Pith. sign in

REVIEW 3 major objections 5 minor 119 references

Learning an Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Per-frame block switching lets a UAV tracker match heavier models while running 17% faster.

desk verdict Useful engineering extension of the authors' ICML 2024 tracker, but the paper leaves the training of its adaptive gate unspecified—without code inspection, the central adaptive claim is unverified. read the letter →

arxiv 2412.20002 v3 pith:IRJDGYAV submitted 2024-12-28 cs.CV

You Wu , Yongxin Li , Mengyuan Liu , Xucheng Wang , Xiangyang Yang , Hengzhou Ye , Dan Zeng , Qijun Zhao
show 1 more author
Shuiwang Li
This is my paper · ORCID
classification cs.CV
keywords UAVtrackingvisiontransformeradaptivecomputationactivationmoduleview-invariantrepresentationmutualinformationmaximizationknowledgedistillationreal-time
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

AVTrack is a single-stream vision-transformer tracker for UAVs that tries to be both accurate and fast enough for real-time onboard use. Its first idea is an Activation Module attached to each transformer block, which looks at a one-token summary of the previous block's output and decides, per input frame, whether that block should run or be skipped. The second idea is to make features view-invariant by maximizing mutual information between the template and the target patch in the search image, which is meant to help with the extreme viewpoint changes typical of aerial tracking. The paper also presents AVTrack-MD, a student model with half the blocks of three AVTrack teachers, trained by maximizing mutual information between the averaged softened teacher features and the student's softened features. The reported payoff is that the student matches the teachers' accuracy while boosting average tracking speed by over 17%, running at 310.6 FPS on a desktop GPU and 46.1 FPS on an edge device.

What carries the argument

The machinery that carries the efficiency claim is the Activation Module (AM): a single linear layer $L_i$ and sigmoid $\sigma$, applied to the first token $r_{i-1}$ of the previous block's output, computing $p_i = \sigma(L_i(r_{i-1}))$. If $p_i$ exceeds the threshold $\beta$, the block runs; otherwise its output is skipped, and the sparsity loss $L_{\mathrm{spar}} = \left| \frac{1}{N-n_f}\sum_{i=n_f+1}^{N} p_i - \zeta \right|$ keeps the average activation low. The robustness claim is carried by $L_{\mathrm{vir}}$, which uses the Deep InfoMax Jensen-Shannon estimator $\hat{I}^{(JSD)}_{\Theta}$ to maximize mutual information between template and search-target features, and the distillation claim is carried by $L_{\mathrm{MD}}$, which maximizes the same estimator between averaged softened teacher features and the student's softened features.

What would settle it

Record the per-frame activation pattern of every AM on a diverse test set such as WebUAV-3M, and compare a variant with a fixed schedule of active blocks (e.g., always the same six blocks) against the reported adaptive model. If the fixed variant matches the speed-accuracy trade-off, or if the active-block set is essentially the same for all frames, the input-adaptive claim is falsified; a clear correlation between scene difficulty (clutter, viewpoint change) and the number of active blocks would support it.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a transformer tracker does not need to run every block for every frame. An Activation Module (AM)—one linear layer followed by a sigmoid—consumes the first token output by the previous block and produces an activation probability $p_i$; if $p_i > \beta$ with $\beta \in (0.5,1)$, the block is activated, otherwise its input is passed straight through. A block sparsity loss $L_{\mathrm{spar}}$ pulls the average of $p_i$ toward a constant $\zeta$, so most blocks are off on average. On top of this, the framework maximizes a Jensen-Shannon mutual-information estimate between the template's features and the target's features in the search image, yielding what the authors call view-invariant representations, and the distilled AVTrack-MD maximizes mutual information between the aggregated softened features of three teachers and the student's softened features. The claim is that this combination delivers performance comparable to the full AVTrack while reducing parameters and FLOPs and increasing speed by more than 17%.

Load-bearing premise

The load-bearing premise is that the Activation Module's on/off decision can be trained by gradient descent, even though the paper does not describe how gradients flow through the hard threshold; if that premise fails, the network may simply learn a fixed schedule of active blocks rather than a truly input-dependent one.

Editorial extensions

If this is right

  • AVTrack-MD-DeiT reports 83.6% precision and 64.1% success averaged over five UAV benchmarks, within 0.5/0.3 points of AVTrack-DeiT, while running at 310.6 FPS on a Titan X GPU and 46.1 FPS on an NVIDIA Jetson AGX Xavier.
  • The three-teacher distillation beats single-teacher and MSE-based distillation: all multi-teacher students stay within 1.6% of their teachers, and two students exceed their corresponding teacher on at least one metric.
  • Increasing the student's ViT blocks from 4 to 6 yields more than 1.0 point gain per block in both precision and success, while going beyond 6 blocks gives diminishing gains with a clear speed penalty.
  • The view-invariant representation term improves feature-map consistency across viewpoints and contributes to top results on 'Viewpoint Change' (UAV123) and 'Camera Motion' (VisDrone2018) attribute subsets.
  • Because activation decisions skip entire blocks rather than randomly discarding tokens, the method avoids the unstructured memory-access overhead that slowed the Aba-ViTrack baseline.

