Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Explicit Context Reasoning with Supervision for Visual Tracking

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

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

Pith's one-line read The paper claims that explicitly supervising temporal reasoning with true target states improves visual tracking and reports top scores on six benchmarks.

desk verdict RSTrack is a solid, well-ablated tracking paper whose central supervision mechanism is genuinely new, but it never checks that the clean state histories used in training survive contact with the tracker's own predictions at inference. read the letter →

arxiv 2507.16191 v2 pith:FULVTOXI submitted 2025-07-22 cs.CV

Fansheng Zeng , Bineng Zhong , Haiying Xia , Yufei Tan , Xiantao Hu , Liangtao Shi , Shuxiang Song This is my paper
classification cs.CV
keywords visualobjecttrackingcontextreasoningtemporalmodelingstatespacemodelforwardsupervisioncompressionRSTrack
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

RSTrack treats temporal context as a supervised reasoning problem instead of stacking historical frames. The tracker compresses each frame's target appearance into compact state tokens, runs a Mamba-based state-space model over the history to predict the current state, and then supervises that prediction against the true target state extracted from the ground-truth crop through an L2 loss. The authors argue this forward supervision suppresses context drift, the silent divergence of implicit temporal models from the real target, and that compression removes cross-frame redundancy so the reasoning chain stays efficient. A sympathetic reader would care because the three mechanisms are modular and the reported gains are consistent across LaSOT, LaSOT$_{ext}$, GOT-10K, TrackingNet, TNL2K, and UAV123, at real-time speeds.

What carries the argument

The load-bearing mechanism is a compression-reconstruction state loop with a supervised reasoning step. A spatial-channel compression module squeezes true or predicted target features into compact state tokens; a Mamba-based state reasoning module $\xi$ reads the history of these tokens, plus an initialized current token, and predicts the current state; a reconstruction module turns the predicted tokens back into predicted target features by fusing them with template features. The forward supervision loss $L_{\text{state}}$ and the reconstruction loss $L_{\text{recon}}$ are what make the reasoning explicit, and a dual-stage cross-attention temporal decoder injects the reconstructed features into the search region before the prediction head.

What would settle it

Measure the distribution gap between training-time true state tokens and test-time self-generated tokens on a held-out split, for instance the mean cosine distance between $S_T$ and $\hat{S}_T$ on high-confidence frames, and check whether tracking failures concentrate where this gap is large. Alternatively, ablate only the $L_{\text{state}}$ term while keeping $L_{\text{recon}}$; if the LaSOT AUC drop is negligible, the forward supervision is not the active ingredient the paper claims.

Watch

Extended reading notes

Core claim

RSTrack's central claim is that explicitly supervising the temporal reasoning process with true target states improves visual tracking. During training, the true target feature is cropped using labels, compressed into state tokens, and used in an L2 loss $L_{\text{state}} = \|S_T - \hat{S}_T\|_2^2 + \|C_T - \hat{C}_T\|_2^2$ against the predicted tokens, with an additional reconstruction loss $L_{\text{recon}}$ tying compressed states back to full features. At inference, the tracker alternates between learned state inference and feature compression, replaces noisy states with the initial state when the classification score falls below 0.4, and sums forward and backward reasoning within a 60-frame interval. On the paper's measurements this gives, for example, LaSOT AUC 74.4 for RSTrack-384 versus 73.6 for MambaLCT-384, with gains on the other five benchmarks.

Load-bearing premise

The gains rest on the assumption that the state tokens the tracker generates from its own predictions at test time are close enough to the true state tokens computed from ground-truth crops during training; if the self-generated states drift, the supervision learned at training will not transfer.

Editorial extensions

If this is right

  • Explicitly supervising the reasoning chain, not just the final box, becomes a viable recipe for reducing context drift in temporal trackers.
  • Compact state tokens let temporal modeling run at lower FLOPs: RSTrack-384 uses 57G FLOPs and runs at 34.7 fps, against 148G and 17.8 fps for SeqTrack-384 in the paper's comparison.
  • The same training recipe transfers across resolutions and benchmarks, with RSTrack-256 already beating several 384-resolution trackers on LaSOT.
  • Because the training signal comes from ground-truth states, the tracker's robustness depends on keeping inference-time state generation close to the trained distribution, which motivates confidence-based state resetting.

