Pith. sign in

REVIEW 3 major objections 8 minor 55 references

UMDATrack: Unified Multi-Domain Adaptive Tracking Under Adverse Weather Conditions

T0 review · 3 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read UMDATrack claims a single tracker with a frozen backbone and a small adapter can hold state-of-the-art accuracy across night, fog, and rain using only cheap synthetic training videos.

desk verdict A genuine system-level advance in multi-weather tracking, but the load-bearing OT loss is under-specified and the synthetic test sets are partly in-distribution; worth a serious referee, not acceptance as-is. read the letter →

arxiv 2507.00648 v1 pith:RLYCUSQZ submitted 2025-07-01 cs.CV

classification cs.CV
keywords visualobjecttrackingdomainadaptationadverseweathertext-to-imagesynthesisoptimaltransportstudent-teacherlearningnighttimefoggyandrainy
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

The paper claims that a single visual tracker, UMDATrack, can hold state-of-the-art accuracy across night, fog, and rain without ever seeing real adverse-weather video during training. It does so by translating a small fraction (less than 2% of source daytime frames) of GOT-10k videos into dark, foggy, and rainy clips with a text-conditioned diffusion generator, then teaching a frozen-backbone tracker to adapt through a compact domain-customized adapter and an optimal-transport-based confidence alignment loss. If the claim holds, weather robustness no longer requires per-domain model retraining or large labeled adverse-weather collections; one adaptation stage covers multiple domains at real-time speed. The paper reports leading AUC and precision on synthetic GOT-10k and DTB70 variants and on the real NAT2021, UAVDark70, and AVisT datasets.

What carries the argument

The mechanism has three parts. Controllable Scenario Generator (CSG): Stable Diffusion-Turbo with text prompts such as "Car in the night/haze/rain/snow" translates daytime GOT-10k frames into unlabeled target-domain videos in one to four steps. Domain-Customized Adapter (DCA): with the ViT backbone frozen, a lightweight ResNet block turns target search features into queries, a learnable token bank is projected to keys and values, and a structural token $S = \mathrm{Softmax}(QK^\top/\sqrt{d_k})V$ is concatenated with source template-search tokens, so the same adapter serves all weather domains. Target-Aware Confidence Alignment (TCA): confidence distributions from student and teacher response maps are compared at each sample's argmax location, with a cost combining normalized confidence difference and position difference, and the resulting position-sensitive optimal transport loss is solved by Sinkhorn. These three components together carry the claim, and the full model runs at 138 FPS with 65M parameters.

What would settle it

Hold out genuine fog or rain videos not used in training, measure the feature-space distance between CSG-generated frames and real adverse-weather frames in the frozen ViT, and compare UMDATrack's AUC with a no-adaptation baseline; if the synthetic-to-real distance is as large as the original daytime-to-adverse distance, or the AUC gain over the baseline vanishes, the domain-overlap premise is wrong.

Watch

Extended reading notes

Core claim

The central discovery is that multi-domain adverse-weather adaptation in tracking can be done with one frozen encoder, one small trainable adapter, and unlabeled synthetic videos, provided the alignment is made localization-aware. The authors build UMDATrack around the claim that the appearance shift produced by text-prompted Stable Diffusion-Turbo is close enough to real night, fog, and rain that a student-teacher EMA setup can transfer daytime tracking knowledge to all three target domains at once. The DCA adapter inserts a learned token bank and cross-attention structural tokens into the frozen ViT, and the TCA module aligns teacher and student confidence maps at their argmax positions through a position-sensitive optimal transport cost. Under this design the paper reports gains over both single-weather adaptation trackers and generic state-of-the-art trackers on synthetic benchmarks, and smaller but consistent gains on real-world night and UAV datasets.

Load-bearing premise

Everything rests on the premise that the synthetic weather made by the text-conditioned generator looks like real night, fog, and rain to the tracker's features; if that overlap is wrong, the adapter and alignment loss are adapting to the wrong domain shift.

