Pith. sign in

REVIEW 3 major objections 4 minor 64 references

A tracker that lets a language description control its temporal memory achieves state-of-the-art results on UAV anti-UAV tracking.

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

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

SATATrack conditions a Mamba-style temporal memory on the target's language description and aligns feature statistics at test time, achieving state-of-the-art on the UAV-Anti-UAV benchmark.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A genuinely novel semantic-modulated SSM tracker with strong ablations, but the SOTA claim is undercut by hyperparameters tuned on the test benchmark and missing code/supplementary. the 3 major comments →

arxiv 2607.26511 v1 pith:D3QWK6BC submitted 2026-07-29 cs.CV

Semantic-Aware Temporal Adaptation for UAV Anti-UAV Tracking

Xiaozhen Qiao , Da Zhang , Yubin Guo , Junyu Gao , Zhiyuan Zhao , Xuelong Li This is my paper
classification cs.CV
keywords UAV anti-UAV trackingvision-language trackingstate-space modeltemporal context propagationtest-time adaptationdistribution alignmentcontrastive learningtarget semantic description
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper tries to establish that in air-to-air UAV tracking — where both the observer and the target are moving — the target's language description is a more stable cue than appearance, and that it should govern the temporal memory of a state-space tracker. It proposes SATATrack, which uses the description to modulate what gets written into and retained by a propagated context, plus a test-time routine that recalibrates feature statistics without touching the learned weights. On the UAV-Anti-UAV benchmark this achieves the best scores on all five metrics, with the largest gain on visually similar distractors, and it remains competitive on related anti-UAV and UAV tracking benchmarks. A sympathetic reader would care because it suggests a concrete way to combat identity drift and domain shift when appearance evidence is unreliable.

Core claim

The central claim is that semantic guidance should be injected directly into the state-space recurrence that carries temporal context, rather than fused as an auxiliary feature or a one-shot query. In SATATrack, the target text embedding modulates the discrete transition of the selective state-space model: it scales the step size (the write/forget gate) per token to amplify target-consistent observations and suppress distractor tokens, and it shifts the log-parameterized transition matrix to set how long target-consistent context is retained. An auxiliary contrastive loss pulls the target region toward the semantic embedding while pushing away the most semantically similar background patches

What carries the argument

The core mechanism is Semantic-Aware Context Propagation (SACP): inside each stage of a selective state-space model, the target language embedding is used to modulate two quantities that control the temporal memory — the input-dependent step size Δ, which acts as a write/forget gate, and the transition matrix A, which sets the retention timescale. This lets semantics decide which observations enter the propagated state and how long they persist. A second mechanism, Temporal-Aware Distribution Alignment (TADA), adapts batch-normalization statistics at test time by computing them only over high-energy spatial locations in recent frames, smoothing them through FIFO queues, and blending with tra

Load-bearing premise

TADA assumes that the highest-energy spatial locations in each search frame carry target-relevant signal; when the background is textured sky or vegetation, those high-energy pixels can be clutter, and the adapted statistics then normalize toward the background rather than the target.

What would settle it

Run TADA with the top-ρ locations replaced by random or low-energy locations of the same count: if AUC and OP75 do not drop, the 'high energy means target' assumption is wrong. A second test is to apply TADA to a curated set of videos with heavy background clutter and check whether the adapted statistics correlate with actual target locations rather than distractor regions.

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

If this is right

  • If the claim holds, state-space trackers should condition their temporal memory on stable semantic anchors rather than only on visual similarity, reducing distractor contamination and drift.
  • Test-time distribution alignment that leaves weights frozen and uses location-selective statistics is a viable and cheap complement to semantic guidance, improving overlap-precision metrics.
  • The same components bring the best average AUC over three anti-UAV benchmarks and competitive results on two generic UAV tracking benchmarks, suggesting the design does not over-specialize to air-to-air scenarios.
  • The reported runtime of about 42 FPS indicates that semantic control of the recurrence need not sacrifice real-time processing, since the text is encoded once per sequence.
  • Because the largest attribute-level gain is on similar distractors, the design suggests that controlling what the state retains is more effective than improving per-frame matching alone.