Reading between the lines

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

  • Inference beyond the paper: the compression-reasoning-supervision loop could transfer to other sequential vision tasks where ground-truth states are available at training time, such as video object segmentation or multi-object tracking, provided inference-time state generation stays reliable.
  • Inference beyond the paper: the fixed confidence threshold of 0.4 and the fixed 60-frame interval look like hand-tuned compensations for inference-time state noise; a learned uncertainty estimate over predicted states could replace them and likely extend the context window beyond the 500-token window tested.
  • Inference beyond the paper: if the train/test state-gap problem is solved, SSM-based reasoning should support substantially longer video context than the window used here, since the paper's state-space model handles long sequences at linear cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes RSTrack, a visual tracker that combines a Fast-iTPN encoder, a Mamba-based state reasoning module, a compression-reconstruction module, and a forward supervision loss. The state reasoning module predicts current target state tokens from a historical state sequence, the compression module reduces target features to compact spatial and channel tokens, and the supervision loss (Eq. 5) forces predicted state tokens to match state tokens extracted from ground-truth target crops. The method is evaluated on LaSOT, LaSOT_ext, GOT-10k, TrackingNet, TNL2K, and UAV123, reporting state-of-the-art results at real-time speed, with code released.

Significance. If the claimed gains hold, the paper offers a practical and conceptually clear way to inject explicit temporal reasoning supervision into transformer/Mamba trackers, and the compression-reconstruction mechanism is a sensible response to the computational cost of long-term context modeling. The paper's strengths include consistent component ablations (Tables 4 and 5), evaluation across six benchmarks, attribute-level analysis, and code release. I do not share the circularity concern: Eq. (5) is a standard supervised loss between predicted states and states derived from ground-truth crops, not a fitted constant. The main weakness is that the training-time state distribution differs from the inference-time state distribution, and the paper does not analyze or mitigate that shift quantitatively.

major comments (3)
  1. [§3.5, Eq. (5)] The central claim that Lssm suppresses context drift is not validated under the inference-time state distribution. Section 3.5 states that during training the historical state sequence is maintained "without relying on the predictions from the state inference model, using only feature compression tokens," whereas at inference it is "maintained using state inference and feature compression" and is augmented with confidence-threshold resets and reverse reasoning. Thus the Mamba in Eq. (4) is trained on clean, ground-truth-derived histories, but at deployment it must reason from histories containing its own imperfect predictions. The paper reports no measurement of the distance between predicted and true state tokens during evaluation, no training-time injection of noisy or self-generated states, and no separate ablation of the confidence reset or reverse-interval heuristics. Consequently, Table 5 (#4 vs #5) demonstrates the value of Lssm in the clean-history training regime, but it does not establish that the learned supervision transfers to self-generated state histories at test time. I recommend adding a quantitative drift analysis, a training variant that simulates inference-time histories, or at minimum an ablation that isolates each inference heuristic.
  2. [Tables 2–3] The state-of-the-art claims rest on single-run evaluation metrics, and the margins over the strongest comparators are small: for example, RSTrack-384 exceeds MambaLCT-384 by 0.8 AUC on LaSOT, by 0.5–0.7 on LaSOT_ext, and by 0.4 AUC on UAV123. Without error bars or multiple evaluation runs, these differences are not shown to be statistically distinguishable, and the ablation deltas in Table 4 are also small (0.4–1.5 AUC). Please report mean and standard deviation over at least three runs for the main comparisons, or provide evidence from the official benchmark protocol that the reported differences are stable.
  3. [§3.5, inference] The inference-time reverse reasoning is described only qualitatively as "summing the forward and backward inferred states," without an equation or a precise description of the sequence reversal, the handling of sequences shorter than the 60-frame interval, or the exact interaction between the confidence score and the state reset. Because these heuristics may be responsible for drift suppression at test time, they need to be specified precisely and ablated separately. As written, the reader cannot determine how much of the final performance comes from the learned supervision versus from these hand-designed inference-time mechanisms.
minor comments (5)
  1. [Table 5] The check/cross notation is ambiguous: rows #2 and #3 do not make it transparent which components are removed, and the text's description of removing "all context reasoning components (only keeping the decoder)" versus "removing the decoder" does not clearly map to the table rows. Please name each configuration explicitly, for example "no SRM, no RE, keep TD" instead of using only symbols.
  2. [§3.3.2, Eq. (4)] Eq. (4) gives only a high-level description of the state reasoning module; please specify the state dimensions, the number of Mamba layers, the sequence lengths, and whether the spatial and channel Mamba networks share parameters.
  3. [§3.5, inference] The reverse reasoning procedure should be described with a concrete equation or pseudocode, including how the backward state is computed and how the forward and backward states are weighted when summed.
  4. [Table 1] The speed comparison includes only SeqTrack; please include at least the direct Mamba-based competitors (TemTrack, MambaLCT) under the same device and resolution for a fair efficiency comparison.
  5. [Keywords and Table 3] There are several typos: "State Space Mode" should be "State Space Model" in the keywords, "UA V123" in Table 3 should be "UAV123", and the affiliation entry "Nanjin" should be "Nanjing".

