Pith. sign in

REVIEW 3 major objections 5 minor 81 references

Decoupled Spatio-Temporal Consistency Learning for Self-Supervised Tracking

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

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

Pith's one-line read SSTrack is a self-supervised visual tracker that trains on unlabeled video using only the first-frame bounding box; it claims to surpass previous self-supervised trackers by wide margins and to come within 7.3 percent AUC of the best…

desk verdict SSTrack reports big self-supervised tracking gains, but the training objective is underspecified and the annotation claims are overstated; worth a careful round of review. read the letter →

arxiv 2507.21606 v1 pith:5UZWKUUA submitted 2025-07-29 cs.CV

Yaozong Zheng , Bineng Zhong , Qihua Liang , Ning Li , Shuxiang Song This is my paper
classification cs.CV
keywords self-supervisedtrackingvisualobjectspatio-temporalconsistencyglobalforwardlocalbackwardinstancecontrastivelearningViTtrackerunlabeledvideopretraining
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 sets out to show that visual object tracking does not need frame-by-frame bounding-box annotations. It introduces SSTrack, a tracker trained from unlabeled videos with only the initial box, using a decoupled spatio-temporal consistency framework that first searches globally for the target and then locally re-associates it across time. An instance-contrastive loss adds label-free instance discrimination. On nine benchmarks the method reports state-of-the-art self-supervised results, including large gains over the prior TADS baseline and a single-digit gap to fully supervised trackers on LaSOT. If this holds, the main bottleneck on tracking data scale, manual box annotation, is substantially relaxed.

What carries the argument

The decoupled spatio-temporal consistency training framework is the central mechanism. During training the model first runs global forward tracking on uncropped search frames to locate the target, then crops those frames around the predicted boxes and runs backward tracking using augmented views of the initial frame as search inputs, closing a spatio-temporal cycle whose classification and regression losses supervise the network without frame-wise annotations. The second component is an instance-contrastive loss that masks the predicted target region, pools its representation, and pulls different views of the same instance together while pushing different instances apart in feature space.

What would settle it

Run the SSTrack training pipeline with the forward tracking outputs replaced by ground-truth boxes, and separately with forward predictions corrupted or fixed to wrong positions; if the cycle loss still decreases while benchmark accuracy collapses, the spatio-temporal cycle is not learning true target identity.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a tracking network can be trained self-supervised by decoupling its cycle into a global forward stage and a local backward stage: first locate the target anywhere in the full frame, then use the forward prediction as the reference to re-associate the same instance in augmented views of the initial frame. With an instance-contrastive loss added to this cycle, SSTrack reports state-of-the-art self-supervised results on nine benchmarks: AO 72.4 on GOT10K (25.7 points above TADS), AUC 65.9 on LaSOT (20.4 above TADS and 7.3 below the supervised ODTrack), and AUC 80.4 on TrackingNet (14.8 above TADS). The paper further claims that at inference only the local tracker is kept, yielding 59 fps on an A100 GPU.

Load-bearing premise

The whole self-supervised signal inherits the forward tracker's localization errors, because backward tracking uses the forward predictions as its pseudo ground truth; if forward localization is wrong, the cycle loss trains on its own mistakes.

Editorial extensions

If this is right

  • Training trackers no longer requires per-frame box annotations; only the initial box plus unlabeled video is enough to reach competitive accuracy.
  • The gap between self-supervised and fully supervised tracking can be reduced to single-digit AUC points, so annotation cost no longer strictly gates tracking performance.
  • The decoupled cycle is a training-time device: at inference only the local tracker is kept, giving 59 fps at 384x384 resolution with a ViT-B backbone.
  • Instance-level contrastive learning with predicted masks adds a consistent improvement over the cycle alone, about 1.6 AO on GOT10K.
  • Temporal-correspondence-aware pretraining is worth 5.9 AO over a plain masked-autoencoder pretraining, so the choice of video pretraining interacts with the self-supervised objective.

