Pith. sign in

REVIEW 5 major objections 6 minor 62 references

CGTrack: Cascade Gating Network with Hierarchical Feature Aggregation for UAV Tracking

T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read CGTrack claims state-of-the-art accuracy on three UAV tracking benchmarks using a cascade gating design that fuses hierarchical features at negligible extra cost.

desk verdict The headline results are internally inconsistent (text vs. plots disagree on two benchmarks), so the SOTA claim is not verifiable as written; the architecture is worth a careful look once the numbers are reconciled. read the letter →

arxiv 2505.05936 v1 pith:IAE5OQKQ submitted 2025-05-09 cs.CV

Weihong Li , Xiaoqiong Liu , Heng Fan , Libo Zhang This is my paper · ORCID
classification cs.CV
keywords UAVtrackingvisualobjecthierarchicalfeaturefusiongatingmechanismlightweightvisiontransformerone-streamtrackerreusereal-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

Unmanned-aerial-vehicle (UAV) tracking must run in real time on limited hardware, which pushes designers toward lightweight networks that often lose the fine detail needed to keep a target through occlusion, low resolution, and sharp viewpoint changes. The paper introduces CGTrack, a one-stream tracker—template and search image are processed jointly in a single network—built on a lightweight hierarchical vision transformer, and argues that two additions close that gap: a Hierarchical Feature Cascade (HFC) module that concatenates multi-scale features and applies residual channel gating, and a Lightweight Gated Center Head (LGCH) that uses Hadamard-product gating to extract target-specific detail from the expanded features. The claim is that this concatenate-then-gate design expands network capacity with almost no added parameters or FLOPs, and the paper reports state-of-the-art numbers on UAV123, UAV123@10fps, and UAVTrack112, with the largest variant running at 42.1 frames per second. If true, the work points to feature reuse and gating as a cheap way to get hierarchical fusion benefits in resource-constrained trackers rather than relying on heavier transformer fusion.

What carries the argument

The load-bearing mechanism is the Hierarchical Feature Cascade (HFC) module: it upsamples each backbone stage's correlation map, concatenates adjacent scales along the channel dimension, and passes the output through a Residual Squeeze-and-Excitation block that rescales channels by a learned gating vector, then repeats the cascade once more. Concatenation is the key operation—it preserves every channel's information where addition would force a lossy sum, and the residual SE applies coarse gating at minimal cost. On top of the HFC runs the Lightweight Gated Center Head (LGCH), the final head that predicts center location and bounding-box size; its Efficient Gating (EG) blocks split an input into a gate branch and a context branch through parallel $1\times1$ convolutions, pass one branch through ReLU6, and combine them with an element-wise (Hadamard) product, mapping the features into a higher-dimensional nonlinear space while computing in the low-dimensional one. The backbone provides three hierarchical correlation maps from a lightweight LeViT, and the loss is the standard one-stream combination of focal, GIoU, and $\ell_1$ terms, so the HFC-plus-LGCH pairing is what carries the argument.

What would settle it

A reader could settle it by training CGTrack-B on the four stated training sets and running the same evaluation scripts on UAV123; if CGTrack's precision and success do not reach the reported 88.0% and 67.2%, or if any of the 13 baselines, re-run under the same script, match or exceed those numbers, the state-of-the-art claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a cascade of feature concatenation and gating, rather than addition-based or transformer-based fusion, lets a lightweight hierarchical ViT compete with heavier trackers in aerial scenarios. The paper reports that CGTrack reaches 88.0% precision and 67.2% success on UAV123, 86.2% precision and 66.1% success on UAV123@10fps, and 82.4% precision and 64.9% success on UAVTrack112, surpassing the 13 published trackers it is compared with in Fig. 5. The design rationale is explicit: concatenation preserves the rich spatial detail of shallow stages alongside semantic depth, Residual Squeeze-and-Excitation blocks reweight the channels, and the gated center head decouples target-oriented coordinates in a higher-dimensional nonlinear feature space. Table I's ablation on UAV123@10fps shows the full cascade-gating configuration 3.34 precision points above addition-based fusion and 2.12 points above concatenation without residual SE, which is the paper's main evidence that the HFC mechanism, not just the backbone, drives the improvement.