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Eq. 5 is a standard supervised loss between ground-truth-derived state tokens and predicted state tokens, and the benchmark comparisons are external evaluation rather than fitted inputs.

full rationale

The paper's central derivation is a supervised sequence-modeling pipeline: historical state tokens are fed to a Mamba-based state reasoning module to predict the current state (Eq. 4), and the true state tokens compressed from the ground-truth crop provide the target for an L2 loss (Eq. 5). This is a conventional supervised training objective, not a parameter fitted to the reported benchmark numbers. The reconstruction loss (Eq. 6) regularizes the compression-reconstruction mechanism and is also independent of the test benchmarks. The ablation in Table 5 measures the contribution of Lssm empirically, and while removing it costs about 1.0 AUC, that is an experimental result rather than a claim forced by construction. Related-work citations to TemTrack and MambaLCT come from the same research group, but they are used only to position the design and not as a load-bearing justification or uniqueness theorem; the method's components (Mamba, Fast-iTPN, loss weights) are additionally grounded in external prior work or standard practice. The training-versus-inference mismatch described in Section 3.5, where training uses compression-only histories while inference alternates with predicted states, is a genuine distribution-shift and robustness concern, but it is not a circular reduction: the supervision is still applied to a quantity computed from ground truth, and the test-time predictions are not fed back into Eq. 5 during evaluation. Overall, no step of the derivation chain reduces to its own inputs, and no fitted parameter is renamed as a prediction. The correct circularity finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on hand-chosen loss weights and inference thresholds, plus four domain assumptions about the faithfulness of the supervision signal, Mamba's suitability, compression-reconstruction fidelity, and train-test transfer. No genuinely new physical or mathematical entities are introduced.

free parameters (6)
  • alpha (state loss weight) = 0.5
    Hand-chosen loss weight in Eq. 7 balancing the state token supervision against reconstruction; no sensitivity analysis is provided.
  • beta (reconstruction loss weight) = 1.0
    Hand-chosen loss weight in Eq. 7 for the reconstruction term; its value affects how strongly the predicted state is tied to true target features.
  • lambda_ssm (temporal supervision loss weight) = 4
    Overall weight of the SSM supervision loss in Eq. 8, chosen without sensitivity analysis.
  • Historical state window size = 500 frames
    Window size for sampling the most recent state tokens in Section 4.3.3; impacts how much temporal context is used by the reasoning module.
  • Confidence reset threshold = 0.4
    Inference threshold in Section 3.5 for replacing the current state with the initial value when the maximum classification score is low.
  • Reverse reasoning interval = 60 frames
    Frame interval in Section 3.5 for modeling the target state in reverse and summing forward and backward inferred states.
assumptions (4)
  • domain assumption The frozen backbone's features of the ground-truth target crop constitute a reliable and sufficient supervision signal for the state reasoning module.
    Used in Section 3.4; if these features do not faithfully represent the target state, the Lstate loss will supervise the predictor toward a poor target.
  • domain assumption Mamba's selective state space model can capture the temporal dynamics of target state token sequences.
    Section 3.3.2 adopts Mamba based on prior successes; no proof is given that target state tokens evolve like a linear state space.
  • domain assumption The compression-reconstruction mechanism preserves the core target information needed for tracking.
    Section 3.3.1; if compression discards discriminative details, the reconstructed features will be impoverished and will not help tracking.
  • domain assumption At test time, alternating state inference and feature compression yields states comparable to the ground-truth states used in training.
    Section 3.5 Inference; this train-test transfer assumption is unverified and is the weakest load-bearing premise of the empirical claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explicit Context Reasoning with Supervision for Visual Tracking." pith.science (2026). https://pith.science/paper/FULVTOXI