Reading between the lines

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

  • The forward-as-teacher design implies the upper bound of the whole pipeline is set by the forward tracker's own accuracy; a stronger forward stage or an uncertainty weighting would likely lift the same architecture further, which the paper does not test.
  • The same decoupled global-search/local-association cycle could transfer to other video correspondence tasks, such as video object segmentation, point tracking, or re-identification, where the supervision bottleneck is also per-frame labels.
  • Scaling the number of views per instance or the batch size should strengthen the instance-contrastive term without any new labels; the paper's fixed batch size and view counts leave that axis unexplored.
  • A direct diagnostic would be to train the identical pipeline with forward boxes replaced by oracle boxes; the performance delta would quantify how much self-generated pseudo-label error currently costs.
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 SSTrack, a self-supervised visual tracker that combines a forward global-search tracking stage, a backward local-association stage, and an instance contrastive loss. Training is claimed to require only the first-frame bounding box, not frame-wise boxes. The method is evaluated on nine benchmarks, where it reports large gains over the prior self-supervised tracker TADS (e.g., GOT-10K AO 72.4 vs 46.7; LaSOT AUC 65.9 vs 45.5; TrackingNet AUC 80.4 vs 65.6) and narrows the gap to fully supervised trackers. The core training objective uses a tracking loss Ltrack and a contrastive loss Lcont, with ablations for each component.

Significance. If the central claim holds, this would be a meaningful step toward reducing annotation cost in visual tracking: it would show that a contrastive/cycle-consistency scheme with one initial box can produce representations close to supervised trackers. The paper's strengths include evaluation on nine external benchmarks, an ablation study isolating the proposed components, public code, and a direct comparison with prior self-supervised methods, which provide independent evidence that the reported behavior is not confined to a single test set. However, the significance is conditional on a precise definition of the self-supervised training target; see major comments.

major comments (3)
  1. [Methodology, Eq. (4) and Algorithm 1] The central training objective is underspecified: Eq. (4) defines Ltrack = Lcls(Bs, B_gt^s) + Lreg(Bs, B_gt^s), but B_gt^s is never defined in the self-supervised setting. Algorithm 1 describes only forward tracking that produces B_t^s, cropping from the search frame based on B_t^s, backward tracking that produces B_t^r, and a final loss 'using Eq.6'; no independent ground-truth box is introduced anywhere in the unlabeled video. If B_gt^s is intended to be the forward tracker's own prediction, the training loss is a pseudo-label bootstrap that reuses the tracker's output as its own target, and the backward stage simply propagates forward localization errors, as the Limitation section concedes. The authors must specify the exact target boxes for both forward and backward branches (for example, transformed versions of the initial box under the data augmentations), and describe how this target avoids collapse or error accumulation. Without this, the reported gains cannot be attributed to a novel self-supervised consistency signal.
  2. [Abstract and Section 'Self-Supervised Pipeline'] The claim that the method 'eliminate[s] the need of box annotations' is contradicted by the paper's own setup. The task definition states that the tracker is given the initial bounding box annotation of the instance, and the Implementation Details describe training on LaSOT, GOT-10K, TrackingNet, and COCO, all of which contain manual bounding-box labels. What is actually proposed is a reduction of per-frame annotation to a single first-frame box. The abstract, introduction, and conclusion should be revised to state 'one-frame annotation' or 'reduced reliance on box annotations' rather than 'eliminate the need of box annotations.'
  3. [Algorithm 1 and Eq. (6)] Line 12 of Algorithm 1 refers to 'Calculate loss using Eq.6', but Eq. (6) is the total objective Lall = Ltrack + Lcont, and the algorithm does not show how Ltrack is evaluated: it does not state which predictions Bs are passed to Eq. (4) and what targets B_gt^s are used for the forward and backward branches. The notation B_t^r is also used in line 9 without being defined as an output of the backward stage. Please rewrite Algorithm 1 to itemize the predictions and the exact loss terms, and align it with the equations in Section 3.
minor comments (5)
  1. [Abstract and Table 2] Abstract reports 'more than 25.3%' AO improvement on GOT-10K, while Table 2 reports 25.7; please align these numbers.
  2. [Equation (5)] Equation (5) writes the summation as a symbol that appears as 'P' rather than a proper sum; the notation should be corrected, and the sets Q, q^+, and q^- should be defined more precisely.
  3. [Table 3 caption] Table 3 caption contains a typo ('Datesets') and the dataset name 'UA V123' has inconsistent spacing.
  4. [Table 2] The footnote '∗ denotes for trackers only trained on GOT10K' is ambiguous for the SSTrack rows; please state which training split is used for each reported SSTrack result on GOT-10K.
  5. [Comparison text, LaSOText, TNL2K, VOT2020 paragraph] The text says CycleSiam is surpassed by 6.8% in VOT2018 accuracy, but the table values (0.562 vs 0.630) correspond to 6.8 percentage points, not 6.8%; please correct the phrasing.