Where Pith is reading between the lines

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

  • If the energy-based location selection in TADA is safe, a similar selective-statistics scheme could be applied to other small-object trackers that suffer from background-dominated normalization.
  • The framework implicitly assumes a stable, informative language description is available; extending it to auto-generated or adaptively refined descriptions would be a testable next step.
  • The idea of using semantics to control state-space write/forget dynamics could transfer to other sequential perception tasks such as video object segmentation or multi-object tracking, where identity descriptions are available.
  • A direct test of the energy-to-importance assumption would be to run TADA with statistics computed from the lowest-energy locations; if gains persist, the mechanism is not about target energy but about something else.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes SATATrack, a vision-language tracker for UAV-to-UAV anti-UAV tracking. The method combines Semantic-Aware Context Propagation (SACP), which injects a frozen CLIP text embedding into a Mamba-style state-space update by modulating the step-size Δ and transition matrix A; a Semantic-Discriminative Contrastive (SDC) loss; and Temporal-Aware Distribution Alignment (TADA), a test-time adaptation scheme that recomputes batch-normalization statistics from top-energy spatial locations, smooths them with FIFO queues, interpolates with training statistics, and blends adapted and frozen outputs. On the UAV-Anti-UAV benchmark the authors report state-of-the-art results on all five metrics (46.40% AUC, 38.63% OP75, etc.), with SACP as the main ablation gain; the tracker is also evaluated on Anti-UAV318, DUT Anti-UAV, UAV123, and UAVDT.

Significance. If the results hold, the paper makes a useful contribution: it proposes a concrete mechanism for using language to control state-space memory write/forget dynamics, and the ablation evidence suggests that modulating Δ and A outperforms simple text fusion. TADA is lightweight, gradient-free, and conceptually simple. The paper also includes cross-benchmark generalization and efficiency comparisons. However, the headline SOTA claim is currently weakened by two load-bearing issues: final hyperparameters are selected on the evaluation benchmark itself, and TADA's core energy-to-importance assumption is asserted without validation. These issues should be fixed before the empirical claims can be taken at face value.

major comments (3)
  1. [Sec. IV-D3, Fig. 8] The final TADA configuration (N=8, n=6, α=1.0, ρ=0.20) is explicitly chosen as the configuration that achieves the reported 46.40% AUC on UAV-Anti-UAV. Because the same benchmark is used for evaluation, the reported headline numbers are the result of test-set model selection, not the performance of a fixed model. This makes the 'first on all five metrics' claim and the +2.37 AUC gain over MambaSTS difficult to interpret, since the comparison baselines are not afforded the same selection procedure. Please retune on a validation split (or use nested/cross-validation) and report the corresponding test performance, or report the mean and spread over the hyperparameter sweep and clearly separate validation from test selection.
  2. [Algorithm 1, Eqs. (21)-(23), Table V] TADA assumes that the top-ρ=20% highest-energy spatial locations are target-relevant and therefore yield reliable per-frame mean/variance estimates. In dual-dynamic air-to-air scenes with tiny, texture-poor UAVs, high-energy locations can be background clutter or distractors; the paper provides no evidence for this energy-to-importance assumption. Table V shows that the top-k statistics are essential (removing them drops AUC from 46.40 to 43.34, below the w/o-TADA baseline), but this ablation only compares against full-map statistics, not against a target-guided oracle. Please add an oracle-controlled experiment in which Ω_t is selected from the ground-truth target region, report attribute-specific results (e.g., similar distractor, small object) for TADA, and specify the drift-rejection test and clipping safeguards that are currently deferred to an absent supplementary document.
  3. [Sec. III-C / III-E] Several hyperparameters controlling the proposed modules are not reported: η in Eq. (8), κ in Eq. (24), τ and M in Eqs. (16)-(18), the SDC weight λ_sdc, and the queue length L. Without these values, the method is not reproducible, and the sensitivity of SACP/SDC cannot be assessed. This is particularly relevant because the claimed gains are large and the final model is selected on the evaluation benchmark; full disclosure of all hyperparameters is necessary.
minor comments (4)
  1. [Sec. III-E, Eq. (24)] c_t is called a 'per-channel confidence factor' but the formula in Eq. (24) suggests a scalar (or a vector over channels). Please clarify the broadcasting and define N before its use in λ_0 = n/(n+N).
  2. [Sec. IV-D3] The sentence 'the final reported model follows the configuration that achieves 46.40% AUC' should be reworded to avoid the appearance of test-set selection; this is related to Major Comment 1.
  3. [Fig. 8] The four panels of Fig. 8 are not fully self-contained; please label the x-axes explicitly (N, n, α, ρ) or add a caption that clarifies each panel.
  4. [Sec. IV-C6, Table IV] The 189.87M parameter count includes the frozen CLIP text encoder while only 126.44M are trainable. Please state this in Table IV or its caption to avoid misleading comparisons.