Load-bearing premise

The state-of-the-art conclusion rests on the assumption that the published scores of the 13 compared trackers were produced under the same training data, resolution, and evaluation protocol as CGTrack; if any baseline came from a different protocol, the reported margins would not be a fair measure of the architecture.

Editorial extensions

If this is right

  • If the reported numbers hold, a lightweight one-stream tracker can reach state-of-the-art aerial accuracy without the transformer-based fusion layers that heavier UAV trackers use, since CGTrack-B's 4.324 GMACs keep it in real-time range.
  • The HFC module is a drop-in style intervention: swapping additive fusion for concatenation-plus-residual-gating is what Table I credits for a 3.34-point precision gain on UAV123@10fps, so the same swap should transfer to other trackers built on hierarchical backbones.
  • The three variants, CGTrack-T, -S, and -B, span 61.4 to 42.1 fps with precision from 80.08 to 86.24 (Table II), giving practitioners a direct speed-accuracy trade-off for different drone-hardware budgets.
  • Because CGTrack trains only on GOT-10k, LaSOT, COCO, and TrackingNet, its evaluation numbers are not in-domain, so the reported SOTA is on unseen aerial test sets rather than benchmarks the model has seen.

Reading between the lines

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

  • A plausible generalization is that the same concat-then-gate cascade could serve as a universal fusion block in other one-stream trackers or multi-scale detection heads, and the strongest test would be reproducing the Table I gain outside this architecture.
  • The attribute results suggest a testable prediction the paper does not fully exploit: the gated cascade should matter most on small-object and occlusion attributes, so a controlled study varying object size could isolate where HFC's contribution comes from.
  • A reader should note an internal inconsistency: Section IV-B's prose reports 83.8% precision on UAV123@10fps and 80.6% on UAVTrack112, while Fig. 5 plots 86.2% and 82.4%; the plotted values should be treated as the canonical claim.
  • If the HFC design transfers, the same fusion trick could be applied to other resource-constrained multi-scale tasks like detection or segmentation, where concatenation and gating are usually studied separately.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes CGTrack, a family of one-stream UAV trackers built on a lightweight hierarchical LeViT backbone. The two main architectural contributions are a Hierarchical Feature Cascade (HFC) module, which concatenates multi-scale backbone features and applies residual Squeeze-and-Excitation gating, and a Lightweight Gated Center Head (LGCH) whose Efficient Gating (EG) blocks use Hadamard-product gating in an expanded feature space. The authors report state-of-the-art results on UAV123, UAV123@10fps, and UAVTrack112, with the largest variant running at 42.1 fps, and support the design with ablations on fusion strategy, model scale, and head configuration. As written, however, the headline numbers are internally inconsistent: Section IV-B reports different precision values from Fig. 5 on two of the three benchmarks, and one ablation increment in Table I is arithmetically wrong. These issues prevent the central SOTA claim from being verified in the current version.

Significance. If the reported results hold, the paper makes a useful and practical contribution: it shows that a simple concatenation-plus-gating fusion of hierarchical ViT features, together with a lightweight gated center head, can exceed more complex fusion designs on challenging UAV benchmarks while remaining real-time. The ablation structure is sensible, and the idea of reusing hierarchical features through cascade concatenation is well motivated by DenseNet. The paper also evaluates on three standard UAV benchmarks and compares with 13 baselines, which is a reasonable scope. The main weakness is that the central quantitative evidence is not currently trustworthy because of the numerical inconsistencies described below; the significance of the architectural claim can only be assessed after those numbers are reconciled and the evaluation protocol is clarified.