Circularity Check

2 steps flagged · score 4.0 of 10

Benchmark gains are externally measured, but the self-supervised training signal is partially self-referential: the contrastive mask is built from the model's own predictions and the tracking-loss target Bgt_s is never defined, so as written the objective can bootstrap from forward-tracker outputs.

  1. self definitional [Instance Contrastive Learning (Eq. 5) and Algorithm 1 (lines 3-12)]
    "we design an additional mask matrix M for each view based on the prediction results to extract the target instance from the background, where 1 represents the target region and 0 represents the background region. ... Calculate loss using Eq.6 and update parameters."

    The instance-contrastive loss is presented as label-free supervision, but the mask that defines the target instance is explicitly computed from the model's own prediction results. Positive views of the same instance are therefore selected by the tracker itself, so the loss trains the representation to agree with the current predictor rather than with any external correspondence. The same forward predictions are used to crop the backward-tracking reference frames, and the Limitation section concedes that the performance of backward tracking somewhat depends on the localization accuracy of forward tracking. The self-supervised loop is thus partially a self-bootstrap.

  2. other [Methodology, Eq. (4) and Algorithm 1]
    "Ltrack = Lcls(Bs, Bgt s ) +Lreg(Bs, Bgt s ); ... Crop I t s based on Bt s yields a new reference frame I t sr"

    Eq. (4) defines the tracking loss against a ground-truth box Bgt_s, but no self-supervised definition of Bgt_s is provided anywhere in the paper. In Algorithm 1, the only boxes produced in the pipeline are the forward tracker's Bt_s, and those same boxes are used to create the reference frames for backward tracking. As written, the objective can therefore reduce to regressing the tracker to its own forward predictions; the missing definition leaves no independently supervised target in the tracking loss.

full rationale

The headline claims of SOTA performance on GOT10K, LaSOT, TrackingNet, and the other benchmarks are not themselves circular: the numbers are measured against held-out manual annotations and against published TADS and supervised baselines, which are external to this paper's fitted values. However, the paper's self-supervised training signal contains a genuine self-referential component. The contrastive mask is generated based on the prediction results, and Eq. (4)'s target Bgt_s is never defined; in Algorithm 1 the forward tracker's output boxes are the only boxes available and are reused to crop backward-tracking references. The Limitation section explicitly concedes that backward tracking depends on forward localization accuracy. This makes the learning loop partially a pseudo-label bootstrap, so the internal claim of supervision without labels is weaker than presented. Because the benchmark evaluation remains external and the method is not reduced to a self-citation or a fitted-parameter trick, the overall circularity score is 4 rather than higher.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The method's central claim rests on a cycle whose pseudo-labels are generated by the model itself, plus a hand-chosen set of hyperparameters (temperature, frame counts, augmentations). No new entities are introduced. The external benchmarks provide the only independent check.