Editorial extensions

If this is right

  • Adding a new weather condition such as snow or haze reduces to writing a new text prompt for the generator and training the small DCA adapter for roughly 50 epochs, without touching the backbone or existing adapters.
  • Deployed trackers on UAVs or dashcams could switch between dark, foggy, and rainy conditions with the same frozen backbone and the same token-bank adapter, rather than loading per-weather models.
  • The reported 138 FPS inference speed with 18 G MACs and 65M parameters means multi-domain adaptation adds almost no run-time cost over a generic tracker.
  • The ablation on NAT2021, moving from 49.11 to 54.58 AUC as CSG, DCA, and TCA are added, shows each component contributes independently to the final result.

Reading between the lines

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

  • The paper does not test whether a DCA trained on three synthetic weather conditions can generalize zero-shot to an unseen weather prompt, but if CSG occupies a continuous semantic weather space, a single token bank might cover new conditions; this is a cheap experiment the authors leave open.
  • The real-world AUC margins (roughly 1.4 to 1.9 points) are much smaller than the synthetic margins (3 to 6 points), and the real evidence is dominated by nighttime data, so the strongest test of the method would be on genuinely collected fog or rain video with no synthetic counterpart.
  • Because the TCA cost normalizes confidence and position differences independently, the same position-sensitive optimal transport loss could be applied to other teacher-student vision tasks with noisy pseudo-labels, such as semi-supervised detection or segmentation, without the weather framing.
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 / 8 minor

Summary. The paper presents UMDATrack, a unified multi-domain adaptive tracker for adverse weather conditions (night, fog, rain). The method has three components: (i) a controllable scenario generator (CSG) that uses Stable Diffusion-Turbo with text prompts to synthesize unlabeled adverse-weather videos from GOT-10k daytime frames; (ii) a domain-customized adapter (DCA) inserted into a frozen ViT backbone, trained in a student-teacher EMA framework; and (iii) a target-aware confidence alignment module (TCA) that adds a position-sensitive optimal transport (PSOT) loss to align localization confidence between source and target domains. The model is trained on four daytime datasets plus three synthetic weather variants, then evaluated on synthetic GOT-10k/DTB70 weather test sets and real-world NAT2021, UAVDark70, and AVisT benchmarks. The paper reports state-of-the-art results on all benchmarks and ablations showing contributions from each module.

Significance. If the results hold, UMDATrack would be a practical and efficient solution for multi-weather tracking: it uses frozen backbones, a lightweight adapter, and only a small amount of synthetic data, and it reports real-time inference (138 FPS). The strengths are the external real-world validation on NAT2021, UAVDark70, and AVisT, the released code, and the careful ablation structure. However, the claimed significance is limited by two concerns: the synthetic test sets are generated by the same pipeline as the training data, making the Table 1 margins in-distribution; and the PSOT loss in Eq. (8) is underspecified and its derivation is relegated to a missing appendix, which is problematic because Table 4 attributes the largest single ablation gain to TCA. The real-world gains over the second-best method are modest (0.94-1.83 AUC points), so the practical significance would be strengthened by a clearer account of statistical significance.