Reading between the lines

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

  • A direct measurement of per-frame activation patterns would separate two readings of the paper: a truly input-dependent architecture, or a learned static pruning of blocks; the reported FLOPs ranges (0.97–2.4G) are consistent with both.
  • If the hard threshold has no differentiable relaxation, the sparsity loss alone cannot teach input-dependence, so the AM's advertised adaptivity may reduce to 'learned static pruning' — a testable gap that the current ablations do not isolate.
  • The MI-based distillation loss is architecture-agnostic, so it could be lifted onto other single-stream trackers that have comparable template and search feature maps, not just AVTrack.
  • One extension worth testing is sweeping the threshold $\beta$ at inference time without retraining: a genuinely adaptive model should trade accuracy for speed smoothly, whereas a static-pruned model should show a sharp cliff.
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 AVTrack, an adaptive-computation single-stream transformer tracker for UAV tracking, in which an Activation Module (AM) attached to each transformer block predicts a probability and hard-gates the block's execution. It further proposes AVTrack-MD, a multi-teacher distillation variant that trains a smaller student with a mutual-information (MI) objective, and claims comparable accuracy to AVTrack while improving speed by over 17%. The paper reports evaluations on six UAV tracking benchmarks, including comparisons with lightweight and deep trackers, and provides qualitative feature-map visualizations and an embedded-device test. The central constructive claims are that the AM learns input-dependent block activation and that the MI-based distillation yields view-invariant representations.

Significance. If the adaptive activation mechanism and the MI-based distillation work as claimed, the paper would provide a practical real-time UAV tracker with a useful trade-off between accuracy and speed, and the release of code is a strength for reproducibility. The benchmark coverage is extensive and includes diverse datasets and hardware settings. However, the core novelty—input-dependent adaptive activation—rests on a training mechanism that the manuscript does not describe, and this undermines the validity of the reported speed/accuracy trade-off as presented. The distillation contribution, while interesting, is also partly confounded with the smaller student backbone, so the quantitative evidence must be separated more carefully. For these reasons, the significance is real but conditional on resolving the trainability issue and providing controlled ablations.

major comments (3)
  1. [Sec. III-B, Eqs. (1)-(2), Sec. III-E] The AM uses a hard threshold pi > beta to decide whether a transformer block is executed, and Section III-E states that the framework is trained end-to-end with L_overall. The manuscript does not describe any differentiable relaxation (e.g., Gumbel-Softmax, straight-through estimator, or a soft-gating scheme) for this discrete decision. Consequently, the task losses L_pred (Eq. 9) and L_vir (Eq. 5) cannot produce gradients for the AM's linear layer L_i. The only term that reaches L_i is L_spar (Eq. 2), which is a scalar penalty on the mean activation probability |mean_i p_i - zeta|; its gradient has the same sign for all samples and can only shift the marginal activation rate toward zeta. It cannot teach the AM to assign different activation decisions to easy versus hard inputs. Without an unstated relaxation, the 'adaptive, input-dependent' activation claimed in Sections I and III-A is not trainable as written, and the reported speed/accuracy trade-off of AVTrack is not explained by a learned gating policy. Please specify the gradient mechanism (or provide code evidence) that allows the task loss to reach the gate, or revise the claims accordingly.
  2. [Tables V and VI] The speed improvements reported for AVTrack-MD are confounded with the reduced backbone size. AVTrack-MD-DeiT uses a 6-block student (Table VI) while the teacher AVTrack-DeiT uses a 12-block backbone, and Table V shows that even single-teacher distillation yields a substantial speedup (e.g., 310.6 vs. 256.8 FPS for the DeiT models). To support the claim that the proposed activation module contributes to the 'over 17%' speed gain, report a controlled comparison where the student's architecture is fixed and the only difference is whether the AM is present and active. Alternatively, state clearly that the speedup relative to the teachers is primarily due to the halved backbone, and isolate any additional speedup attributable to the AM.
  3. [Sec. IV-D, Fig. 7] The view-invariant representation (VIR) component lacks a quantitative ablation. The only evidence for its effectiveness is the qualitative feature-map visualization in Fig. 7, which compares AVTrack-DeiT* (without both VIR and AM) to AVTrack-DeiT; this confounds the effect of VIR with the effect of the AM. Add an ablation removing VIR while keeping the AM, and vice versa, so that the contribution of the VIR loss (Eq. 5) to the viewpoint-change robustness can be quantified and attributed.
minor comments (5)
  1. [Sec. III-B] The notation eT_1 ti-1_1:K := ri-1 is dimensionally inconsistent: eT_1 is a row vector in R^K, ti-1_1:K is in R^{K x d}, so their product should be a d-dimensional row vector, not an element of R^K. Please rewrite the selector (e.g., using a slice operator) to make the extracted token ri-1 correctly typed.
  2. [Abstract] "Codes is available" should be "Code is available."
  3. [Table III caption] The caption contains a typo: "Compwith" should be "Comparison with".
  4. [Reference [35]] The author name "H. R.D." appears malformed; this should be a proper citation to the Deep InfoMax paper (e.g., Hjelm et al.).
  5. [Sec. IV-B] The subsection numbering is inconsistent: "5) Attribute-Based Evaluation" is followed by "4) Qualitative evaluation"; renumber these subsections sequentially.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's performance claims are external benchmark measurements, and self-citation to the prior ICML 2024 paper is used only as a legitimate baseline/teacher source, not as a load-bearing derivation step.