@misc{pith2026250716191,
  author       = {Pith},
  title        = {Pith review of: Explicit Context Reasoning with Supervision for Visual Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FULVTOXI}},
  note         = {Machine review of arXiv:2507.16191}
}
read the original abstract

Contextual reasoning with constraints is crucial for enhancing temporal consistency in cross-frame modeling for visual tracking. However, mainstream tracking algorithms typically associate context by merely stacking historical information without explicitly supervising the association process, making it difficult to effectively model the target's evolving dynamics. To alleviate this problem, we propose RSTrack, which explicitly models and supervises context reasoning via three core mechanisms. \textit{1) Context Reasoning Mechanism}: Constructs a target state reasoning pipeline, converting unconstrained contextual associations into a temporal reasoning process that predicts the current representation based on historical target states, thereby enhancing temporal consistency. \textit{2) Forward Supervision Strategy}: Utilizes true target features as anchors to constrain the reasoning pipeline, guiding the predicted output toward the true target distribution and suppressing drift in the context reasoning process. \textit{3) Efficient State Modeling}: Employs a compression-reconstruction mechanism to extract the core features of the target, removing redundant information across frames and preventing ineffective contextual associations. These three mechanisms collaborate to effectively alleviate the issue of contextual association divergence in traditional temporal modeling. Experimental results show that RSTrack achieves state-of-the-art performance on multiple benchmark datasets while maintaining real-time running speeds. Our code is available at https://github.com/GXNU-ZhongLab/RSTrack.

Figures

Figures reproduced from arXiv: 2507.16191 by the authors.

Figure 1
Figure 1. Our framework. The historical state tokens [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The architecture of RSTrack. (a) Main Pipeline: The encoder extracts visual features from video frames. The state [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Compression-Reconstruction Mechanism. The detailed structures of the channel compression module and spatial [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: State Reasoning Module. The current state is first [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: AUC scores of different attributes on LaSOT. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visual Comparison. Our tracker vs. three SOTA trackers on LaSOT under three challenges [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 41 canonical work pages

  1. [1]

    Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. 2024. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 19048–19057

    work page 2024

  2. [2]

    UT Benchmark. 2016. A benchmark and simulator for uav tracking. In European conference on computer vision , Vol. 7

    work page 2016

  3. [3]

    Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. 2016. Fully-convolutional siamese networks for object tracking. InComputer Vision–ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15- 16, 2016, Proceedings, Part II 14 . Springer, 850–865

    work page 2016

  4. [4]

    Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. 2023. Robust object modeling for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision. 9589–9600

    work page 2023

  5. [5]

    Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. 2023. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14572–14581

    work page 2023

  6. [6]

    Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu

  7. [7]

    Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, Rongrong Ji, Zhenjun Tang, and Xianxian Li. 2022. SiamBAN: Target-aware tracking with Siamese box adaptive network. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 4 (2022), 5158–5173

    work page 2022

  8. [8]

    Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. 2022. Mixformer: End- to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 13608–13618

    work page 2022

Show all 57 references
  1. [9]

    Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg

  2. [10]

    Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  3. [11]

    Heng Fan, Hexin Bai, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Harshit, Mingzhen Huang, Juehuan Liu, et al . 2021. Lasot: A high-quality large-scale single object tracking benchmark. International Journal of Computer Vision 129 (2021), 439–461

    2021

  4. [12]

    Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. 2019. 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

  5. [13]

    Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. 2021. Stmtrack: Template- free visual tracking with space-time memory networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 13774–13783

    2021

  6. [14]

    Albert Gu and Tri Dao. 2024. Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv:2312.00752 [cs.LG] https://arxiv.org/abs/2312.00752

    2024 arXiv

  7. [15]

    Albert Gu, Ankit Gupta, Karan Goel, and Christopher Ré. 2022. On the Parameteri- zation and Initialization of Diagonal State Space Models. arXiv:2206.11893 [cs.LG] https://arxiv.org/abs/2206.11893