Circularity Check

1 steps flagged

No equation-level circularity; the main circularity-adjacent issue is TADA hyperparameters tuned on the same benchmark whose SOTA is then reported.

specific steps
  1. fitted input called prediction [Section IV-D.3 (TADA Analysis), hyperparameter sensitivity discussion and Fig. 8]
    "Based on these observations, we use N=8, n=6, α=1.0, and ρ=0.20 as the final hyperparameter values for TADA on UAV-Anti-UAV, and the final reported model follows the configuration that achieves 46.40% AUC."

    The four TADA hyperparameters are selected by sweeping on the UAV-Anti-UAV benchmark itself (Fig. 8), and the final reported SOTA model is explicitly chosen as the configuration that achieves 46.40% AUC on that same benchmark. Thus the headline result (46.40 AUC, +2.37 over MambaSTS) is the maximum of the tuning curve, not an independent evaluation of a fixed model; TADA's claimed gain (44.07 to 46.40 in Table V) is partly forced by selection on the same test set. This is a selection artifact rather than an equation identity, and it does not make SACP's separate, externally validated contribution circular.

full rationale

The paper's central derivation is SACP: target semantics enter the state-space update by modulating the write gate Δ (Eq. 8) and the retention matrix A (Eq. 9), followed by the standard discretized SSM recurrence (Eqs. 10-13). This is a compositional design choice, not a result defined in terms of its own output; its benefit is tested on external benchmarks (Anti-UAV318, DUT Anti-UAV, UAV123, UAVDT) and through component ablations, so the main claim has independent content. The only reduction-by-construction I can exhibit is in TADA: the final hyperparameters are chosen by sweeping on UAV-Anti-UAV, and then the same benchmark's best configuration is reported as the SOTA result. That is a fitted-selection issue, not an equation identity, and it is limited to the TADA margin; SACP+SDC without TADA already reaches 44.07 AUC versus MambaSTS's 44.03. The self-citations [33,34] merely motivate avoiding online parameter updates and are not load-bearing. There is no imported uniqueness theorem, no ansatz smuggled via citation, and no renaming of a known result. The absent supplementary details on drift rejection and clipping are a completeness gap, not evidence of circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard SSM discretization, the stability of language descriptions, the high-energy-location proxy for target foreground, and a first-order BN-statistics shift model. No new physical entities are introduced. The main free parameters are TADA's hyperparameters, which are tuned on the evaluation benchmark itself, plus several unspecified constants (η, κ, τ, M) that appear directly in the method equations.

free parameters (8)
  • TADA candidate number N = 8
    Selected on the UAV-Anti-UAV test benchmark as best AUC in Fig. 8(a); used to set base blend λ0 = n/(n+N).
  • TADA current-frame weight n = 6
    Selected on the same benchmark; Fig. 8(b) shows saturation after n=6.
  • TADA output blend α = 1.0
    Selected on the same benchmark; Fig. 8(c) shows best AUC at α=1.0.
  • TADA top-k ratio ρ = 0.20
    Selected on the same benchmark; Fig. 8(d) shows optimum at ρ=0.20.
  • SACP semantic-gate scale η = not specified
    Introduced in Eq. (8) as a modulation strength on the semantic gate, but no value or sensitivity study is given.
  • TADA confidence-gate scale κ = not specified
    Introduced in Eq. (24) as a scale hyperparameter; no default or sweep is reported.
  • SDC temperature τ = not specified
    Used in Eq. (17) for contrastive cosine similarities; no value reported.
  • SDC hard-negative count M = not specified
    Used in Eq. (16) to select top-M background patches; no value reported.
axioms (5)
  • standard math Exponential discretization of the selective state-space update yields a stable temporal memory.
    Eqs. (10)-(11) use the standard Mamba discretization Ā=exp(Δ̃ Ã) and the recurrence h = Ā⊙h + B̄⊙u; treated as standard background.
  • domain assumption A CLIP text embedding of the target description remains a stable, identity-specific semantic signal across frames.
    Core to SACP (Sec. III-C); if the description is outdated or mismatched, the semantic gate can amplify the wrong temporal memory.
  • ad hoc to paper The top-ρ highest-energy spatial locations are target-relevant for estimating test-time batch-norm statistics.
    Algorithm 1 selects Ω_t by per-location energy; no validation that high-energy locations correspond to the small UAV rather than clutter.
  • domain assumption Test-time domain shift in anti-UAV tracking can be modeled by first-order batch-norm mean/variance shifts only.
    TADA (Sec. III-E) recalibrates only BN statistics; higher-order distribution changes are ignored.
  • domain assumption Natural-language target descriptions are available at inference time in the deployed setting.
    The UAV-Anti-UAV benchmark [3] provides descriptions, but real counter-UAV deployments may not have them.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic-Aware Temporal Adaptation for UAV Anti-UAV Tracking." pith.science (2026). https://pith.science/paper/D3QWK6BC