free parameters (3)
  • temperature tau in instance contrastive loss (Eq. 5) = not reported
    Controls the sharpness of the contrastive distribution; no value or sensitivity study is given, so the reported gains from L_cont cannot be separated from this hyperparameter.
  • number of forward and backward tracking steps = forward: 1 reference + 3 global search frames; backward: 3 reference + 2 cropped search frames
    Manual design choice in Algorithm 1 and Implementation Details; no ablation tests other numbers of frames.
  • data augmentation set (shear, blur, LSJ) = shear, blur, LSJ
    Ablated in Table 8 and combined by performance; the chosen combination is a hand-selected configuration.
assumptions (6)
  • domain assumption Unlabeled videos contain a persistent target instance whose identity can be recovered by cycle consistency from a single initial bounding box.
    The entire forward-backward training loop in Algorithm 1 presupposes that the initial box identifies a trackable instance across later frames.
  • ad hoc to paper The forward tracker's predictions are accurate enough to serve as pseudo references for the backward stage.
    The loss in Eq. (4) needs targets for B_s; in the self-supervised setting these can only come from the model's own forward outputs, an assumption the Limitation section concedes.
  • ad hoc to paper Ground-truth boxes B_gt^s exist for the self-supervised loss in Eq. (4).
    The paper never defines where the pseudo ground truth comes from; it uses the symbol without construction in the unlabeled setting.
  • domain assumption DropMAE pretrained weights from external video data are a valid initialization.
    Implementation Details says the ViT-B encoder starts from DropMAE; the method is not trained from scratch and inherits that pretraining's biases.
  • domain assumption Evaluation protocols of the nine benchmarks are satisfied.
    SSTrack is trained on LaSOT, GOT-10k, TrackingNet, and COCO jointly, while the GOT-10k one-shot protocol typically restricts training to the GOT-10k train split; the paper does not reconcile this.
  • standard math Standard loss functions (focal, GIoU, L1, InfoNCE-style contrastive) are appropriate and implemented correctly.
    The paper relies on standard deep learning objectives without derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupled Spatio-Temporal Consistency Learning for Self-Supervised Tracking." pith.science (2026). https://pith.science/paper/5UZWKUUA

@misc{pith2026250721606,
  author       = {Pith},
  title        = {Pith review of: Decoupled Spatio-Temporal Consistency Learning for Self-Supervised Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5UZWKUUA}},
  note         = {Machine review of arXiv:2507.21606}
}
read the original abstract

The success of visual tracking has been largely driven by datasets with manual box annotations. However, these box annotations require tremendous human effort, limiting the scale and diversity of existing tracking datasets. In this work, we present a novel Self-Supervised Tracking framework named \textbf{{\tracker}}, designed to eliminate the need of box annotations. Specifically, a decoupled spatio-temporal consistency training framework is proposed to learn rich target information across timestamps through global spatial localization and local temporal association. This allows for the simulation of appearance and motion variations of instances in real-world scenarios. Furthermore, an instance contrastive loss is designed to learn instance-level correspondences from a multi-view perspective, offering robust instance supervision without additional labels. This new design paradigm enables {\tracker} to effectively learn generic tracking representations in a self-supervised manner, while reducing reliance on extensive box annotations. Extensive experiments on nine benchmark datasets demonstrate that {\tracker} surpasses \textit{SOTA} self-supervised tracking methods, achieving an improvement of more than 25.3\%, 20.4\%, and 14.8\% in AUC (AO) score on the GOT10K, LaSOT, TrackingNet datasets, respectively. Code: https://github.com/GXNU-ZhongLab/SSTrack.

Figures

Figures reproduced from arXiv: 2507.21606 by the authors.

Figure 1
Figure 1. The annotation requirements of different tracking [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. SSTrack training and inference pipeline. 1) Forward Tracking: Given an initial frame and a global search frame, our [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of our tracker with self [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 68 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bai, Y.; Zhao, Z.; Gong, Y.; and Wei, X. 2024. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19048--19057

    work page 2024

  4. [4]

    F.; Vedaldi, A.; and Torr, P

    Bertinetto, L.; Valmadre, J.; Henriques, J. F.; Vedaldi, A.; and Torr, P. H. S. 2016. Fully-Convolutional Siamese Networks for Object Tracking. In ECCV Workshops , 850--865

    work page 2016

  5. [5]

    V.; and Timofte, R

    Bhat, G.; Danelljan, M.; Gool, L. V.; and Timofte, R. 2019. Learning Discriminative Model Prediction for Tracking. In ICCV , 6181--6190

    work page 2019

  6. [6]

    Cai, W.; Liu, Q.; and Wang, Y. 2024. HIPTrack: Visual Tracking with Historical Prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19258--19267

    work page 2024

  7. [7]

    Chen, B.; Li, P.; Bai, L.; Qiao, L.; Shen, Q.; Li, B.; Gan, W.; Wu, W.; and Ouyang, W. 2022. Backbone is All Your Need: A Simplified Architecture for Visual Object Tracking. In ECCV (22) , 375--392

    work page 2022

  8. [8]