full rationale

The paper's central claims—AVTrack's adaptive block activation, view-invariant representation learning via MI maximization, and AVTrack-MD's distillation-based speed/accuracy trade-off—are validated by experiments on external UAV tracking benchmarks (DTB70, UAVDT, VisDrone2018, UAV123, UAV123@10fps, WebUAV-3M, UAV20L, UAVTrack112L) against many external trackers. The reported precision, success rate, and FPS values are measurements on held-out data, not quantities derived from the method's own definitions. AVTrack-MD trains a smaller student using three off-the-shelf AVTrack models as teachers, but distillation to a teacher does not by construction guarantee the measured comparable performance; the comparison is empirical. Self-citation to the ICML 2024 AVTrack paper [31] identifies the baseline and teacher models, but it does not force the present results. The hard-threshold activation decision (pi > beta) with no described differentiable relaxation is a genuine optimization concern: as written, L_pred and L_vir cannot backpropagate through the gate, and L_spar only controls the average activation probability, so input-dependent adaptivity may not actually be learned. That is a correctness or reproducibility risk, not circularity, because the speed and accuracy numbers are not constructed to equal the method's inputs. Likewise, the unreported values of beta and zeta make the speed/accuracy trade-off partly a design choice, but this is a reporting issue, not an equivalence between inputs and outputs. No circular step is exhibited in the paper's equations or claims.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The framework relies on pretrained ViT backbones, several hand-set hyperparameters, and an unstated mechanism for training the discrete block activation. No new physical entities are introduced. The most load-bearing element is the assumption that the hard threshold decision can be trained despite the absence of a described relaxation.

free parameters (8)
  • Activation threshold beta = not reported
    Controls the hard on/off decision for each transformer block; its value directly sets the speed/accuracy trade-off, but the paper never states it.
  • Sparsity target zeta = not reported
    Target average activation probability in L_spar; smaller zeta gives sparser model. Not reported.
  • Loss weight gamma = 50
    Weight for block sparsity loss in total loss.
  • Loss weight kappa = 0.0001
    Weight for view-invariant representation loss.
  • Distillation loss weight eta = 0.1e-3
    Weight for MD loss.
  • Softmax temperature tau = 2
    Temperature for softening features in MD.
  • Number of always-on initial blocks nf = not reported
    First nf blocks are always activated; value not reported.
  • Student ViT block count = 6
    Student backbone has half the teacher's blocks, chosen from ablation (Table VI) as the default balance.
assumptions (5)
  • domain assumption The first token from the previous block carries global contextual information sufficient for an activation decision.
    Used in Eq. (1) where AM reads only the first token to decide whether to run the block; not justified beyond the claim that it carries global info.
  • ad hoc to paper The hard threshold activation is trainable via gradient descent.
    The paper does not specify any relaxation (Gumbel-Softmax, straight-through, REINFORCE); if none is used, L_pred gradients cannot reach the AM.
  • standard math Jensen-Shannon MI estimator with a critic network provides a suitable training objective for view invariance and distillation.
    Deep InfoMax is a known estimator, but the critic's architecture and update rule are not specified.
  • domain assumption Softmax over feature maps with temperature tau creates meaningful soft targets for distillation.
    Used in Eq. (6); softening features with softmax is borrowed from classification KD and its suitability for spatial features is untested.
  • domain assumption Pre-trained ViT weights are a valid starting point.
    All backbones are initialized with pretrained ViT-tiny, DeiT-tiny, or EVA-tiny weights; standard practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning an Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking." pith.science (2026). https://pith.science/paper/IRJDGYAV

@misc{pith2026241220002,
  author       = {Pith},
  title        = {Pith review of: Learning an Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRJDGYAV}},
  note         = {Machine review of arXiv:2412.20002}
}
read the original abstract

Transformer-based models have improved visual tracking, but most still cannot run in real time on resource-limited devices, especially for unmanned aerial vehicle (UAV) tracking. To achieve a better balance between performance and efficiency, we propose AVTrack, an adaptive computation tracking framework that adaptively activates transformer blocks through an Activation Module (AM), which dynamically optimizes the ViT architecture by selectively engaging relevant components. To address extreme viewpoint variations, we propose to learn view-invariant representations via mutual information (MI) maximization. In addition, we propose AVTrack-MD, an enhanced tracker incorporating a novel MI maximization-based multi-teacher knowledge distillation framework. Leveraging multiple off-the-shelf AVTrack models as teachers, we maximize the MI between their aggregated softened features and the corresponding softened feature of the student model, improving the generalization and performance of the student, especially under noisy conditions. Extensive experiments show that AVTrack-MD achieves performance comparable to AVTrack's performance while reducing model complexity and boosting average tracking speed by over 17\%. Codes is available at: https://github.com/wuyou3474/AVTrack.

Figures

Figures reproduced from arXiv: 2412.20002 by the authors.

Figure 1
Figure 1. Compared to SOTA lightweight trackers, our AVTrack-DeiT sets a new record with 84.1% precision and a speed of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (left) Overview of the proposed AVTrack’s framework, which consists of a single-stream backbone and a prediction head. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An overview of the AM’s internal working mechanism. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: An overview of the proposed MI maximization-based [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The precision plots and success plots of attribute-based comparison are presented for the attribute subsets of DTB70 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The qualitative evaluations were performed on three [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 8
Figure 8. Figure 8: Real-world UAV tracking test on an embedded device: [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 7
Figure 7. Figure 7: For each group, we present input images from different [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 9
Figure 9. Figure 9: Analysis of failure cases: AVTrack-DeiT struggles [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

119 extracted references · 75 canonical work pages

  1. [1]

    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,” in ICCV, 2023, pp. 13 989–14 000

    2023

  2. [2]

    Autotrack: Towards high-performance visual tracking for uav with automatic spatio-temporal regularization,

    Y . Li, C. Fu, and et al, “Autotrack: Towards high-performance visual tracking for uav with automatic spatio-temporal regularization,” in CVPR, 2020

    2020

  3. [3]

    Hift: Hierarchical feature transformer for aerial tracking,

    Z. Cao, C. Fu, and et al., “Hift: Hierarchical feature transformer for aerial tracking,” in ICCV, 2021

    2021

  4. [4]