    2022 arXiv

  8. [16]

    Ankit Gupta, Albert Gu, and Jonathan Berant. 2022. Diagonal State Spaces are as Effective as Structured State Spaces. arXiv:2203.14343 [cs.LG] https: //arxiv.org/abs/2203.14343

    2022 arXiv

  9. [17]

    Xiantao Hu, Ying Tai, Xu Zhao, Chen Zhao, Zhenyu Zhang, Jun Li, Bineng Zhong, and Jian Yang. 2024. Exploiting Multimodal Spatial-temporal Patterns for Video Object Tracking. arXiv preprint arXiv:2412.15691 (2024)

    2024 arXiv

  10. [18]

    Xiantao Hu, Bineng Zhong, Qihua Liang, Zhiyi Mo, Liangtao Shi, Ying Tai, and Jian Yang. 2025. Adaptive Perception for Unified Visual Multi-modal Object Tracking. arXiv preprint arXiv:2502.06583 (2025)

    2025 arXiv

  11. [19]

    Xiantao Hu, Bineng Zhong, Qihua Liang, Shengping Zhang, Ning Li, and Xianxian Li. 2024. Toward Modalities Correlation for RGB-T Tracking. IEEE Transactions on Circuits and Systems for Video Technology 34, 10 (2024), 9102–9111. doi:10. 1109/TCSVT.2024.3396289

    2024

  12. [20]

    Xiantao Hu, Bineng Zhong, Qihua Liang, Shengping Zhang, Ning Li, Xianxian Li, and Rongrong Ji. 2024. Transformer Tracking via Frequency Fusion. IEEE Transactions on Circuits and Systems for Video Technology 34, 2 (2024), 1020–1031. doi:10.1109/TCSVT.2023.3289624

    2024

  13. [21]

    Lianghua Huang, Xin Zhao, and Kaiqi Huang. 2019. Got-10k: A large high- diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence 43, 5 (2019), 1562–1577

    2019

  14. [22]

    Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. 2019. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 4282–4291

    2019

  15. [23]

    Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. 2018. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE conference on computer vision and pattern recognition . 8971–8980

    2018

  16. [24]

    Xiaohai Li, Bineng Zhong, Qihua Liang, Guorong Li, Zhiyi Mo, and Shuxiang Song. 2024. MambaLCT: Boosting Tracking via Long-term Context State Space Model. arXiv preprint arXiv:2412.13615 (2024)

    2024 arXiv

  17. [25]

    T Lin. 2017. Focal Loss for Dense Object Detection.arXiv preprint arXiv:1708.02002 (2017)

    2017 arXiv

  18. [26]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proce...

    2014

  19. [27]

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. 2024. VMamba: Visual State Space Model. arXiv:2401.10166 [cs.CV] https://arxiv.org/abs/2401.10166

    2024 arXiv

  20. [28]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs.LG] https://arxiv.org/abs/1711.05101

    2019 arXiv

  21. [29]

    Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. 2018. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In Proceedings of the European conference on computer vision (ECCV). 300–317

    2018

  22. [30]

    Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. 2019. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 658–666

    2019

  23. [31]

    Liangtao Shi, Bineng Zhong, Qihua Liang, Xiantao Hu, Zhiyi Mo, and Shuxiang Song. 2025. Mamba Adapter: Efficient Multi-Modal Fusion for Vision-Language Tracking. IEEE Transactions on Circuits and Systems for Video Technology (2025), 1–1. doi:10.1109/TCSVT.2025.3557570

    2025

  24. [32]

    Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Shengping Zhang, and Xianx- ian Li. 2024. Explicit Visual Prompts for Visual Object Tracking. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 4838–4846

    2024

  25. [33]

    Jimmy T. H. Smith, Shalini De Mello, Jan Kautz, Scott W. Linderman, and Wonmin Byeon. 2023. Convolutional State Space Models for Long-Range Spatiotemporal Modeling. arXiv:2310.19694 [cs.LG] https://arxiv.org/abs/2310.19694

    2023 arXiv

  26. [34]

    Zhangyong Tang, Tianyang Xu, Hui Li, Xiao-Jun Wu, Xuefeng Zhu, and Josef Kittler. 2023. Exploring fusion strategies for accurate RGBT visual object tracking. Information Fusion 99 (2023), 101881

    2023

  27. [35]

    Zhangyong Tang, Tianyang Xu, Xiaojun Wu, Xue-Feng Zhu, and Josef Kittler

  28. [36]

    Yunjie Tian, Lingxi Xie, Jihao Qiu, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. 2024. Fast-iTPN: Integrally pre-trained transformer pyramid net- work with token migration. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  29. [37]

    Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. 2021. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogniti...