major comments (5)
  1. [Section IV-B, Fig. 5] The headline SOTA numbers are internally inconsistent. Section IV-B reports UAV123@10fps Precision 83.8% and UAVTrack112 Precision 80.6%, while the precision plots in Fig. 5 label CGTrack as 86.2% and 82.4% on the same benchmarks, respectively. Only UAV123 (88.0%) is consistent. The difference is not cosmetic: with the text value, CGTrack would trail TCTrack++ and SiamAPN (both 81.4) on UAVTrack112, and the UAV123@10fps margin over SiamTPN would shrink from 6.3 to 3.9 points. Please correct the tables/text/figures to one consistent set of numbers and, if possible, release per-sequence results or raw logs so the reader can verify them.
  2. [Section IV-D, Table I] The claimed gain of the Residual SE gating is misreported. The text states that Row 3 exhibits a 3.34% increase in Precision over Row 2, but 86.24 - 84.12 = 2.12 percentage points. The 1.22% gain from Row 1 to Row 2 is correct (84.12 - 82.90). Please correct the stated increment and re-evaluate the conclusion that cascade gating is the main driver of the HFC improvement.
  3. [Section IV-D, Table III] The final model configuration appears to be selected on the same benchmark used for the headline results. Section IV-D states that all ablations are performed on UAV123@10fps, and Table III selects the EG upsampling ratio 2 by the highest AUC on that benchmark, while Fig. 5 reports CGTrack's UAV123@10fps success and precision on the same dataset. If the selection and the reported result share the same test set, the reported SOTA numbers are optimistically biased. Please state explicitly whether the ablations were done on a held-out validation split and whether the final evaluation is on a disjoint test protocol.
  4. [Fig. 5, Section IV-A] Comparability of the 13 baseline numbers is not established. The manuscript does not state, for each baseline, the training data, template/search resolution, and evaluation protocol used to produce the plotted scores, nor does it say whether all baselines were re-run under a common codebase or quoted from the original papers. Because several UAV trackers are trained on different data (some on UAV123 itself) and CGTrack is trained on GOT-10k/LaSOT/COCO/TrackingNet, the claimed margins (e.g., 88.0 vs 82.3 precision on UAV123) may reflect protocol differences rather than architectural merit. Please provide a table of baseline settings or re-evaluate all methods in a common protocol.
  5. [Section III-C] The efficiency claim 'without additional parameters or FLOPs' for the HFC module is contradicted by the paper's own ablation. Table I shows that adding Residual SE (the gating part of HFC) increases parameters from 40.668M to 41.219M and MACs from 4.323 to 4.324G; the Residual SE block itself contains FC layers, as shown in Eqs. (2)-(5). Please restate the claim to refer only to the concatenation step, or provide a corrected analysis.
minor comments (6)
  1. [Throughout] The spacing 'UA V' appears in the title, abstract, and main text (e.g., 'UA V Tracking', 'UA V scenarios'); it should be 'UAV'.
  2. [Table III] The row numbering is duplicated: both 'EG block-3x' and 'EG block-4x' are labeled as row 4. The rows should be numbered 1 through 5.
  3. [Eq. (5)] The symbol S in Eq. (5) is used without a definition; presumably it denotes the channel-wise scaling tensor whose entries are s_c from Eq. (4), but this should be stated explicitly.
  4. [Section II-C] The heading 'Gating Mechenism' contains a typo; it should be 'Gating Mechanism'.
  5. [Fig. 5 caption] The caption writes 'UA V123@10fp' in the second column label; this should be 'UAV123@10fps'.
  6. [Section IV-A] The phrase 'excluding 1k sequences as convention' is ambiguous; please clarify whether this refers to the standard GOT-10k validation split or to some other exclusion.

Circularity Check

1 steps flagged · score 6.0 of 10

Final configuration selected by AUC on UAV123@10fps is then reported as SOTA on the same benchmark, making that prediction partly circular; the other two benchmarks remain independent.