    Learning motion blur robust vision transformers with dynamic early exit for real- time uav tracking,

    Y . Wu, X. Wang, D. Zeng, H. Ye, X. Xie, Q. Zhao, and S. Li, “Learning motion blur robust vision transformers with dynamic early exit for real- time uav tracking,” arXiv preprint arXiv:2407.05383 , 2024

    arXiv 2024

  5. [5]

    High-speed tracking with kernelized correlation filters,

    J. F. Henriques, R. Caseiro, and et al, “High-speed tracking with kernelized correlation filters,” IEEE TPAMI, 2015

    2015

  6. [6]

    Learning spatially regularized correlation filters for visual tracking,

    M. Danelljan, G. Hager, F. Shahbaz Khan, and M. Felsberg, “Learning spatially regularized correlation filters for visual tracking,” in ICCV, 2015, pp. 4310–4318

    2015

  7. [7]

    Learning aberrance repressed correlation filters for real-time uav tracking,

    Z. Huang, C. Fu, and et al., “Learning aberrance repressed correlation filters for real-time uav tracking,” ICCV, pp. 2891–2900, 2019

    2019

  8. [8]

    Learning temporary block-based bidirectional incongruity-aware correlation filters for efficient uav object tracking,

    F. Lin, C. Fu, and et al., “Learning temporary block-based bidirectional incongruity-aware correlation filters for efficient uav object tracking,” IEEE TCSVT , vol. 31, no. 6, pp. 2160–2174, 2020

    2020

Show all 119 references
  1. [9]

    Learning background- aware correlation filters for visual tracking,

    H. Kiani Galoogahi, A. Fagg, and S. Lucey, “Learning background- aware correlation filters for visual tracking,” in ICCV, 2017, pp. 1135– 1143

    2017

  2. [10]

    Visual tracking via adaptive spatially-regularized correlation filters,

    K. Dai, D. Wang, H. Lu, C. Sun, and J. Li, “Visual tracking via adaptive spatially-regularized correlation filters,” in CVPR, 2019, pp. 4670–4679

    2019

  3. [11]

    Fully-convolutional siamese networks for object tracking,

    L. Bertinetto, J. Valmadre, J. F. Henriques, A. Vedaldi, and P. H. Torr, “Fully-convolutional siamese networks for object tracking,” in ECCV, 2016, pp. 850–865

    2016

  4. [12]

    High performance visual tracking with siamese region proposal network,

    B. Li, J. Yan, W. Wu, Z. Zhu, and X. Hu, “High performance visual tracking with siamese region proposal network,” in CVPR, 2018, pp. 8971–8980

    2018

  5. [13]

    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,” in CVPR, 2020, pp. 6668–6677

    2020

  6. [14]

    Tctrack: Temporal contexts for aerial tracking,

    Z. Cao, Z. Huang, and et al., “Tctrack: Temporal contexts for aerial tracking,” in CVPR, 2022

    2022

  7. [15]

    Bactrack: Building appearance collection for aerial tracking,

    X. Liu, T. Xu, and et al., “Bactrack: Building appearance collection for aerial tracking,” IEEE TCSVT , vol. 34, no. 6, pp. 5002–5017, 2023

    2023

  8. [16]

    Multi-step temporal modeling for uav tracking,

    X. Yuan, T. Xu, and et al., “Multi-step temporal modeling for uav tracking,” IEEE TCSVT , 2024

    2024

  9. [17]

    Vit spatio-temporal feature fusion for aerial object tracking,

    C. Guo, K. Liu, D. Deng, and X. Li, “Vit spatio-temporal feature fusion for aerial object tracking,” IEEE TCSVT, vol. 34, no. 8, pp. 6749–6761, 2023

    2023

  10. [18]

    Learning spatial-frequency transformer for visual object tracking,

    C. Tang, X. Wang, Y . Bai, Z. Wu, J. Zhang, and Y . Huang, “Learning spatial-frequency transformer for visual object tracking,” IEEE TCSVT, vol. 33, no. 9, pp. 5102–5116, 2023

    2023

  11. [19]

    Object tracking via spatial-temporal memory network,

    Z. Zhou, X. Li, T. Zhang, H. Wang, and Z. He, “Object tracking via spatial-temporal memory network,” IEEE TCSVT , vol. 32, no. 5, pp. 2976–2989, 2021

    2021

  12. [20]

    Target-aware tracking with spatial-temporal context attention,

    K.-J. He, C.-L. Zhang, S. Xie, Z.-X. Li, Z.-W. Wang, and R.-G. Qin, “Target-aware tracking with spatial-temporal context attention,” IEEE TCSVT, vol. 34, no. 8, pp. 7176–7189, 2024

    2024

  13. [21]

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

    B. Ye and et al., “Joint feature learning and relation modeling for tracking: A one-stream framework,” in ECCV, 2022, pp. 341–357

    2022

  14. [22]

    Backbone is all your need: A simplified architecture for visual object tracking,

    B. Chen, P. Li, and et al., “Backbone is all your need: A simplified architecture for visual object tracking,” in ECCV, 2022, pp. 375–392

    2022

  15. [23]

    Mixformer: End-to-end tracking with iterative mixed attention,

    Y . Cui, C. Jiang, and et al., “Mixformer: End-to-end tracking with iterative mixed attention,” in CVPR, 2022, pp. 13 608–13 618

    2022

  16. [24]

    Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks,

    Q. Wu, T. Yang, and et al., “Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks,” in CVPR, 2023, pp. 14 561–14 571

    2023

  17. [25]