    Chen, X.; and He, K. 2021. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 15750--15758

    2021

Show all 81 references
  1. [9]

    Chen, X.; Peng, H.; Wang, D.; Lu, H.; and Hu, H. 2023. SeqTrack: Sequence to Sequence Learning for Visual Object Tracking. CVPR, abs/2304.14394

    2023 arXiv

  2. [10]

    Chen, X.; Yan, B.; Zhu, J.; Wang, D.; Yang, X.; and Lu, H. 2021. Transformer Tracking. In CVPR , 8126--8135

    2021

  3. [11]

    Chen, Z.; Zhong, B.; Li, G.; Zhang, S.; and Ji, R. 2020. Siamese Box Adaptive Network for Visual Tracking. In CVPR , 6667--6676

    2020

  4. [12]

    Cheng, S.; Zhong, B.; Li, G.; Liu, X.; Tang, Z.; Li, X.; and Wang, J. 2021. Learning To Filter: Siamese Relation Network for Robust Tracking. In CVPR , 4421--4431

    2021

  5. [13]

    Cui, Y.; Jiang, C.; Wang, L.; and Wu, G. 2022. MixFormer: End-to-End Tracking with Iterative Mixed Attention. In CVPR , 13598--13608

    2022

  6. [14]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; Uszkoreit, J.; and Houlsby, N. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR

    2021

  7. [15]

    Fan, H.; Bai, H.; Lin, L.; Yang, F.; Chu, P.; Deng, G.; Yu, S.; Harshit; Huang, M.; Liu, J.; Xu, Y.; Liao, C.; Yuan, L.; and Ling, H. 2021. LaSOT: A High-quality Large-scale Single Object Tracking Benchmark. Int. J. Comput. Vis., 439--461

    2021

  8. [16]

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

    2019

  9. [17]

    Fu, Z.; Liu, Q.; Fu, Z.; and Wang, Y. 2021. STMTrack: Template-Free Visual Tracking With Space-Time Memory Networks. In CVPR , 13774--13783

    2021

  10. [18]

    Gao, D.; Lu, S.; Walters, S.; Zhou, W.; Chu, J.; Zhang, J.; Zhang, B.; Jia, M.; Zhao, J.; Fan, Z.; et al. 2024. EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers. arXiv preprint arXiv:2412.20413

    2024 arXiv

  11. [19]

    Gao, S.; Zhou, C.; Ma, C.; Wang, X.; and Yuan, J. 2022. AiATrack: Attention in Attention for Transformer Visual Tracking. In ECCV (22) , 146--164

    2022

  12. [20]

    Gao, S.; Zhou, C.; and Zhang, J. 2023. Generalized Relation Modeling for Transformer Tracking. CVPR, abs/2303.16580

    2023 arXiv

  13. [21]

    D.; Le, Q

    Ghiasi, G.; Cui, Y.; Srinivas, A.; Qian, R.; Lin, T.-Y.; Cubuk, E. D.; Le, Q. V.; and Zoph, B. 2021. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2918--2928

    2021

  14. [22]

    Gong, Y.; Huang, L.; and Chen, L. 2021. Eliminate deviation with deviation for data augmentation and a general multi-modal data learning method. arXiv preprint arXiv:2101.08533

    2021 arXiv

  15. [23]

    Gong, Y.; Huang, L.; and Chen, L. 2022. Person re-identification method based on color attack and joint defence. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4313--4322

    2022

  16. [24]

    Gong, Y.; Zhong, Z.; Qu, Y.; Luo, Z.; Ji, R.; and Jiang, M. 2024. Cross-Modality Perturbation Synergy Attack for Person Re-identification. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  17. [25]

    Guo, M.; Zhang, Z.; Fan, H.; Jing, L.; Lyu, Y.; Li, B.; and Hu, W. 2022. Learning Target-aware Representation for Visual Tracking via Informative Interactions. In IJCAI , 927--934

    2022

  18. [26]

    He, C.; Fang, C.; Zhang, Y.; Li, K.; Tang, L.; You, C.; Xiao, F.; Guo, Z.; and Li, X. 2025 a . Reti-diff: Illumination degradation image restoration with retinex-based latent diffusion model. ICLR

    2025

  19. [27]