major comments (3)
  1. [Section 3.3, Eq. (8)] The PSOT loss is written as Lp = <µ, dT/||dT||1> + <ν, dS/||dS||1>, but the potentials µ and ν are not defined, the cost matrix C_ij constructed in Eqs. (5)-(7) does not appear in the loss, and the constraint set of the OT dual (e.g., µ_i + ν_j ≤ C_ij) is not stated. The text defers to Appendix B, which is absent from the preprint. Because Table 4 attributes 2.31 AUC and 3.29 precision on NAT2021 to TCA, and this is comparable to the final margin over ARTrackV2 (54.58 vs 53.13), the central empirical claim depends on an unverifiable training objective. Please provide the complete OT formulation and the derivation, or reconcile Eq. (8) with a standard Sinkhorn dual.
  2. [Section 4.2, Table 1] The synthetic test sets (GOT-10k-Foggy/Dark/Rainy, DTB70-Foggy/Dark/Rainy) are produced by the same CSG pipeline (Stable Diffusion-Turbo with text prompts, Section 3.1) that generates the synthetic training videos. The large margins over baselines in Table 1 therefore measure in-distribution performance with respect to the synthetic degradation model, not generalization to genuine fog, night, or rain. This limitation should be stated, and the 'significant margin' claim on synthetic data should be either supported by an independently synthesized test set or removed.
  3. [Section 4.1, Training Details] The sentence 'we only need to train the DCA module for each weather condition' is ambiguous. If separate DCA weights are trained per weather condition, the framework is not a single unified model as claimed in the abstract and introduction; if a single DCA is trained jointly on all conditions, the text should state this and clarify whether the token bank B is shared or domain-specific. This distinction is central to the 'unified multi-domain adaptation' claim and should be resolved in the revised manuscript.
minor comments (8)
  1. [Abstract/Introduction] The word 'unfiied' should be 'unified'.
  2. [Section 4.1, Loss Function] The loss is referred to as 'PDOT loss' here, but it is called 'PSOT loss' earlier; please use one consistent name.
  3. [Section 3.3, Eq. (4)] The notation dS = exp(rS_i,pi) is ambiguous; define dS_i = exp(rS_{i,pi}) for i = 1,...,N, with rS_{i,pi} being the score at the argmax position of the teacher response for sample i.
  4. [Section 3.3] The statement 'we solve the OT problem by a fast Sinkhorn distances algorithm [8]' is inconsistent with writing the loss in dual form; Sinkhorn typically solves the entropic primal. Clarify which optimization problem is actually solved and how the dual potentials are obtained.
  5. [Table 3] 'AQATrack-' should be 'AQATrack' (without the trailing hyphen).
  6. [Tables 1 and 2] 'A VTrack' should be 'AVTrack' (or the intended tracker name) for consistency.
  7. [Section 1] The claim 'first unified multi-domain adaptation tracker in VOT community' is not substantiated by the related-work discussion; either provide a more thorough comparison with existing multi-domain or multi-weather trackers or soften the claim.
  8. [Section 4.1, Training Details] The relationship between 'less than 2% frames in source daytime datasets' (Abstract) and the training sampling ratio 1:1:1:1:4:4:4 is not explained; the oversampling of synthetic data in each batch may offset the small total fraction, so please clarify how the 2% figure is computed and how it relates to the batch composition.

Circularity Check

1 steps flagged · score 3.0 of 10

Synthetic GOT-10k test sets come from the same CSG training pool, making part of the Table 1 'prediction' in-distribution; real-world results remain external anchors.

  1. fitted input called prediction [Section 4.1 (Training Details) + Section 4.2 (Comparisons with SOTA), Table 1]
    "we first synthesize the videos in adverse weather conditions only using GOT-10k dataset, the synthesized datasets includes GOT-10k-Dark, GOT-10k-Foggy and GOT-10k-Rainy. ... we evaluate the tracking performance using the GOT-10k-Foggy and DTB70-Foggy datasets."

    Training and test are named identically: the synthetic GOT-10k videos produced by CSG for training are the same dataset labels later reported as GOT-10k-Foggy/Dark/Rainy in Table 1. The paper specifies no held-out synthesis, no split, and no separate source frames for those test sets. Therefore the Table 1 AUC/precision gains on the GOT-10k synthetic columns measure the tracker's fit to the CSG-generated training distribution, not a prediction on unseen adverse-weather data. This is a fitted-input-called-prediction pattern. The effect does not extend to DTB70 synthetic (different source videos) or to the real-world NAT2021/UAVDark70/AVisT tables, which are external.

full rationale

The real-world claims are self-contained: Table 2 compares on NAT2021, UAVDark70, and AVisT, which were not used to fit the model, so the headline 'SOTA on real-world adverse weather' does not reduce to the method's inputs. The one concrete circularity is limited to the synthetic GOT-10k columns of Table 1. Section 4.1 states that CSG synthesized GOT-10k-Dark/Foggy/Rainy for training, and Section 4.2 evaluates on the same-named GOT-10k-Foggy/Dark/Rainy sets without reporting any train/test split or separate holdout generation. Since the test frames come from the same generator and the same source GOT-10k video pool, those margins are in-distribution by construction rather than a transfer prediction. DTB70 synthetic sets and all real-world sets are independent anchors, so this is a partial, not total, circularity. The PSOT loss in Eq. 8 references an absent Appendix B and omits the cost matrix from the displayed dual objective; that is an omitted derivation/reproducibility defect, not a circular reduction, and I do not add to the score for it. No load-bearing self-citation is present; references [40], [44]-[46] are background/loss citations.