    Zoomtrack: target-aware non-uniform resizing for efficient visual tracking,

    Y . Kou, J. Gao, and et al., “Zoomtrack: target-aware non-uniform resizing for efficient visual tracking,” NIPS, vol. 36, 2024

    2024

  18. [26]

    Autoregressive queries for adaptive tracking with spatio-temporal transformers,

    J. Xie and et al., “Autoregressive queries for adaptive tracking with spatio-temporal transformers,” in CVPR, 2024, pp. 19 300–19 309

    2024

  19. [27]

    Explicit visual prompts for visual object tracking,

    L. Shi, B. Zhong, Q. Liang, N. Li, S. Zhang, and X. Li, “Explicit visual prompts for visual object tracking,” in AAAI, 2024

    2024

  20. [28]

    Hiptrack: Visual tracking with historical prompts,

    W. Cai, Q. Liu, and Y . Wang, “Hiptrack: Visual tracking with historical prompts,” in CVPR, 2024, pp. 19 258–19 267

    2024

  21. [29]

    Autoregressive visual tracking,

    X. Wei, Y . Bai, Y . Zheng, D. Shi, and Y . Gong, “Autoregressive visual tracking,” in CVPR, 2023, pp. 9697–9706

    2023

  22. [30]

    Exploring lightweight hierarchical vision transformers for efficient visual track- ing,

    B. Kang, X. Chen, D. Wang, H. Peng, and H. Lu, “Exploring lightweight hierarchical vision transformers for efficient visual track- ing,” in CVPR, 2023, pp. 9612–9621

    2023

  23. [31]

    Learning adaptive and view-invariant vision transformer for real-time uav tracking,

    Y . Li, M. Liu, Y . Wu, and et al., “Learning adaptive and view-invariant vision transformer for real-time uav tracking,” in ICML, 2024

    2024

  24. [32]

    The mutual information: detecting and evaluating dependencies between variables,

    R. Steuer, J. Kurths, C. O. Daub, J. Weise, and J. Selbig, “The mutual information: detecting and evaluating dependencies between variables,” Bioinformatics, vol. 18, no. suppl 2, pp. S231–S240, 2002. 14

    2002

  25. [33]

    Temporal feature alignment and mutual information maximization for video-based human pose estimation,

    Z. Liu, R. Feng, and et al., “Temporal feature alignment and mutual information maximization for video-based human pose estimation,” in CVPR, 2022, pp. 11 006–11 016

    2022

  26. [34]

    Learning deep generative clustering via mutual information maximization,

    X. Yang, J. Yan, and et al., “Learning deep generative clustering via mutual information maximization,” IEEE TNNLS , vol. 34, no. 9, pp. 6263–6275, 2022

    2022

  27. [35]

    Learning deep representations by mutual informa- tion estimation and maximization,

    H. R.D. and et al., “Learning deep representations by mutual informa- tion estimation and maximization,” in ICLR, 2019

    2019

  28. [36]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,” IJCV, vol. 129, no. 6, pp. 1789–1819, 2021

    2021

  29. [37]

    Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks,

    L. Wang and K.-J. Yoon, “Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks,” IEEE TPAMI, vol. 44, no. 6, pp. 3048–3068, 2021

    2021

  30. [38]

    Mtkd: Multi-teacher knowledge distilla- tion for image super-resolution,

    Y . Jiang, C. Feng, and et al., “Mtkd: Multi-teacher knowledge distilla- tion for image super-resolution,” in ECCV, 2024, pp. 364–382

    2024

  31. [39]

    Class incremental learning with multi- teacher distillation,

    H. Wen, L. Pan, and et al., “Class incremental learning with multi- teacher distillation,” in CVPR, 2024, pp. 28 443–28 452

    2024

  32. [40]

    Let all be whitened: Multi-teacher distillation for efficient visual retrieval,

    Z. Ma, J. Dong, and et al., “Let all be whitened: Multi-teacher distillation for efficient visual retrieval,” in AAAI, vol. 38, no. 5, 2024, pp. 4126–4135

    2024

  33. [41]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

    2015 arXiv

  34. [42]

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

    C. Zhang, G. Huang, and et al., “Webuav-3m: A benchmark for unveiling the power of million-scale deep uav tracking,” IEEE TPAMI, vol. 45, no. 7, pp. 9186–9205, 2022

    2022

  35. [43]

    Onboard real-time aerial tracking with efficient siamese anchor proposal network,

    C. Fu, Z. Cao, Y . Li, J. Ye, and C. Feng, “Onboard real-time aerial tracking with efficient siamese anchor proposal network,” IEEE TGRS, vol. 60, pp. 1–13, 2021

    2021

  36. [44]

    A benchmark and simulator for uav tracking,

    M. Mueller, N. Smith, and B. Ghanem, “A benchmark and simulator for uav tracking,” in ECCV, 2016

    2016

  37. [45]

    Learning residue-aware correlation filters and refining scale estimates with the grabcut for real-time uav tracking,

    S. Li, Y . Liu, and et al., “Learning residue-aware correlation filters and refining scale estimates with the grabcut for real-time uav tracking,” in 3DV, 2021, pp. 1238–1248

    2021

  38. [46]

    Siamese anchor proposal network for high- speed aerial tracking,

    C. Fu, Z. Cao, and et al., “Siamese anchor proposal network for high- speed aerial tracking,” in ICRA, 2021, pp. 510–516

    2021

  39. [47]

    Rank-based filter pruning for real-time uav tracking,

    X. Wang, D. Zeng, and et al., “Rank-based filter pruning for real-time uav tracking,” in ICME, 2022

    2022

  40. [48]