@misc{pith2026260726511,
  author       = {Pith},
  title        = {Pith review of: Semantic-Aware Temporal Adaptation for UAV Anti-UAV Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D3QWK6BC}},
  note         = {Machine review of arXiv:2607.26511}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

UAV Anti-UAV tracking is an emerging low-altitude security task for localizing an adversarial UAV using the onboard camera of a moving observer UAV. It differs from conventional UAV tracking and ground-based Anti-UAV tracking because both the camera platform and the target move simultaneously. This dual-dynamic setting induces rapid viewpoint changes, motion blur, scale variation, and visually similar distractors, making reliable appearance matching difficult. Under such rapidly changing conditions, fixed visual representations are often insufficient because target appearance becomes unreliable and feature distributions may deviate from the training domain. The target language description remains stable across frames and can therefore serve as a semantic anchor for temporal state propagation, while online feature-distribution alignment can reduce video-specific test-time shifts. In this paper, we propose \emph{SATATrack}, a Semantic-Aware Temporal Adaptation framework for UAV Anti-UAV tracking. SATATrack introduces Semantic-Aware Context Propagation (SACP), which uses the target description to guide temporal context propagation across backbone stages and preserve target identity under rapid appearance changes. An auxiliary contrastive regularizer is used during training to discourage responses to semantically similar background regions. During inference, Temporal-Aware Distribution Alignment (TADA) aligns feature distributions online without updating model parameters, combining recent-frame estimates with training-time statistics for stability. SATATrack achieves state-of-the-art performance on the UAV-Anti-UAV benchmark while remaining competitive in Anti-UAV and UAV object tracking tasks. The code will be available at https://github.com/XiaozhenQiao/SATATrack.

Figures

Figures reproduced from arXiv: 2607.26511 by Da Zhang, Junyu Gao, Xiaozhen Qiao, Xuelong Li, Yubin Guo, Zhiyuan Zhao.

Figure 1
Figure 1. Figure 1: (a) Traditional visual trackers. (b) Existing vision-language trackers. (c) The proposed SATATrack, which adds SACP and TADA. (d) Radar of the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed pipeline. Given a template image, a search image, and a semantic description of the target, the framework extracts multi-level [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pipeline of the proposed SACP module, where the target semantics [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with ODTrack, MambaLCT, MCITrack, and MambaTrack on UAV-Anti-UAV sequences involving fast motion, motion blur, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Temporal stability comparison on three representative UAV-Anti-UAV sequences. The first row shows frame-wise IoU curves, and the second row [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Bar-chart visualization of the component ablation on UAV-Anti-UAV. Each panel shows one metric under progressively added components, from the [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Response-map visualization over short UAV-Anti-UAV video clips (eight consecutive frames each), comparing the pure-vision baseline, Ours w/o [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Hyperparameter sensitivity analysis of TADA on UAV-Anti-UAV. From left to right, (a) to (d) vary the candidate number [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ablation of SACP designs on UAV-Anti-UAV without TADA. The [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Representative challenging cases where our tracker may still [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

64 extracted references · 5 linked inside Pith

  1. [1]