    He, C.; Li, K.; Xu, G.; Yan, J.; Tang, L.; Zhang, Y.; Wang, Y.; and Li, X. 2023 a . Hqg-net: Unpaired medical image enhancement with high-quality guidance. TNNLS

    2023

  20. [28]

    He, C.; Li, K.; Xu, G.; Zhang, Y.; Hu, R.; Guo, Z.; and Li, X. 2023 b . Degradation-Resistant Unfolding Network for Heterogeneous Image Fusion. In ICCV, 12611--12621

    2023

  21. [29]

    He, C.; Li, K.; Zhang, Y.; Tang, L.; Zhang, Y.; Guo, Z.; and Li, X. 2023 c . Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, 22046--22055

    2023

  22. [30]

    He, C.; Li, K.; Zhang, Y.; Xu, G.; Tang, L.; Zhang, Y.; Guo, Z.; and Li, X. 2024 a . Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. NeurIPS, 36

    2024

  23. [31]

    He, C.; Li, K.; Zhang, Y.; Yang, Z.; Tang, L.; Zhang, Y.; Kong, L.; and Farsiu, S. 2025 b . Segment concealed object with incomplete supervision. TPAMI

    2025

  24. [32]

    He, C.; Li, K.; Zhang, Y.; Zhang, Y.; Guo, Z.; Li, X.; Danelljan, M.; and Yu, F. 2024 b . Strategic Preys Make Acute Predators: Enhancing Camouflaged Object Detectors by Generating Camouflaged Objects. ICLR

    2024

  25. [33]

    He, C.; Shen, Y.; Fang, C.; Xiao, F.; Tang, L.; Zhang, Y.; Zuo, W.; Guo, Z.; and Li, X. 2025 c . Diffusion Models in Low-Level Vision: A Survey. TPAMI

    2025

  26. [34]

    He, C.; Zhang, R.; Xiao, F.; Fang, C.; Tang, L.; Zhang, Y.; and Farsiu, S. 2025 d . UnfoldIR: Rethinking Deep Unfolding Network in Illumination Degradation Image Restoration. arXiv preprint arXiv:2505.06683

    2025 arXiv

  27. [35]

    He, C.; Zhang, R.; Xiao, F.; Fang, C.; Tang, L.; Zhang, Y.; Kong, L.; Fan, D.-P.; Li, K.; and Farsiu, S. 2025 e . RUN: Reversible Unfolding Network for Concealed Object Segmentation. ICML

    2025

  28. [36]

    He, K.; Fan, H.; Wu, Y.; Xie, S.; and Girshick, R. 2020. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9729--9738

    2020

  29. [37]

    Huang, L.; Zhao, X.; and Huang, K. 2021. GOT-10k: A Large High-Diversity Benchmark for Generic Object Tracking in the Wild. IEEE Trans. Pattern Anal. Mach. Intell. , 43(5): 1562--1577

    2021

  30. [38]

    Kristan, M.; Leonardis, A.; and et.al. 2020. The Eighth Visual Object Tracking VOT2020 Challenge Results. In ECCV Workshops (5) , volume 12539 of Lecture Notes in Computer Science, 547--601. Springer

    2020

  31. [39]

    Kristan, M.; Leonardis, A.; Matas, J.; Felsberg, M.; Pflugfelder, R.; ˇCehovin Zajc, L.; Vojir, T.; Bhat, G.; Lukezic, A.; Eldesokey, A.; et al. 2018. The sixth visual object tracking vot2018 challenge results. In Proceedings of the European conference on computer vision (ECCV...

    2018

  32. [40]

    Li, B.; Wu, W.; Wang, Q.; Zhang, F.; Xing, J.; and Yan, J. 2019. SiamRPN++: Evolution of Siamese Visual Tracking With Very Deep Networks. In CVPR , 4282--4291

    2019

  33. [41]

    Li, B.; Yan, J.; Wu, W.; Zhu, Z.; and Hu, X. 2018. High Performance Visual Tracking With Siamese Region Proposal Network. In CVPR , 8971--8980

    2018

  34. [42]

    Li, L.; Lu, S.; Ren, Y.; and Kong, A. W.-K. 2025. Set you straight: Auto-steering denoising trajectories to sidestep unwanted concepts. arXiv preprint arXiv:2504.12782