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

The paper is an empirical system contribution, so the ledger records hand-chosen loss weights, hyperparameters tuned on the headline benchmark NAT2021 (Tables 5 and 6), and the three domain-overlap and pseudo-label assumptions that carry synthetic-to-real transfer. No invented entities appear.

free parameters (6)
  • lambda_psot (PSOT loss weight) = 10.0
    Hand-chosen weight for the position-sensitive optimal transport loss in Eq. 9; no sensitivity analysis is reported.
  • beta_l1 (L1 loss weight) = 5.0
    Hand-chosen weight for localization L1 loss in Eq. 10.
  • gamma_giou (GIoU loss weight) = 2.0
    Hand-chosen weight for GIoU loss in Eq. 10.
  • dataset_sampling_ratio = 1:1:1:1:4:4:4 (LaSOT:TrackingNet:COCO:GOT-10k:Dark:Foggy:Rainy)
    Selected in Table 6 as the best of four scanned ratios, evaluated on NAT2021, the same dataset used for the headline real-world result.
  • ema_update_frequency = each epoch
    Selected in Table 5 as the best of four settings (each epoch, every 3, every 5, each batch), evaluated on NAT2021.
  • ema_momentum_alpha = 0.99
    EMA coefficient in Eq. 2; hand-chosen, no sensitivity study.
assumptions (4)
  • standard math The PSOT loss in Eq. 8 is a correct dual formulation of the optimal transport problem with the cost matrix C in Eq. 7.
    Invoked in Section 3.3 with the derivation deferred to appendix B, which is absent from the preprint; Eq. 8 as written does not visibly contain the cost matrix C.
  • domain assumption Stable Diffusion-Turbo translation preserves target identity, spatial structure, and bounding-box validity in synthesized frames.
    CSG (Section 3.1) generates the unlabeled target-domain videos used for adaptation; structure preservation is assumed via Zero-Conv and skip connections, with no identity-preservation metric reported.
  • domain assumption Synthetic weather degradation is a faithful proxy for real night, fog, and rain domain shift in the tracker's feature space.
    The adaptation in Sections 3.2 and 3.3 is trained only on CSG output; the supporting evidence (Table 2 real-world gains) is itself the conclusion being validated.
  • domain assumption Teacher pseudo-labels are sufficiently reliable for student training once the TCA loss is applied.
    Section 3.3 states pseudo-labels may be noisy and proposes TCA to mitigate; the residual noise level is not measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UMDATrack: Unified Multi-Domain Adaptive Tracking Under Adverse Weather Conditions." pith.science (2026). https://pith.science/paper/RLYCUSQZ

@misc{pith2026250700648,
  author       = {Pith},
  title        = {Pith review of: UMDATrack: Unified Multi-Domain Adaptive Tracking Under Adverse Weather Conditions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RLYCUSQZ}},
  note         = {Machine review of arXiv:2507.00648}
}
read the original abstract

Visual object tracking has gained promising progress in past decades. Most of the existing approaches focus on learning target representation in well-conditioned daytime data, while for the unconstrained real-world scenarios with adverse weather conditions, e.g. nighttime or foggy environment, the tremendous domain shift leads to significant performance degradation. In this paper, we propose UMDATrack, which is capable of maintaining high-quality target state prediction under various adverse weather conditions within a unified domain adaptation framework. Specifically, we first use a controllable scenario generator to synthesize a small amount of unlabeled videos (less than 2% frames in source daytime datasets) in multiple weather conditions under the guidance of different text prompts. Afterwards, we design a simple yet effective domain-customized adapter (DCA), allowing the target objects' representation to rapidly adapt to various weather conditions without redundant model updating. Furthermore, to enhance the localization consistency between source and target domains, we propose a target-aware confidence alignment module (TCA) following optimal transport theorem. Extensive experiments demonstrate that UMDATrack can surpass existing advanced visual trackers and lead new state-of-the-art performance by a significant margin. Our code is available at https://github.com/Z-Z188/UMDATrack.

Figures

Figures reproduced from arXiv: 2507.00648 by the authors.