    Learning tracking representations via dual-branch fully transformer networks,

    F. Xie, C. Wang, and et al., “Learning tracking representations via dual-branch fully transformer networks,” in ICCVW, 2021

    2021

  41. [49]

    Correlation-aware deep tracking,

    F. Xie, C. Wang, G. Wang, Y . Cao, W. Yang, and W. Zeng, “Correlation-aware deep tracking,” in CVPR, 2022

    2022

  42. [50]

    Learning target-aware vision transformers for real-time uav tracking,

    S. Li, X. Yang, and et al., “Learning target-aware vision transformers for real-time uav tracking,” IEEE TGRS , 2024

    2024

  43. [51]

    Efficient template distinction modeling tracker with temporal contexts for aerial tracking,

    H. Zhang, W. Xing, and et al., “Efficient template distinction modeling tracker with temporal contexts for aerial tracking,” IEEE TGRS , 2024

    2024

  44. [52]

    Linformer: Self-attention with linear complexity,

    S. Wang, B. Z. Li, and et al., “Linformer: Self-attention with linear complexity,” arXiv preprint arXiv:2006.04768 , 2020

    2006 arXiv

  45. [53]

    Minivit: Compressing vision transform- ers with weight multiplexing,

    J. Zhang, H. Peng, and et al., “Minivit: Compressing vision transform- ers with weight multiplexing,” in CVPR, 2022

    2022

  46. [54]

    Efficientformer: Vision transformers at mobilenet speed,

    Y . Li, G. Yuan, and et al., “Efficientformer: Vision transformers at mobilenet speed,” in NIPS, 2022

    2022

  47. [55]

    Skeleton neural networks via low-rank guided filter pruning,

    L. Yang, S. Gu, and et al., “Skeleton neural networks via low-rank guided filter pruning,” IEEE TCSVT , vol. 33, no. 12, pp. 7197–7211, 2023

    2023

  48. [56]

    Dynamicvit: Efficient vision transformers with dynamic token sparsification,

    Y . Rao, W. Zhao, and et al., “Dynamicvit: Efficient vision transformers with dynamic token sparsification,” in NIPS, 2021

    2021

  49. [57]

    A-vit: Adaptive tokens for efficient vision transformer,

    H. Yin, A. Vahdat, and et al., “A-vit: Adaptive tokens for efficient vision transformer,” in CVPR, 2022

    2022

  50. [58]

    Dual-attention network for view-invariant action recognition,

    G. A. Kumie, M. A. Habtie, and et al., “Dual-attention network for view-invariant action recognition,” Complex & Intelligent Systems , vol. 10, no. 1, pp. 305–321, 2024

    2024

  51. [59]

    View-invariant represen- tation of hand postures in the human lateral occipitotemporal cortex,

    S. Bracci, A. Caramazza, and M. V . Peelen, “View-invariant represen- tation of hand postures in the human lateral occipitotemporal cortex,” NeuroImage, 2018

    2018

  52. [60]

    Deepgait: A learning deep convolutional representation for view-invariant gait recognition using joint bayesian,

    C. Li, X. Min, S. Sun, W. Lin, and Z. Tang, “Deepgait: A learning deep convolutional representation for view-invariant gait recognition using joint bayesian,” Applied Sciences , 2017

    2017

  53. [61]

    View-invariant representation and recognition of actions,

    C. Rao, A. Yilmaz, and M. Shah, “View-invariant representation and recognition of actions,” IJCV, 2002

    2002

  54. [62]

    View invariant human action recognition using histograms of 3d joints,

    L. Xia, C.-C. Chen, and J. K. Aggarwal, “View invariant human action recognition using histograms of 3d joints,” in CVPRW, 2012

    2012

  55. [63]

    Advances in view-invariant human motion analysis: A review,

    X. Ji and H. Liu, “Advances in view-invariant human motion analysis: A review,” IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews) , 2010

    2010

  56. [64]

    Focalized contrastive view-invariant learning for self-supervised skeleton-based action recog- nition,

    Q. Men, E. S. Ho, H. P. Shum, and H. Leung, “Focalized contrastive view-invariant learning for self-supervised skeleton-based action recog- nition,” Neurocomputing, vol. 537, pp. 198–209, 2023

    2023

  57. [65]

    View-invariant human action recognition via view transformation network (vtn),

    L. Gao, Y . Ji, and et al., “View-invariant human action recognition via view transformation network (vtn),” IEEE TMM , vol. 24, pp. 4493– 4503, 2021

    2021

  58. [66]

    Geinet: View-invariant gait recognition using a convolutional neural network,

    K. Shiraga, Y . Makihara, and et al., “Geinet: View-invariant gait recognition using a convolutional neural network,” in ICB, 2016

    2016

  59. [67]

    Learning clothing and pose invariant 3d shape representation for long-term person re-identification,

    F. Liu, M. Kim, and et al., “Learning clothing and pose invariant 3d shape representation for long-term person re-identification,” in CVPR, 2023, pp. 19 617–19 626

    2023

  60. [68]

    Aedet: Azimuth-invariant multi-view 3d object detection,

    C. Feng, Z. Jie, and et al., “Aedet: Azimuth-invariant multi-view 3d object detection,” in CVPR, 2022

    2022

  61. [69]

    Learning from multiple teacher networks,

    S. You, C. Xu, C. Xu, and D. Tao, “Learning from multiple teacher networks,” in KDD, 2017, pp. 1285–1294

    2017

  62. [70]

    M2kd: Multi-teacher multi-modal knowledge distillation for aerial view object classification,

    Z. Lan and et al., “M2kd: Multi-teacher multi-modal knowledge distillation for aerial view object classification,” in IJCNN, 2024, pp. 1–8