    2021

  30. [38]

    Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. 2023. Autore- gressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 9697–9706

    2023

  31. [39]

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. 2018. CBAM: Convolutional Block Attention Module. arXiv:1807.06521 [cs.CV] https://arxiv. org/abs/1807.06521

    2018 arXiv

  32. [40]

    Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. 2022. Correlation-Aware Deep Tracking. arXiv:2203.01666 [cs.CV] https: //arxiv.org/abs/2203.01666

    2022 arXiv

  33. [41]

    Jinxia Xie, Bineng Zhong, Qihua Liang, Ning Li, Zhiyi Mo, and Shuxiang Song

  34. [42]

    Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. 2024. Autoregressive Queries for Adaptive Tracking with Spatio-Temporal Transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 19300–19309

    2024

  35. [43]

    Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. 2020. SiamFC++: Towards Robust and Accurate Visual Tracking with Target Estimation Guidelines. arXiv:1911.06188 [cs.CV] https://arxiv.org/abs/1911.06188

    2020 arXiv

  36. [44]

    Chaocan Xue, Bineng Zhong, Qihua Liang, Haiying Xia, and Shuxiang Song

  37. [45]

    arXiv preprint arXiv:2412.13611 (2024)

    Robust Tracking via Mamba-based Context-aware Token Learning. arXiv preprint arXiv:2412.13611 (2024)

    2024 arXiv

  38. [46]

    Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. 2021. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision . 10448–10457

    2021

  39. [47]

    Dawei Yang, Jianfeng He, Yinchao Ma, Qianjin Yu, and Tianzhu Zhang. 2023. Foreground-Background Distribution Modeling Transformer for Visual Object Tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 10117–10127

    2023

  40. [48]

    Yijun Yang, Zhaohu Xing, Lequan Yu, Chunwang Huang, Huazhu Fu, and Lei Zhu. 2024. Vivim: a Video Vision Mamba for Medical Video Segmentation. arXiv:2401.14168 [cs.CV] https://arxiv.org/abs/2401.14168

    2024 arXiv

  41. [49]

    Unifying Motion and Appearance Cues for Visual Tracking via Shared MM ’25, October 27–31, 2025, Dublin, Ireland Fansheng Zeng et al. Queries. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

  42. [50]

    Chaocan Xue, Bineng Zhong, Qihua Liang, Yaozong Zheng, Ning Li, Yuanliang Xue, and Shuxiang Song. 2025. Similarity-Guided Layer-Adaptive Vision Trans- former for UAV Tracking. arXiv:2503.06625 [cs.CV] https://arxiv.org/abs/2503. 06625

    2025 arXiv

  43. [51]

    Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. 2024. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 7588–7596

    2024

  44. [52]

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. 2024. Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model. arXiv:2401.09417 [cs.CV] https://arxiv. org/abs/2401.09417

    2024 arXiv

  45. [54]

    Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. 2022. Joint feature learning and relation modeling for tracking: A one-stream framework. In European Conference on Computer Vision . Springer, 341–357

    2022

  46. [55]

    Zhipeng Zhang, Houwen Peng, Jianlong Fu, Bing Li, and Weiming Hu. 2020. Ocean: Object-aware anchor-free tracking. In European conference on computer vision. Springer, 771–787

    2020

  47. [2017]

    In Proceedings of the IEEE conference on computer vision and pattern recognition

    Eco: Efficient convolution operators for tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition . 6638–6646

  48. [2021]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Transformer tracking. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8126–8135

  49. [2024]

    In Proceed- ings of the AAAI Conference on Artificial Intelligence , Vol

    Generative-based fusion mechanism for multi-modal tracking. In Proceed- ings of the AAAI Conference on Artificial Intelligence , Vol. 38. 5189–5197

Pith tools

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