Figure 1
Figure 1. Three representative tracking pipelines under adverse [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed UMDATrack. It first utilizes a controllable scenarios generator (CSG) to synthesize the video frames in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Details of the Controllable Scenario Generation (CSG) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 7
Figure 7. Figure 7: Visualization comparison of our approach and other ex [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 6
Figure 6. Figure 6: Feature visualization by t-SNE of dark, foggy, and rainy [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 50 canonical work pages

  1. [1]

    Ar- trackv2: Prompting autoregressive tracker where to look and how to describe

    Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Ar- trackv2: Prompting autoregressive tracker where to look and how to describe. In IEEE Conference on Computer Vision and Pattern Recognition, pages 19048–19057, 2024. 6, 7

  2. [2]

    Learning discriminative model prediction for track- ing

    Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for track- ing. In IEEE International Conference on Computer Vision, pages 6182–6191, 2019. 6

  3. [3]

    Hiptrack: Vi- sual tracking with historical prompts

    Wenrui Cai, Qingjie Liu, and Yunhong Wang. Hiptrack: Vi- sual tracking with historical prompts. In IEEE Conference on Computer Vision and Pattern Recognition, pages 19258– 19267, 2024. 6

  4. [4]

    Robust object modeling for visual tracking

    Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In IEEE International Conference on Computer Vision, pages 9589–9600, 2023. 6, 7

  5. [5]

    Seqtrack: Sequence to sequence learning for visual object tracking

    Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14572–14581, 2023. 6, 7

  6. [6]

    Siamese box adaptive network for visual tracking

    Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, and Rongrong Ji. Siamese box adaptive network for visual tracking. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6667–6676, 2020. 2

  7. [7]

    Optimal transport for domain adaptation

    Nicolas Courty, R´emi Flamary, Devis Tuia, and Alain Rako- tomamonjy. Optimal transport for domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1853–1865, 2017. 3

  8. [8]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Annual Conference on Neural Infor- mation Processing Systems, pages 2292–2300, 2013. 5

Show all 55 references
  1. [9]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

  2. [10]

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

    Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5374–5383, 2019. 1, 5

  3. [11]

    Highlightnet: Highlighting low-light potential features for real-time UA V tracking

    Changhong Fu, Haolin Dong, Junjie Ye, Guangze Zheng, Si- hang Li, and Jilin Zhao. Highlightnet: Highlighting low-light potential features for real-time UA V tracking. InIEEE/RSJ International Conference on Intelligent Robots and Systems, pages 12146–12153, 2022. 2

  4. [12]

    SAM-DA: UA V Tracks Anything at Night with SAM-Powered Domain Adaptation

    Changhong Fu, Liangliang Yao, Haobo Zuo, Guangze Zheng, and Jia Pan. SAM-DA: UA V Tracks Anything at Night with SAM-Powered Domain Adaptation. In Proceedings of the IEEE International Conference on Advanced Robotics and Mechatronics, pages 1–8, 2024. 6, 7

  5. [13]

    Expressive text-to-image generation with rich text

    Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In IEEE International Conference on Computer Vision, pages 7511– 7522, 2023. 2

  6. [14]

    Separable self and mixed attention transformers for efficient object tracking

    Goutam Yelluru Gopal and Maria A Amer. Separable self and mixed attention transformers for efficient object tracking. In IEEE Winter Conference on Applications of Computer Vision, pages 6708–6717, 2024. 7

  7. [15]

    Visualizing data using t-sne journal of machine learning research

    G Hinton and L Van Der Maaten. Visualizing data using t-sne journal of machine learning research. Journal of Machine Learning Research, 9:2579–2605, 2008. 8

  8. [16]

    SOOD: towards semi- supervised oriented object detection

    Wei Hua, Dingkang Liang, Jingyu Li, Xiaolong Liu, Zhikang Zou, Xiaoqing Ye, and Xiang Bai. SOOD: towards semi- supervised oriented object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 15558– 15567, 2023. 3

  9. [17]

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

    Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1562–1577, 2019. 5

  10. [18]

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial net- works. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5967–5976, 2017. 2

  11. [19]