    2024

  63. [71]

    Multi-teacher knowledge distillation for compressed video action recognition on deep neural networks,

    M.-C. Wu, C.-T. Chiu, and K.-H. Wu, “Multi-teacher knowledge distillation for compressed video action recognition on deep neural networks,” in ICASSP, 2019, pp. 2202–2206

    2019

  64. [72]

    Mask-guided self-distillation for visual tracking,

    L. Li, C. Chen, and et al., “Mask-guided self-distillation for visual tracking,” in ICME, 2022, pp. 1–6

    2022

  65. [73]

    Siamohot: A lightweight dual siamese network for onboard hyperspectral object tracking via joint spatial- spectral knowledge distillation,

    C. Sun, X. Wang, and et al., “Siamohot: A lightweight dual siamese network for onboard hyperspectral object tracking via joint spatial- spectral knowledge distillation,” IEEE TGRS , 2023

    2023

  66. [74]

    Model compression,

    C. Bucilu ˇa, R. Caruana, and A. Niculescu-Mizil, “Model compression,” in KDD, 2006, pp. 535–541

    2006

  67. [75]

    Do deep nets really need to be deep?

    J. Ba and R. Caruana, “Do deep nets really need to be deep?” NIPS, vol. 27, 2014

    2014

  68. [76]

    Multiple teacher distillation for robust and greener models,

    A. Ilichev, N. Sorokin, and et al., “Multiple teacher distillation for robust and greener models,” in RANLP, 2021, pp. 601–610

    2021

  69. [77]

    Collaborative learning for deep neural net- works,

    G. Song and W. Chai, “Collaborative learning for deep neural net- works,” NIPS, vol. 31, 2018

    2018

  70. [78]

    D. J. MacKay, Information theory, inference and learning algorithms . Cambridge university press, 2004

    2004

  71. [79]

    On variational bounds of mutual information,

    B. Poole, S. Ozair, A. Van Den Oord, A. Alemi, and G. Tucker, “On variational bounds of mutual information,” in ICML, 2019

    2019

  72. [80]

    Comparing kullback- leibler divergence and mean squared error loss in knowledge distilla- tion,

    T. Kim, J. Oh, N. Kim, S. Cho, and S.-Y . Yun, “Comparing kullback- leibler divergence and mean squared error loss in knowledge distilla- tion,” arXiv preprint arXiv:2105.08919 , 2021

    2021 arXiv

  73. [81]

    How to train the teacher model for effective knowledge distillation,

    S. M. Hamidi, X. Deng, R. Tan, L. Ye, and A. H. Salamah, “How to train the teacher model for effective knowledge distillation,” arXiv preprint arXiv:2407.18041, 2024

    2024 arXiv

  74. [82]

    Cornernet: Detecting objects as paired key- points,

    H. Law and J. Deng, “Cornernet: Detecting objects as paired key- points,” in ECCV, 2018

    2018

  75. [83]

    Generalized intersection over union: A metric and a loss for bounding box regression,

    S. H. Rezatofighi, N. Tsoi, and et al., “Generalized intersection over union: A metric and a loss for bounding box regression,” in CVPR, 2019

    2019

  76. [84]

    Visual object tracking for unmanned aerial vehicles: A benchmark and new motion models,

    S. Li and D. Y . Yeung, “Visual object tracking for unmanned aerial vehicles: A benchmark and new motion models,” in AAAI, 2017

    2017

  77. [85]

    The unmanned aerial vehicle benchmark: Object detection and tracking,

    D. Du, Y . Qi, and et al., “The unmanned aerial vehicle benchmark: Object detection and tracking,” in ECCV, 2018

    2018

  78. [86]

    Visdrone-sot2018: The vision meets drone single-object tracking challenge results,

    P. Zhu, L. Wen, and et al., “Visdrone-sot2018: The vision meets drone single-object tracking challenge results,” in ECCV, 2018

    2018

  79. [87]

    Discriminative scale space track- ing,

    M. Danelljan, G. Hager, and et al., “Discriminative scale space track- ing,” IEEE TPAMI, 2017

    2017

  80. [88]

    Eco: Efficient convolution operators for tracking,

    M. Danelljan, G. Bhat, and et al., “Eco: Efficient convolution operators for tracking,” in CVPR, 2017

    2017

  81. [89]

    Multi-cue correlation filters for robust visual tracking,

    N. Wang, W. gang Zhou, and et al., “Multi-cue correlation filters for robust visual tracking,” in CVPR, 2018

    2018

  82. [90]

    Learning spatial-temporal regularized correlation filters for visual tracking,

    F. Li, C. Tian, and et al., “Learning spatial-temporal regularized correlation filters for visual tracking,” in CVPR, 2018, pp. 4904–4913

    2018

  83. [91]

    Learning residue-aware correlation filters and refining scale for real-time uav tracking,

    S. Li, Y . Liu, and et al., “Learning residue-aware correlation filters and refining scale for real-time uav tracking,” PR, 2022

    2022

  84. [92]

    Towards real-world visual tracking with temporal contexts,

    Z. Cao, Z. Huang, and et al., “Towards real-world visual tracking with temporal contexts,” IEEE TPAMI, 2023

    2023

  85. [93]

    Sgdvit: Saliency-guided dynamic vision transformer for uav tracking,

    L. Yao, C. Fu, and et al, “Sgdvit: Saliency-guided dynamic vision transformer for uav tracking,” in ICRA, 2023

    2023

  86. [94]