  1. fitted input called prediction [Section IV-D (Ablation Study), Tables I and III; Section IV-B (UAV123@10fps results)]
    "In this subsection, we present the ablation studies on UA V123@10fps. ... When setting the upsampling ratio to 2, the highest Success score is achieved."

    The ablation studies in Tables I and III are explicitly run on UAV123@10fps, and the grey 'final configuration' is chosen by the best Precision/AUC on that same benchmark. Section IV-B then reports CGTrack achieving 'SOTA performance, with the highest Precision (83.8%) and Success score (66.1%)' on UAV123@10fps. Thus the UAV123@10fps headline number is not an independent evaluation of a fixed architecture: the HFC fusion design, the EG upsampling ratio, and the head configuration were selected using the test benchmark, so the SOTA claim on that benchmark is in part forced by the selection criterion. UAV123 and UAVTrack112 were not used for these ablations, so the architecture claim retains independent content there.

full rationale

The architecture derivation itself is self-contained: the HFC concatenation (Eq. 1, 6-7), Residual SE (Eq. 2-5), EG block (Eq. 8-12), and training loss (Eq. 13) are defined from the paper's own components and external works like DenseNet, SENet, and StarNet, with no uniqueness theorem or load-bearing self-citation. The only self-citation is the LaSOT dataset reference, which is a training resource, not an argument. The circular element is confined to evaluation: Section IV-D states that ablations are performed on UAV123@10fps, and the final configuration is selected by the best score on that benchmark (highest AUC in Table III, gray row; highest Precision in Table I). Section IV-B then presents UAV123@10fps precision/success as part of the SOTA claim. This is a fitted-input-called-prediction pattern for that benchmark. Separately, there is a correctness/verifiability defect not scored here as circularity: Section IV-B reports UAV123@10fps Precision 83.8% and UAVTrack112 Precision 80.6%, while Fig. 5 labels 86.2% and 82.4% for the same quantities, and the Table I ablation arithmetic (86.24-84.12=2.12) contradicts the stated 3.34% gain. These inconsistencies mean the exact SOTA margins cannot be verified from the paper as written, but they are not a derivation-level circularity.

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

CGTrack is built from published components: LeViT/HiT for the one-stream backbone, DenseNet-style concatenation, SENet-style channel gating, and StarNet-style Hadamard gating. The empirical claims rest on inherited and hand-chosen hyperparameters, an ablation-selected upsampling ratio, and the unstated assumption that published baseline numbers are protocol-comparable. No new physical entities are introduced; the named modules are the method itself, not explanatory props.

free parameters (4)
  • EG upsampling ratio = 2x
    Selected in Table III by AUC on UAV123@10fps, the same benchmark where SOTA is later claimed; the choice is tuned to the evaluation target.
  • Loss weights lambda_G and lambda_l = 2 and 5
    Adopted from OSTrack [15] as stated after Eq. (13); not fitted here, but the final result depends on these hand-chosen values.
  • Training hyperparameters = LR 4e-5, weight decay 1e-4, batch size 128, 10% LR decay in final 20% of epochs
    Reported in Section IV-A; epoch count and warmup are unstated, so the exact recipe is incomplete.
  • Flagship variant selection = CGTrack-B (LeViT-384)
    The headline SOTA numbers correspond to the largest variant in Table II; the variant choice is made on the benchmarks and the paper does not show which variant produced Fig. 5's curves.
assumptions (4)
  • domain assumption LeViT, adapted via the HiT conventions (position encoding, one-stream pipeline, training recipe), provides hierarchical features suitable for tracking.
    Section III-B explicitly defers to LeViT [47] and HiT [32] for the backbone and its adaptation; CGTrack inherits all their assumptions.
  • domain assumption The quoted baseline numbers from the 13 compared trackers are protocol-comparable to CGTrack's runs.
    Section IV-B and Fig. 5 quote prior published scores without specifying each tracker's training data or evaluation protocol; mixed protocols would invalidate the SOTA margins.
  • domain assumption One-pass benchmark evaluation on UAV123, UAV123@10fps, and UAVTrack112 is the accepted evidence for tracking quality.
    Section IV-B treats these three benchmarks as authoritative; the claim type is empirical benchmark performance, not a theoretical bound.
  • domain assumption Benchmark annotations are treated as ground truth and the training splits used (GOT-10k, LaSOT, COCO, TrackingNet) have no leakage into the test benchmarks.
    Section IV-A lists training data; the UAV benchmarks are not used for training, but the paper does not discuss annotation bias or protocol details.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CGTrack: Cascade Gating Network with Hierarchical Feature Aggregation for UAV Tracking." pith.science (2026). https://pith.science/paper/IAE5OQKQ

