracking with adaptive memory- based prompts,

    X. Feng, X. Li, S. Hu, D. Zhang, M. Wu, J. Zhang, X. Chen, and K. Huang, “Memvlt: Vision-language tracking with adaptive memory- based prompts,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2024

    2024

  2. [10]

    Transvg: End-to-end visual grounding with transformers,

    J. Deng, Z. Yang, T. Chen, W. Zhou, and H. Li, “Transvg: End-to-end visual grounding with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 1769–1779

    2021

  3. [11]

    Grounding dino: Marrying dino with grounded pre- training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhu et al., “Grounding dino: Marrying dino with grounded pre- training for open-set object detection,” arXiv preprint arXiv:2303.05499, 2023

    2023 arXiv

  4. [12]

    Grounded language-image pre- training,

    L. H. Li, P. Zhang, H. Zhang, J. Yang, C. Li, Y . Zhong, L. Wang, L. Yuan, L. Zhang, J.-N. Hwang et al., “Grounded language-image pre- training,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 10 965–10 975

    2022

  5. [13]

    Towards more flexible and accurate object tracking with natural lan- guage: Algorithms and benchmark,

    X. Wang, X. Shu, Z. Zhang, B. Jiang, Y . Wang, Y . Tian, and F. Wu, “Towards more flexible and accurate object tracking with natural lan- guage: Algorithms and benchmark,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 13 763–13 773

    2021

  6. [14]

    Lasot: A high-quality benchmark for large-scale single object tracking,

    H. Fan, L. Lin, F. Yang, P. Chu, G. Deng, S. Yu, H. Bai, Y . Xu, C. Liao, and H. Ling, “Lasot: A high-quality benchmark for large-scale single object tracking,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 5369–5378

    2019

  7. [15]

    Siamese natural lan- guage tracker: Tracking by natural language descriptions with siamese trackers,

    Q. Feng, V . Ablavsky, Q. Bai, and S. Sclaroff, “Siamese natural lan- guage tracker: Tracking by natural language descriptions with siamese trackers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 5847–5856

    2021

  8. [16]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  9. [17]

    All in one: Exploring unified vision-language tracking with multi-modal alignment,

    C. Zhang, X. Sun, Y . Yang, L. Liu, Q. Liu, X. Zhou, and Y . Wang, “All in one: Exploring unified vision-language tracking with multi-modal alignment,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 5552–5561

    2023

  10. [18]

    Context- aware integration of language and visual references for natural language tracking,

    Y . Shao, S. He, Q. Ye, Y . Feng, W. Luo, and J. Chen, “Context- aware integration of language and visual references for natural language tracking,” arXiv preprint arXiv:2403.19975 , 2024

    2024 arXiv

  11. [19]

    Textual tokens classification for multi-modal alignment in vision-language tracking,

    Z. Mao, Y . Wang, X. Chen, and J. Yan, “Textual tokens classification for multi-modal alignment in vision-language tracking,” in ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 8025–8029

    2024

  12. [20]

    One-stream stepwise decreasing for vision-language tracking,

    G. Zhang, B. Zhong, Q. Liang, Z. Mo, N. Li, and S. Song, “One-stream stepwise decreasing for vision-language tracking,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

    2024

  13. [21]

    Joint feature learning and relation modeling for tracking: A one-stream framework,

    B. Ye, H. Chang, B. Ma, S. Shan, and X. Chen, “Joint feature learning and relation modeling for tracking: A one-stream framework,” in Proceedings of the European Conference on Computer Vision , 2022, pp. 341–357

    2022

  14. [22]

    Autore- gressive queries for adaptive tracking with spatio-temporaltransformers,

    J. Xie, B. Zhong, Z. Mo, S. Zhang, L. Shi, S. Song, and R. Ji, “Autore- gressive queries for adaptive tracking with spatio-temporaltransformers,” arXiv preprint arXiv:2403.10574 , 2024

    2024 arXiv

  15. [23]

    Odtrack: Online dense temporal token learning for visual tracking,

    Y . Zheng, B. Zhong, Q. Liang, Z. Mo, S. Zhang, and X. Li, “Odtrack: Online dense temporal token learning for visual tracking,” arXiv preprint arXiv:2401.01686, 2024

    2024 arXiv

  16. [24]