    AI flow: Perspectives, scenarios, and approaches,

    H. An, W. Hu, S. Huang, S. Huang, R. Li, Y . Liang, J. Shao, Y . Song, Z. Wang, C. Yuan, C. Zhang, H. Zhang, W. Zhuang, and X. Li, “AI flow: Perspectives, scenarios, and approaches,”Vicinagearth, vol. 3, no. 1, p. 1, 2026

    2026

  2. [2]

    Generative video compression: Towards 0.01% compression rate for video transmission,

    X. Chen, J. Luo, J. Xu, F. Yi, C. Zhang, and X. Li, “Generative video compression: Towards 0.01% compression rate for video transmission,” Vicinagearth, vol. 3, no. 1, p. 7, 2026

    2026

  3. [3]

    How far are modern trackers from uav-anti-uav? a million- scale benchmark and new baseline,

    C. Zhang, L. Liu, Z. Zhang, Y . Wang, H. Wen, X. Zhou, S. Ge, and Y . Wang, “How far are modern trackers from uav-anti-uav? a million- scale benchmark and new baseline,”arXiv preprint arXiv:2512.07385, 2025

    arXiv 2025

  4. [4]

    A benchmark and simulator for UA V tracking,

    M. Mueller, N. Smith, and B. Ghanem, “A benchmark and simulator for UA V tracking,” inECCV, 2016, pp. 445–461

    2016

  5. [5]

    The unmanned aerial vehicle benchmark: Object detection and tracking,

    H. Yu, G. Li, W. Zhanget al., “The unmanned aerial vehicle benchmark: Object detection and tracking,” inECCV, 2018, pp. 370–386

    2018

  6. [6]

    Detection and tracking meet drones challenge,

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, and H. Ling, “Detection and tracking meet drones challenge,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 7380–7399, 2021

    2021

  7. [7]

    Webuav-3m: A benchmark for unveiling the power of million- scale deep uav tracking,

    C. Zhang, G. Huang, L. Liu, S. Huang, Y . Yang, X. Wan, S. Ge, and D. Tao, “Webuav-3m: A benchmark for unveiling the power of million- scale deep uav tracking,”IEEE TPAMI, vol. 45, no. 7, pp. 9186–9205, 2023

    2023

  8. [8]

    Anti-uav: A large-scale benchmark for vision-based uav tracking,

    N. Jiang, K. Wang, X. Peng, X. Yu, Q. Wang, J. Xing, G. Li, G. Guo, Q. Ye, J. Jiaoet al., “Anti-uav: A large-scale benchmark for vision-based uav tracking,”IEEE TMM, vol. 25, pp. 486–500, 2021

    2021

  9. [9]

    Vision-based anti-uav detection and tracking,

    J. Zhao, J. Zhang, D. Li, and D. Wang, “Vision-based anti-uav detection and tracking,”IEEE TITS, vol. 23, no. 12, pp. 25 323–25 334, 2022

    2022

  10. [10]

    Anti-uav410: A thermal infrared benchmark and customized scheme for tracking drones in the wild,

    B. Huang, J. Li, J. Chen, G. Wang, J. Zhao, and T. Xu, “Anti-uav410: A thermal infrared benchmark and customized scheme for tracking drones in the wild,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  11. [11]

    Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-uav system,

    X.-F. Zhu, T. Xu, J. Zhao, J.-W. Liu, K. Wang, G. Wang, J. Li, Q. Wang, L. Jin, Z. Zhuet al., “Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-uav system,” arXiv preprint arXiv:2306.15767, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    A tri-modal dataset and a baseline system for tracking unmanned aerial vehicles,

    T. Xu, J. Gu, X. Zhu, X. Wu, and J. Kittler, “A tri-modal dataset and a baseline system for tracking unmanned aerial vehicles,”arXiv preprint arXiv:2511.18344, 2025

    arXiv 2025

  13. [13]

    Fully-convolutional siamese networks for object tracking,

    L. Bertinetto, J. Valmadre, J. F. Henriqueset al., “Fully-convolutional siamese networks for object tracking,” inECCV, 2016, pp. 850–865

    2016

  14. [14]

    Atom: Accurate tracking by overlap maximization,

    M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg, “Atom: Accurate tracking by overlap maximization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 4660–4669