@misc{pith2026250505936,
  author       = {Pith},
  title        = {Pith review of: CGTrack: Cascade Gating Network with Hierarchical Feature Aggregation for UAV Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IAE5OQKQ}},
  note         = {Machine review of arXiv:2505.05936}
}
read the original abstract

Recent advancements in visual object tracking have markedly improved the capabilities of unmanned aerial vehicle (UAV) tracking, which is a critical component in real-world robotics applications. While the integration of hierarchical lightweight networks has become a prevalent strategy for enhancing efficiency in UAV tracking, it often results in a significant drop in network capacity, which further exacerbates challenges in UAV scenarios, such as frequent occlusions and extreme changes in viewing angles. To address these issues, we introduce a novel family of UAV trackers, termed CGTrack, which combines explicit and implicit techniques to expand network capacity within a coarse-to-fine framework. Specifically, we first introduce a Hierarchical Feature Cascade (HFC) module that leverages the spirit of feature reuse to increase network capacity by integrating the deep semantic cues with the rich spatial information, incurring minimal computational costs while enhancing feature representation. Based on this, we design a novel Lightweight Gated Center Head (LGCH) that utilizes gating mechanisms to decouple target-oriented coordinates from previously expanded features, which contain dense local discriminative information. Extensive experiments on three challenging UAV tracking benchmarks demonstrate that CGTrack achieves state-of-the-art performance while running fast. Code will be available at https://github.com/Nightwatch-Fox11/CGTrack.

Figures

Figures reproduced from arXiv: 2505.05936 by the authors.

Figure 1
Figure 1. Comparison of success rate and precision between CGTrack [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the popular hierarchical feature fusion [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed CGTrack, which comprises three main components: a lightweight hierarchical backbone, an HFC module, [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Detailed architectures of LGCH. The left part illustrates the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Overall performance of CGTrack and prevailing SOTA trackers on UAV123 [ [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of CGTrack with other trackers on [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 57 canonical work pages

  1. [1]

    Ieee tsmc,

    S. Kim and I. Moon, “Ieee tsmc,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 49, no. 1, pp. 42–52, 2019

    work page 2019

  2. [2]

    Artificial noise aided scheme to secure uav-assisted internet of things with wireless power transfer,

    Q. Wang, H. N. Dai, X. Li, M. K. Shukla, and M. Imran, “Artificial noise aided scheme to secure uav-assisted internet of things with wireless power transfer,”Computer Communications, vol. 164, 2020

    work page 2020

  3. [3]

    Control of an aerial manipulator using a quadrotor with a replaceable robotic arm,

    Z. Ouyang, R. Mei, Z. Liu, M. Wei, Z. Zhou, and H. Cheng, “Control of an aerial manipulator using a quadrotor with a replaceable robotic arm,” inICRA. IEEE, 2021, pp. 153–159

    work page 2021

  4. [4]

    High-speed tracking with kernelized correlation filters,

    J. F. Henriques, R. Caseiro, P. Martins, and J. Batista, “High-speed tracking with kernelized correlation filters,”IEEE TPAMI, vol. 37, no. 3, pp. 583–596, 2014

    work page 2014

  5. [5]