    Adtrack: Target-aware dual filter learning for real-time anti-dark UA V tracking

    Bowen Li, Changhong Fu, Fangqiang Ding, Junjie Ye, and Fuling Lin. Adtrack: Target-aware dual filter learning for real-time anti-dark UA V tracking. In IEEE International Conference on Robotics and Automation , pages 496–502,

  12. [20]

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

    Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4282–4291,

  13. [21]

    High performance visual tracking with siamese region proposal network

    Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8971–8980, 2018. 6

  14. [22]

    Visual object tracking for un- manned aerial vehicles: A benchmark and new motion models

    Siyi Li and Dit-Yan Yeung. Visual object tracking for un- manned aerial vehicles: A benchmark and new motion models. In AAAI Conference on Artificial Intelligence, pages 4140– 4146, 2017. 6

  15. [23]

    Training-free model merging for multi- target domain adaptation

    Wenyi Li, Huan-ang Gao, Mingju Gao, Beiwen Tian, Rong Zhi, and Hao Zhao. Training-free model merging for multi- target domain adaptation. arXiv preprint arXiv:2407.13771,

  16. [24]

    Sigma: Semantic- complete graph matching for domain adaptive object detec- tion

    Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic- complete graph matching for domain adaptive object detec- tion. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5291–5300, 2022. 3

  17. [25]

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

    Yongxin Li, Mengyuan Liu, You Wu, Xucheng Wang, Xi- angyang Yang, and Shuiwang Li. Learning adaptive and view-invariant vision transformer for real-time uav tracking. In International Conference on Machine Learning. 6, 7 9

  18. [26]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740–755,

  19. [27]

    Breuel, and Jan Kautz

    Ming-Yu Liu, Thomas M. Breuel, and Jan Kautz. Unsu- pervised image-to-image translation networks. In Annual Conference on Neural Information Processing Systems, pages 700–708, 2017. 8

  20. [28]

    Mutual-learning knowledge distillation for night- time UA V tracking.arXiv preprint arXiv:2312.07884, 2023

    Yufeng Liu. Mutual-learning knowledge distillation for night- time UA V tracking.arXiv preprint arXiv:2312.07884, 2023. 6, 7

  21. [29]

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

    Matthias M ¨uller, Adel Bibi, Silvio Giancola, Salman Al- Subaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In European Conference on Computer Vision, pages 310–327,

  22. [30]

    GLIDE: towards photorealistic image gen- eration and editing with text-guided diffusion models

    Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image gen- eration and editing with text-guided diffusion models. In International Conference on Machine Learning, p...

  23. [31]

    Avist: A benchmark for visual object tracking in adverse visibility

    Mubashir Noman, Wafa Al Ghallabi, Daniya Kareem, Christoph Mayer, Akshay Dudhane, Martin Danelljan, Hisham Cholakkal, Salman Khan, Luc Van Gool, and Fa- had Shahbaz Khan. Avist: A benchmark for visual object tracking in adverse visibility. In British Machine Vision Con- ferenc...

  24. [32]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Confer- ence on Machine Learning, pages 8821–8831, 2021. 2

  25. [33]

    Generalized in- tersection over union: A metric and a loss for bounding box regression

    Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized in- tersection over union: A metric and a loss for bounding box regression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 658–666, 2019. 6

  26. [34]

    Focal loss for dense ob- ject detection

    T-YLPG Ross and GKHP Doll ´ar. Focal loss for dense ob- ject detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2980–2988, 2017. 5

  27. [35]

    Curriculum graph co-teaching for multi- target domain adaptation

    Subhankar Roy, Evgeny Krivosheev, Zhun Zhong, Nicu Sebe, and Elisa Ricci. Curriculum graph co-teaching for multi- target domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5351–5360, 2021. 3

  28. [36]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87–103, 2024. 3

  29. [37]

    Explicit visual prompts for visual object tracking

    Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Sheng- ping Zhang, and Xianxian Li. Explicit visual prompts for visual object tracking. In AAAI Conference on Artificial Intel- ligence, pages 4838–4846, 2024. 6, 7

  30. [38]

    Cross-modal pattern- propagation for RGB-T tracking

    Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal pattern- propagation for RGB-T tracking. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7062–7071,

  31. [39]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, 2004. 8

  32. [40]

    Lvptrack: High per- formance domain adaptive UA V tracking with label aligned visual prompt tuning