    2019

  15. [15]

    Siamrpn++: Evolution of siamese visual tracking with very deep networks,

    B. Li, W. Wu, Q. Wang, F. Zhang, J. Xing, and J. Yan, “Siamrpn++: Evolution of siamese visual tracking with very deep networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 4282–4291

    2019

  16. [16]

    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,” inEuropean Conference on Computer Vision, 2022, pp. 341–357

    2022

  17. [17]

    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,” inProceed- ings of the AAAI conference on artificial intelligence, vol. 38, no. 7, 2024, pp. 7588–7596

    2024

  18. [18]

    To- wards more flexible and accurate object tracking with natural language: Algorithms and benchmark,

    X. Wang, X. Shu, Z. Zhang, B. Jiang, Y . Wang, Y . Tian, and F. Wu, “To- wards more flexible and accurate object tracking with natural language: Algorithms and benchmark,” inIEEE CVPR, 2021, pp. 13 763–13 773

    2021

  19. [19]

    Divert more attention to vision- language tracking,

    M. Guo, Z. Zhang, H. Fan, and L. Jing, “Divert more attention to vision- language tracking,”NeurIPS, 2022

    2022

  20. [20]

    Joint visual grounding and track- ing with natural language specification,

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

    2023

  21. [21]

    Citetracker: Correlating image and text for visual tracking,

    X. Li, Y . Huang, Z. He, Y . Wang, H. Lu, and M.-H. Yang, “Citetracker: Correlating image and text for visual tracking,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 9974–9983

    2023

  22. [22]

    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,” inProceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 5552–5561

    2023

  23. [23]

    Unifying visual and vision-language tracking via contrastive learning,

    Y . Ma, Y . Tang, W. Yang, T. Zhang, J. Zhang, and M. Kang, “Unifying visual and vision-language tracking via contrastive learning,”arXiv preprint arXiv:2401.11228, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Sutrack: Towards simple and unified single object tracking,

    X. Chen, B. Kang, W. Geng, J. Zhu, Y . Liu, D. Wang, and H. Lu, “Sutrack: Towards simple and unified single object tracking,” inAAAI, vol. 39, no. 2, 2025, pp. 2239–2247

    2025

  25. [25]

    Dynamic updates for language adaptation in visual-language tracking,

    X. Li, B. Zhong, Q. Liang, Z. Mo, J. Nong, and S. Song, “Dynamic updates for language adaptation in visual-language tracking,” inCVPR, 2025, pp. 19 165–19 174

    2025

  26. [26]

    Atctrack: Aligning target-context cues with dynamic target states for robust vision-language tracking,

    X. Feng, S. Hu, X. Li, D. Zhang, M. Wu, J. Zhang, X. Chen, and K. Huang, “Atctrack: Aligning target-context cues with dynamic target states for robust vision-language tracking,” inICCV, 2025, pp. 19 850– 19 861

    2025

  27. [27]

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

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

    Pith/arXiv arXiv 2023

  28. [28]

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

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

    Pith/arXiv arXiv 2024

  29. [29]

    Mambanut: Nighttime uav tracking via mamba-based adaptive curriculum learning,

    Y . Wu, X. Yang, X. Wang, H. Ye, D. Zeng, and S. Li, “Mambanut: Nighttime uav tracking via mamba-based adaptive curriculum learning,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 18 864–18 871

    2025

  30. [30]

    Mambalct: Boosting tracking via long-term context state space model,

    X. Li, B. Zhong, Q. Liang, G. Li, Z. Mo, and S. Song, “Mambalct: Boosting tracking via long-term context state space model,” inProceed- ings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 5, 2025, pp. 4986–4994

    2025

  31. [31]

    Exploring enhanced contextual information for video-level object tracking,

    B. Kang, X. Chen, S. Lai, Y . Liu, Y . Liu, and D. Wang, “Exploring enhanced contextual information for video-level object tracking,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 4194–4202. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    2025

  32. [32]

    Mambatrack: Exploiting dual-enhancement for night uav tracking,

    C. Zhang, L. Liu, H. Wen, X. Zhou, and Y . Wang, “Mambatrack: Exploiting dual-enhancement for night uav tracking,” inICASSPP. IEEE, 2025, pp. 1–5