    ECO: Efficient Convolution Operators for Tracking,

    M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg, “ECO: Efficient Convolution Operators for Tracking,” inCVPR, 2017, pp. 6931–6939

    work page 2017

  6. [6]

    Atom: Accurate tracking by overlap maximization,

    ——, “Atom: Accurate tracking by overlap maximization,” inCVPR, 2019, pp. 4660–4669

    work page 2019

  7. [7]

    Learning Discriminative Model Prediction for Tracking,

    G. Bhat, M. Danelljan, L. V . Gool, and R. Timofte, “Learning Discriminative Model Prediction for Tracking,” inICCV, 2019, pp. 6181–6190

    work page 2019

  8. [8]

    Probabilistic Regression for Visual Tracking,

    M. Danelljan, L. V . Gool, and R. Timofte, “Probabilistic Regression for Visual Tracking,” inCVPR, 2020, pp. 7181–7190

    work page 2020

Show all 62 references
  1. [9]

    Know your surroundings: Exploiting scene information for object tracking,

    G. Bhat, M. Danelljan, L. Van Gool, and R. Timofte, “Know your surroundings: Exploiting scene information for object tracking,” in ECCV. Springer, 2020, pp. 205–221

    2020

  2. [10]

    Fully-Convolutional Siamese Networks for Object Tracking,

    L. Bertinetto, J. Valmadre, J. F. Henriques, A. Vedaldi, and P. H. S. Torr, “Fully-Convolutional Siamese Networks for Object Tracking,” in ECCV, 2016, pp. 850–865

    2016

  3. [11]

    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,” inCVPR, 2018, pp. 8971–8980

    2018

  4. [12]

    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

  5. [13]

    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,” inICCV, 2021, pp. 10 428–10 437

    2021

  6. [14]

    Transformer Tracking,

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

    2021

  7. [15]

    Joint Feature Learning and Relation Modeling for Tracking: A One-Stream Framework,

    B. Ye, H. Chang, B. Ma, S. Shan, and X. Chen, “Joint Feature Learning and Relation Modeling for Tracking: A One-Stream Framework,” in ECCV, 2022, pp. 341–357

    2022

  8. [16]

    Seqtrack: Sequence to sequence learning for visual object tracking,

    X. Chen, H. Peng, D. Wang, H. Lu, and H. Hu, “Seqtrack: Sequence to sequence learning for visual object tracking,” inCVPR, 2023, pp. 14 572–14 581

    2023

  9. [17]

    SiamFC++: Towards Ro- bust and Accurate Visual Tracking with Target Estimation Guidelines

    Y . Xu, Z. Wang, Z. Li, Y . Yuan, and G. Yu, “SiamFC++: Towards Ro- bust and Accurate Visual Tracking with Target Estimation Guidelines.” inAAAI, 2020, pp. 12 549–12 556

    2020

  10. [18]

    SiamCAR: Siamese Fully Convolutional Classification and Regression for Visual Tracking,

    D. Guo, J. Wang, Y . Cui, Z. Wang, and S. Chen, “SiamCAR: Siamese Fully Convolutional Classification and Regression for Visual Tracking,” inCVPR, 2020, pp. 6268–6276

    2020

  11. [19]

    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 CVPR, 2019, pp. 4282–4291

    2019

  12. [20]

    Fast Online Object Tracking and Segmentation: A Unifying Approach,

    Q. Wang, L. Zhang, L. Bertinetto, W. Hu, and P. H. S. Torr, “Fast Online Object Tracking and Segmentation: A Unifying Approach,” in CVPR, 2019, pp. 1328–1338

    2019

  13. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in NIPS, 2017, pp. 5998–6008

    2017

  14. [22]

    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,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 457– 15 466

    2021

  15. [23]

    SGDViT: saliency-guided dynamic vision transformer for uav tracking,

    L. Yao, C. Fu, S. Li, G. Zheng, and J. Ye, “SGDViT: saliency-guided dynamic vision transformer for uav tracking,” inICRA. IEEE, 2023, pp. 3353–3359