    Adversarial blur-deblur network for robust uav tracking,

    H. Zuo, C. Fu, and et al., “Adversarial blur-deblur network for robust uav tracking,” IEEE RAL , 2023

    2023

  87. [95]

    Towards discriminative representations with contrastive instances for real-time uav tracking,

    D. Zeng, M. Zou, and et al., “Towards discriminative representations with contrastive instances for real-time uav tracking,” in ICME, 2023

    2023

  88. [96]

    Progressive representation learning for real-time uav tracking,

    C. Fu, X. Lei, H. Zuo, L. Yao, G. Zheng, and J. Pan, “Progressive representation learning for real-time uav tracking,” arXiv preprint arXiv:2409.16652, 2024

    2024 arXiv

  89. [97]

    Litetrack: Layer pruning with asyn- chronous feature extraction for lightweight and efficient visual track- ing,

    Q. Wei, B. Zeng, and et al., “Litetrack: Layer pruning with asyn- chronous feature extraction for lightweight and efficient visual track- ing,” in ICRA, 2024, pp. 4968–4975. 15

    2024

  90. [98]

    Separable self and mixed attention transformers for efficient object tracking,

    G. Y . Gopal and et al., “Separable self and mixed attention transformers for efficient object tracking,” in WACV, 2024, pp. 6708–6717

    2024

  91. [99]

    Lightweight full- convolutional siamese tracker,

    Y . Li, B. Wang, X. Wu, Z. Liu, and Y . Li, “Lightweight full- convolutional siamese tracker,” KBS, vol. 286, p. 111439, 2024

    2024

  92. [100]

    Correlation-embedded trans- former tracking: A single-branch framework,

    F. Xie, W. Yang, C. Wang, and et al., “Correlation-embedded trans- former tracking: A single-branch framework,” IEEE TPAMI, vol. 46, no. 12, pp. 10 681–10 696, 2024

    2024

  93. [101]

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

    A. Dosovitskiy, L. Beyer, and A. K. andet al., “An image is worth 16x16 words: Transformers for image recognition at scale,” in ICLR, 2021

    2021

  94. [102]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, and et al., “Training data-efficient image transformers & distillation through attention,” in ICML, 2021

    2021

  95. [103]

    Eva: Exploring the limits of masked visual representation learning at scale,

    Y . Fang, W. Wang, and et al., “Eva: Exploring the limits of masked visual representation learning at scale,” in CVPR, 2023

    2023

  96. [104]

    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, 2021

    2021

  97. [105]

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

    H. Fan, L. Lin, F. Yang, and et al., “Lasot: A high-quality benchmark for large-scale single object tracking,” in CVPR, 2019

    2019

  98. [106]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, and et al., “Microsoft coco: Common objects in context,” in ECCV, 2014

    2014

  99. [107]

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

    M. Muller, A. Bibi, S. Giancola, and et al., “Trackingnet: A large-scale dataset and benchmark for object tracking in the wild,” in ECCV, 2018

    2018

  100. [108]

    Ocean: Object-aware anchor-free tracking,

    Z. Zhang, H. Peng, and et al., “Ocean: Object-aware anchor-free tracking,” in ECCV, 2020

    2020

  101. [109]

    Learning discriminative model prediction for tracking,

    G. Bhat, M. Danelljan, and et al., “Learning discriminative model prediction for tracking,” in ICCV, 2019, pp. 6182–6191

    2019

  102. [110]

    Probabilistic regression for visual tracking,

    M. Danelljan, L. V . Gool, and R. Timofte, “Probabilistic regression for visual tracking,” CVPR, pp. 7181–7190, 2020

    2020

  103. [111]

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

    N. Wang, W. Zhou, and et al., “Transformer meets tracker: Exploiting temporal context for robust visual tracking,” in CVPR, 2021, pp. 1571– 1580

    2021

  104. [112]

    Transformer tracking,

    X. Chen, B. Yan, J. Zhu, and et al., “Transformer tracking,” in CVPR, 2021, pp. 8126–8135

    2021

  105. [113]

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

    Z. Zhang, Y . Liu, X. Wang, B. Li, and W. Hu, “Learn to match: Automatic matching network design for visual tracking,” ICCV, pp. 13 319–13 328, 2021

    2021

  106. [114]

    Sparsett: Visual tracking with sparse transformers,

    Z. Fu, Z. Fu, Q. Liu, W. Cai, and Y . Wang, “Sparsett: Visual tracking with sparse transformers,” arXiv e-prints , 2022

    2022

  107. [115]

    Transformer tracking with cyclic shifting window attention,

    Z. Song, J. Yu, and et al., “Transformer tracking with cyclic shifting window attention,” in CVPR, 2022, pp. 8791–8800

    2022

  108. [116]

    Seqtrack: Sequence to sequence learning for visual object tracking,

    X. C. et al., “Seqtrack: Sequence to sequence learning for visual object tracking,” ArXiv, vol. abs/2304.14394, 2023

    2023 arXiv

  109. [117]

    Representation learning for visual object tracking by masked appearance transfer,

    H. Zhao, D. Wang, and H. Lu, “Representation learning for visual object tracking by masked appearance transfer,” in CVPR, 2023, pp. 18 696–18 705

    2023

  110. [118]

    Robust object modeling for visual tracking,

    Y . Cai, J. Liu, J. Tang, and G. Wu, “Robust object modeling for visual tracking,” in ICCV, 2023, pp. 9589–9600

    2023

  111. [119]

    Dcpt: Darkness clue-prompted tracking in nighttime uavs,

    J. Zhu, H. Tang, and et al., “Dcpt: Darkness clue-prompted tracking in nighttime uavs,” in ICRA, 2024, pp. 7381–7388

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.