    Beyond accuracy: Tracking more like human via visual search,

    D. Zhang, S. Hu, X. Feng, X. Li, M. Wu, J. Zhang, and K. Huang, “Beyond accuracy: Tracking more like human via visual search,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems

  17. [25]

    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,” Proceedings of the Advances in Neural Information Processing Systems , vol. 30, 2017

    2017

  18. [26]

    A hierarchical theme recognition model for sandplay therapy,

    X. Feng, S. Hu, X. Chen, and K. Huang, “A hierarchical theme recognition model for sandplay therapy,” in Chinese Conference on Pattern Recognition and Computer Vision (PRCV) . Springer, 2023, pp. 241–252

    2023

  19. [27]

    Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models,

    J. Luo, S. Khandelwal, L. Sigal, and B. Li, “Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4029–4040

    2024

  20. [28]

    Grounding everything: Emerging localization properties in vision-language trans- formers,

    W. Bousselham, F. Petersen, V . Ferrari, and H. Kuehne, “Grounding everything: Emerging localization properties in vision-language trans- formers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 3828–3837

    2024

  21. [29]

    Robust visual object tracking with natural language region proposal network,

    Q. Feng, V . Ablavsky, Q. Bai, and S. Sclaroff, “Robust visual object tracking with natural language region proposal network,” arXiv preprint arXiv:1912.02048, vol. 1, no. 7, p. 8, 2019

    1912 arXiv

  22. [30]

    Real-time visual object tracking with natural language description,

    Q. Feng, V . Ablavsky, Q. Bai, G. Li, and S. Sclaroff, “Real-time visual object tracking with natural language description,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2020, pp. 700–709

    2020

  23. [31]

    Grounding-tracking- integration,

    Z. Yang, T. Kumar, T. Chen, J. Su, and J. Luo, “Grounding-tracking- integration,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 9, pp. 3433–3443, 2021

    2021

  24. [32]

    Cross-modal target retrieval for tracking by natural language,

    Y . Li, J. Yu, Z. Cai, and Y . Pan, “Cross-modal target retrieval for tracking by natural language,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 4931–4940

    2022

  25. [33]

    Divert more attention to vision-language tracking,

    M. Guo, Z. Zhang, H. Fan, and L. Jing, “Divert more attention to vision-language tracking,” in Proceedings of the Advances in Neural Information Processing Systems , S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35, 2022, pp. 4446–4460

    2022

  26. [34]

    Transformer vision- language tracking via proxy token guided cross-modal fusion,

    H. Zhao, X. Wang, D. Wang, H. Lu, and X. Ruan, “Transformer vision- language tracking via proxy token guided cross-modal fusion,” Pattern Recognition Letters, vol. 168, pp. 10–16, 2023

    2023

  27. [35]

    Joint visual grounding and tracking with natural language specification,

    L. Zhou, Z. Zhou, K. Mao, and Z. He, “Joint visual grounding and tracking with natural language specification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 23 151–23 160

    2023

  28. [36]

    Unified transformer with isomorphic branches for natural language tracking,

    R. Wang, Z. Tang, Q. Zhou, X. Liu, T. Hui, Q. Tan, and S. Liu, “Unified transformer with isomorphic branches for natural language tracking,” IEEE Transactions on Circuits and Systems for Video Technology, 2023

    2023

  29. [37]

    Tracking by natural language specification with long short-term context decoupling,

    D. Ma and X. Wu, “Tracking by natural language specification with long short-term context decoupling,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 14 012–14 021

    2023

  30. [38]

    Towards unified token learning for vision-language tracking,

    Y . Zheng, B. Zhong, Q. Liang, G. Li, R. Ji, and X. Li, “Towards unified token learning for vision-language tracking,” IEEE Transactions on Circuits and Systems for Video Technology , 2023

    2023

  31. [39]

    Onetracker: Unifying visual object tracking with foundation models and efficient tuning,

    L. Hong, S. Yan, R. Zhang, W. Li, X. Zhou, P. Guo, K. Jiang, Y . Chen, J. Li, Z. Chen et al. , “Onetracker: Unifying visual object tracking with foundation models and efficient tuning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024...

    2024

  32. [40]

    Revealing the dark secrets of extremely large kernel convnets on robustness,

    H. Chen, Y . Zhang, X. Feng, X. Chu, and K. Huang, “Revealing the dark secrets of extremely large kernel convnets on robustness,” arXiv preprint arXiv:2407.08972, 2024

    2024 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.