    2023

  16. [24]

    Continuity-aware latent interframe information mining for reliable uav tracking,

    C. Fu, M. Cai, S. Li, K. Lu, H. Zuo, and C. Liu, “Continuity-aware latent interframe information mining for reliable uav tracking,” in ICRA. IEEE, 2023, pp. 1327–1333

    2023

  17. [25]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” inICLR, 2021

    2021

  18. [26]

    Learning Transferable Visual Models From Natural Language Supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning Transferable Visual Models From Natural Language Supervision,” inICML, 2021, pp. 8748–8763

    2021

  19. [27]

    Context autoencoder for self-supervised representation learning,

    X. Chen, M. Ding, X. Wang, Y . Xin, S. Mo, Y . Wang, S. Han, P. Luo, G. Zeng, and J. Wang, “Context autoencoder for self-supervised representation learning,”International Journal of Computer Vision, pp. 1–16, 2023

    2023

  20. [28]

    Emerging properties in self-supervised vision trans- formers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision trans- formers,” inICCV, 2021

    2021

  21. [29]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  22. [30]

    Backbone is All Your Need: A Simplified Architecture for Visual Object Tracking,

    B. Chen, P. Li, L. Bai, L. Qiao, Q. Shen, B. Li, W. Gan, W. Wu, and W. Ouyang, “Backbone is All Your Need: A Simplified Architecture for Visual Object Tracking,” inECCV, 2022, pp. 375–392

    2022

  23. [31]

    Robust object modeling for visual tracking,

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

    2023

  24. [32]

    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,” inICCV, 2023

    2023

  25. [33]

    Conformer: Local features coupling global representations for visual recognition,

    Z. Peng, W. Huang, S. Gu, L. Xie, Y . Wang, J. Jiao, and Q. Ye, “Conformer: Local features coupling global representations for visual recognition,” inICCV, 2021, pp. 367–376

    2021

  26. [34]

    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

  27. [35]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” inCVPR, 2017

    2017

  28. [36]

    Generalized relation modeling for transformer tracking,

    S. Gao, C. Zhou, and J. Zhang, “Generalized relation modeling for transformer tracking,” inCVPR, 2023, pp. 18 686–18 695

    2023

  29. [37]

    Visual prompt multi- modal tracking,

    J. Zhu, S. Lai, X. Chen, D. Wang, and H. Lu, “Visual prompt multi- modal tracking,” inCVPR, 2023, pp. 9516–9526

    2023

  30. [38]

    Hiptrack: Visual tracking with historical prompts,

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

    2024

  31. [39]

    Rewrite the stars,

    X. Ma, X. Dai, Y . Bai, Y . Wang, and Y . Fu, “Rewrite the stars,” in CVPR, 2024

    2024

  32. [40]

    Free-form image inpainting with gated convolution,

    J. Yu, Z. Lin, J. Yang, X. Shen, X. Lu, and T. S. Huang, “Free-form image inpainting with gated convolution,” inICCV, 2019, pp. 4471– 4480

    2019

  33. [41]

    Siamapn++: Siamese attentional aggregation network for real-time uav tracking,

    Z. Cao, C. Fu, J. Ye, B. Li, and Y . Li, “Siamapn++: Siamese attentional aggregation network for real-time uav tracking,” inIROS. IEEE, 2021, pp. 3086–3092

    2021

  34. [42]

    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,” inCVPR, 2022, pp. 14 798– 14 808

    2022

  35. [43]

    Towards real-world visual tracking with temporal contexts,

    ——, “Towards real-world visual tracking with temporal contexts,” IEEE TPAMI, 2023

    2023

  36. [44]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in CVPR, 2018, pp. 7132–7141

    2018

  37. [45]

    Moganet: Multi-order gated aggregation network,

    S. Li, Z. Wang, Z. Liu, C. Tan, H. Lin, D. Wu, Z. Chen, J. Zheng, and S. Z. Li, “Moganet: Multi-order gated aggregation network,” in ICLR, 2022