    2025

  33. [33]

    Bidirectional prototype-reward co-evolution for test-time adaptation of vision-language models,

    X. Qiao, P. Huang, J. Yuan, X. Guo, B. Ye, C. Xue, Y . Zheng, Z. Sun, and X. Li, “Bidirectional prototype-reward co-evolution for test-time adaptation of vision-language models,”IEEE Transactions on Multimedia, 2026

    2026

  34. [34]

    Class- aware prototype learning with negative contrast for test-time adaptation of vision-language models,

    X. Qiao, J. Zhao, Y . Jiang, X. Guo, Z. Sun, H. Zhang, and X. Li, “Class- aware prototype learning with negative contrast for test-time adaptation of vision-language models,”arXiv preprint arXiv:2510.19802, 2025

    arXiv 2025

  35. [35]

    Got-10k: A large high-diversity benchmark for generic object tracking in the wild,

    L. Huang, X. Zhao, and K. Huang, “Got-10k: A large high-diversity benchmark for generic object tracking in the wild,”IEEE TPAMI, vol. 43, no. 5, pp. 1562–1577, 2019

    2019

  36. [36]

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

    H. Fan, L. Lin, F. Yanget al., “Lasot: A high-quality benchmark for large-scale single object tracking,” inCVPR, 2019, pp. 5374–5383

    2019

  37. [37]

    Track- ingnet: A large-scale dataset and benchmark for object tracking in the wild,

    M. Muller, A. Bibi, S. Giancola, S. Alsubaihi, and B. Ghanem, “Track- ingnet: A large-scale dataset and benchmark for object tracking in the wild,” inECCV, 2018, pp. 300–317

    2018

  38. [38]

    Eco: Efficient convolution operators for tracking,

    M. Danelljan, G. Bhat, F. Shahbaz Khan, and M. Felsberg, “Eco: Efficient convolution operators for tracking,” inCVPR, 2017, pp. 6638– 6646

    2017

  39. [39]

    Transformer tracking,

    X. Chen, B. Yan, J. Zhu, D. Wang, X. Yang, and H. Lu, “Transformer tracking,” inCVPR, 2021, pp. 8126–8135

    2021

  40. [40]

    Hift: Hierarchical feature transformer for aerial tracking,

    Z. Cao, C. Fu, J. Ye, B. Li, and Y . Li, “Hift: Hierarchical feature transformer for aerial tracking,” inICCV, 2021, pp. 15 457–15 466

    2021

  41. [41]

    Tctrack: Temporal contexts for aerial tracking,

    Z. Cao, Z. Huang, L. Pan, S. Zhang, Z. Liu, and C. Fu, “Tctrack: Temporal contexts for aerial tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 14 798–14 808

    2022

  42. [42]

    Adaptive and background- aware vision transformer for real-time uav tracking,

    S. Li, Y . Yang, D. Zeng, and X. Wang, “Adaptive and background- aware vision transformer for real-time uav tracking,” inICCV, 2023, pp. 13 989–14 000

    2023

  43. [43]

    Similarity-guided layer-adaptive vision transformer for uav tracking,

    C. Xue, B. Zhong, Q. Liang, Y . Zheng, N. Li, Y . Xue, and S. Song, “Similarity-guided layer-adaptive vision transformer for uav tracking,” inProceedings of the Computer Vision and Pattern Recognition Confer- ence, 2025, pp. 6730–6740

    2025

  44. [44]

    Learning occlusion-robust vision transformers for real-time uav tracking,

    Y . Wu, X. Wang, X. Yang, M. Liu, D. Zeng, H. Ye, and S. Li, “Learning occlusion-robust vision transformers for real-time uav tracking,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 103–17 113

    2025

  45. [45]

    Securing the skies: A comprehensive survey on anti-uav methods, benchmarking, and future directions,

    Y . Dong, F. Wu, S. Zhang, G. Chen, Y . Hu, M. Yano, J. Sun, S. Huang, F. Liu, Q. Daiet al., “Securing the skies: A comprehensive survey on anti-uav methods, benchmarking, and future directions,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 6659–6673

    2025

  46. [46]

    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

    Pith/arXiv arXiv 2018

  47. [47]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, andet al., “Learning transferable visual models from natural language supervision,” inInternational Conference on Machine Learning, 2021, pp. 8748–8763