    Hongjing Wu, Siyuan Yao, Feng Huang, Shu Wang, Linchao Zhang, Zhuoran Zheng, and Wenqi Ren. Lvptrack: High per- formance domain adaptive UA V tracking with label aligned visual prompt tuning. In AAAI Conference on Artificial Intel- ligence, pages 8395–8403, 2025. 1

  33. [41]

    Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In IEEE Conference on Computer Vision and Pattern Recognition , pages 14561–14571, 2023. 6, 7

  34. [42]

    Autore- gressive queries for adaptive tracking with spatio-temporal transformers

    Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autore- gressive queries for adaptive tracking with spatio-temporal transformers. In IEEE Conference on Computer Vision and Pattern Recognition, pages 19300–19309, 2024. 6, 7

  35. [43]

    Depthtrack: Un- veiling the power of RGBD tracking

    Song Yan, Jinyu Yang, Jani K ¨apyl¨a, Feng Zheng, Ales Leonardis, and Joni-Kristian K¨am¨ar¨ainen. Depthtrack: Un- veiling the power of RGBD tracking. In IEEE International Conference on Computer Vision, pages 10705–10713, 2021. 2

  36. [44]

    Unctrack: Reliable visual object tracking with uncertainty-aware prototype memory network

    Siyuan Yao, Yang Guo, Yanyang Yan, Wenqi Ren, and Xi- aochun Cao. Unctrack: Reliable visual object tracking with uncertainty-aware prototype memory network. IEEE Trans- actions on Image Processing, 34:3533–3546, 2025. 6

  37. [45]

    Learning deep lucas-kanade siamese network for visual tracking

    Siyuan Yao, Xiaoguang Han, Hua Zhang, Xiao Wang, and Xiaochun Cao. Learning deep lucas-kanade siamese network for visual tracking. IEEE Transactions on Image Processing, 30:4814–4827, 2021. 2

  38. [46]

    Robust online tracking via con- trastive spatio-temporal aware network

    Siyuan Yao, Hua Zhang, Wenqi Ren, Chao Ma, Xiaoguang Han, and Xiaochun Cao. Robust online tracking via con- trastive spatio-temporal aware network. IEEE Transactions on Image Processing, 30:1989–2002, 2021. 2

  39. [47]

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

    Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European Conference on Computer Vision, pages 341–357, 2022. 5, 6

  40. [48]

    Darklighter: Light up the darkness for UA V track- ing

    Junjie Ye, Changhong Fu, Guangze Zheng, Ziang Cao, and Bowen Li. Darklighter: Light up the darkness for UA V track- ing. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3079–3085, 2021. 1, 2

  41. [49]

    Unsupervised domain adaptation for nighttime aerial tracking

    Junjie Ye, Changhong Fu, Guangze Zheng, Danda Pani Paudel, and Guang Chen. Unsupervised domain adaptation for nighttime aerial tracking. In IEEE Conference on Com- puter Vision and Pattern Recognition, pages 8886–8895, 2022. 1, 2, 6, 7

  42. [50]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE International Conference on Computer Vision, pages 3813–3824, 2023. 2

  43. [51]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Com- puter Vision and Pattern Recognition, pages 586–595, 2018. 8 10

  44. [52]

    Domain adaptive siamrpn++ for object tracking in the wild

    Zhongzhou Zhang and Lei Zhang. Domain adaptive siamrpn++ for object tracking in the wild. arXiv preprint arXiv:2106.07862, 2021. 1, 2

  45. [53]

    Odtrack: Online dense temporal token learning for visual tracking

    Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. In AAAI Confer- ence on Artificial Intelligence, 2024. 6, 7

  46. [54]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In IEEE International Conference on Computer Vision, pages 2242–2251, 2017. 8

  47. [55]

    DCPT: darkness clue-prompted tracking in nighttime uavs

    Jiawen Zhu, Huayi Tang, Zhi-Qi Cheng, Jun-Yan He, Bin Luo, Shihao Qiu, Shengming Li, and Huchuan Lu. DCPT: darkness clue-prompted tracking in nighttime uavs. In IEEE International Conference on Robotics and Automation, pages 7381–7388, 2024. 1, 6, 7 11

Pith tools

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