    2022

  38. [46]

    Autoregressive visual tracking,

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

    2023

  39. [47]

    LeViT: a Vision Transformer in ConvNet’s Clothing for Faster Inference,

    B. Graham, A. El-Nouby, H. Touvron, P. Stock, A. Joulin, H. J ´egou, and M. Douze, “LeViT: a Vision Transformer in ConvNet’s Clothing for Faster Inference,” inICCV, 2021, pp. 12 239–12 249

    2021

  40. [48]

    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 TGARS, vol. 60, pp. 1–13, 2021

    2021

  41. [49]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inICCV, 2017, pp. 2980–2988

    2017

  42. [50]

    Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression,

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. D. Reid, and S. Savarese, “Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression,” inCVPR, 2019, pp. 658–666

    2019

  43. [51]

    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, 2021

    2021

  44. [52]

    LaSOT: A High-Quality Benchmark for Large-Scale Single Object Tracking,

    H. Fan, L. Lin, F. Yang, P. Chu, G. Deng, S. Yu, H. Bai, Y . Xu, C. Liao, and H. Ling, “LaSOT: A High-Quality Benchmark for Large-Scale Single Object Tracking,” inCVPR, 2019, pp. 5374–5383

    2019

  45. [53]

    Microsoft COCO: Common Objects in Context,

    T.-Y . Lin, M. Maire, S. J. Belongie, L. D. Bourdev, R. B. Girshick, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft COCO: Common Objects in Context,” inECCV, 2014, pp. 740–755

    2014

  46. [54]

    TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild,

    M. Muller, A. Bibi, S. Giancola, S. Alsubaihi, and B. Ghanem, “TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild,” inECCV, 2018, pp. 310–327

    2018

  47. [55]

    Decoupled Weight Decay Regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regulariza- tion,” inICLR, 2019

    2019

  48. [56]

    End-to-end feature decontaminated network for uav tracking,

    H. Zuo, C. Fu, S. Li, J. Ye, and G. Zheng, “End-to-end feature decontaminated network for uav tracking,” inIROS. IEEE, 2022, pp. 12 130–12 137

    2022

  49. [57]

    Local perception-aware transformer for aerial tracking,

    C. Fu, W. Peng, S. Li, J. Ye, and Z. Cao, “Local perception-aware transformer for aerial tracking,” inIROS. IEEE, 2022, pp. 12 122– 12 129

    2022

  50. [58]

    Deconnet: End-to- end decontaminated network for vision-based aerial tracking,

    H. Zuo, C. Fu, S. Li, J. Ye, and G. Zheng, “Deconnet: End-to- end decontaminated network for vision-based aerial tracking,”IEEE TGARS, vol. 60, pp. 1–12, 2022

    2022

  51. [59]

    Siamese trans- former pyramid networks for real-time uav tracking,

    D. Xing, N. Evangeliou, A. Tsoukalas, and A. Tzes, “Siamese trans- former pyramid networks for real-time uav tracking,” inWACV, 2022, pp. 2139–2148

    2022

  52. [60]

    Deeper and Wider Siamese Networks for Real-Time Visual Tracking,

    Z. Zhang and H. Peng, “Deeper and Wider Siamese Networks for Real-Time Visual Tracking,” inCVPR, 2019, pp. 4591–4600

    2019

  53. [61]

    Scale equivariance improves siamese tracking,

    I. Sosnovik, A. Moskalev, and A. W. Smeulders, “Scale equivariance improves siamese tracking,” inWACV, 2021, pp. 2765–2774

    2021

  54. [62]

    Ocean: Object-aware Anchor-free Tracking,

    Z. Zhang, H. Peng, J. Fu, B. Li, and W. Hu, “Ocean: Object-aware Anchor-free Tracking,” inECCV, 2020, pp. 771–787

    2020

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.