    2021

  48. [48]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, andet al., “An image is worth 16x16 words: Transformers for image recognition at scale,” in International Conference on Learning Representations, 2020

    2020

  49. [49]

    Transformer meets tracker: Exploiting temporal context for robust visual tracking,

    N. Wang, W. Zhouet al., “Transformer meets tracker: Exploiting temporal context for robust visual tracking,” inCVPR, 2021, pp. 1571– 1580

    2021

  50. [50]

    Learning spatio-temporal transformer for visual tracking,

    B. Yan, H. Peng, J. Fu, D. Wang, and H. Lu, “Learning spatio-temporal transformer for visual tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 10 448–10 457

    2021

  51. [51]

    Generalized intersection over union: A metric and a loss for bound- ing box regression,

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized intersection over union: A metric and a loss for bound- ing box regression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 658–666

    2019

  52. [52]

    Positive-incentive noise,

    X. Li, “Positive-incentive noise,”IEEE TNNLS, vol. 35, no. 6, pp. 8708– 8714, 2024

    2024

  53. [53]

    Variational positive-incentive noise: How noise benefits models,

    H. Zhang, S. Huang, Y . Guo, and X. Li, “Variational positive-incentive noise: How noise benefits models,”IEEE TPAMI, vol. 47, no. 9, pp. 8313–8320, 2025

    2025

  54. [54]

    Improving accuracy and generalization for efficient visual tracking,

    R. Zaveri, S. Patel, Y . Gu, and G. Doretto, “Improving accuracy and generalization for efficient visual tracking,” inProceedings of the Winter Conference on Applications of Computer Vision, 2025, pp. 9450–9460

    2025

  55. [55]

    Pura: Parameter update-recovery test-time adaption for rgb-t tracking,

    Z. Shao, Y . Hu, B. Fan, and H. Liu, “Pura: Parameter update-recovery test-time adaption for rgb-t tracking,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22 089–22 098

    2025

  56. [56]

    Fast online object tracking and segmentation: A unifying approach,

    Q. Wang, L. Zhang, L. Bertinetto, W. Hu, and P. H. Torr, “Fast online object tracking and segmentation: A unifying approach,” inCVPR, 2019, pp. 1328–1338

    2019

  57. [57]

    Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines,

    Y . Xu, Z. Wang, Z. Li, Y . Yuan, and G. Yu, “Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines,” inAAAI, vol. 34, no. 07, 2020, pp. 12 549–12 556

    2020

  58. [58]

    Siamese box adaptive network for visual tracking,

    Z. Chen, B. Zhong, G. Li, S. Zhang, and R. Ji, “Siamese box adaptive network for visual tracking,” inCVPR, 2020, pp. 6667–6676

    2020

  59. [59]

    Siamcar: Siamese fully convolutional classifica- tion and regression for visual tracking,

    D. Guo, J. Wanget al., “Siamcar: Siamese fully convolutional classifica- tion and regression for visual tracking,” inCVPR, 2020, pp. 6269–6277

    2020

  60. [60]

    Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search,

    B. Yan, H. Peng, K. Wu, D. Wang, J. Fu, and H. Lu, “Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search,” inCVPR, 2021, pp. 15 180–15 189

    2021

  61. [61]

    Graph attention tracking,

    D. Guo, Y . Shao, Y . Cui, Z. Wang, L. Zhang, and C. Shen, “Graph attention tracking,” inCVPR, 2021, pp. 9543–9552

    2021

  62. [62]

    Learn to match: Automatic matching network design for visual tracking,

    Z. Zhang, Y . Liu, X. Wanget al., “Learn to match: Automatic matching network design for visual tracking,” inICCV, 2021, pp. 13 339–13 348

    2021

  63. [63]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongieet al., “Microsoft coco: Common objects in context,” inECCV, 2014, pp. 740–755

    2014

  64. [64]

    Fast-itpn: Integrally pre-trained transformer pyramid network with token migra- tion,

    Y . Tian, L. Xie, J. Qiu, J. Jiao, Y . Wang, Q. Tian, and Q. Ye, “Fast-itpn: Integrally pre-trained transformer pyramid network with token migra- tion,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 9766–9779, 2024

    2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.