    2025 arXiv

  35. [43]

    Li, X.; Pei, W.; Wang, Y.; He, Z.; Lu, H.; and Yang, M.-H. 2023. Self-supervised tracking via target-aware data synthesis. IEEE Transactions on Neural Networks and Learning Systems

    2023

  36. [44]

    B.; He, K.; and Doll \' a r, P

    Lin, T.; Goyal, P.; Girshick, R. B.; He, K.; and Doll \' a r, P. 2017. Focal Loss for Dense Object Detection. In ICCV , 2999--3007

    2017

  37. [45]

    J.; Hays, J.; Perona, P.; Ramanan, D.; Doll \' a r, P.; and Zitnick, C

    Lin, T.; Maire, M.; Belongie, S. J.; Hays, J.; Perona, P.; Ramanan, D.; Doll \' a r, P.; and Zitnick, C. L. 2014. Microsoft COCO: Common Objects in Context. In ECCV , 740--755

    2014

  38. [46]

    Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In ICLR

    2019

  39. [47]

    Lu, S.; Liu, Y.; and Kong, A. W.-K. 2023. Tf-icon: Diffusion-based training-free cross-domain image composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2294--2305

    2023

  40. [48]

    Lu, S.; Wang, Z.; Li, L.; Liu, Y.; and Kong, A. W.-K. 2024 a . Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6430--6440

    2024

  41. [49]

    Lu, S.; Zhou, Z.; Lu, J.; Zhu, Y.; and Kong, A. W.-K. 2024 b . Robust watermarking using generative priors against image editing: From benchmarking to advances. arXiv preprint arXiv:2410.18775

    2024 arXiv

  42. [50]

    P.; and Gool, L

    Mayer, C.; Danelljan, M.; Paudel, D. P.; and Gool, L. V. 2021. Learning Target Candidate Association to Keep Track of What Not to Track. In ICCV , 13424--13434. IEEE

    2021

  43. [51]

    Mueller, M.; Smith, N.; and Ghanem, B. 2016. A Benchmark and Simulator for UAV Tracking. In ECCV , 445--461

    2016

  44. [52]

    M \" u ller, M.; Bibi, A.; Giancola, S.; Al - Subaihi, S.; and Ghanem, B. 2018. TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild. In ECCV , 310--327

    2018

  45. [53]

    Oord, A. v. d.; Li, Y.; and Vinyals, O. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

    2018 arXiv

  46. [54]

    Peng, L.; Cao, Y.; Sun, Y.; and Wang, Y. 2024. Lightweight adaptive feature de-drifting for compressed image classification. IEEE Transactions on Multimedia, 26: 6424--6436

    2024

  47. [55]

    Peng, L.; Di, X.; Feng, Z.; Li, W.; Pei, R.; Wang, Y.; Fu, X.; Cao, Y.; and Zha, Z.-J. 2025 a . Directing Mamba to Complex Textures: An Efficient Texture-Aware State Space Model for Image Restoration. arXiv preprint arXiv:2501.16583

    2025 arXiv

  48. [56]

    Peng, L.; Li, W.; Pei, R.; Ren, J.; Xu, J.; Wang, Y.; Cao, Y.; and Zha, Z.-J. 2025 b . Towards Realistic Data Generation for Real-World Super-Resolution. In The Thirteenth International Conference on Learning Representations

    2025

  49. [57]

    Peng, L.; Wang, Y.; Di, X.; Fu, X.; Cao, Y.; Zha, Z.-J.; et al. 2025 c . Boosting image de-raining via central-surrounding synergistic convolution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 6470--6478

    2025

  50. [58]

    D.; and Savarese, S

    Rezatofighi, H.; Tsoi, N.; Gwak, J.; Sadeghian, A.; Reid, I. D.; and Savarese, S. 2019. Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression. In CVPR , 658--666

    2019

  51. [59]

    Shi, L.; Zhong, B.; Liang, Q.; Li, N.; Zhang, S.; and Li, X. 2024. Explicit Visual Prompts for Visual Object Tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 4838--4846

    2024

  52. [60]

    H.; Ma, Y.-J.; Shuai, H.-H.; Chen, J.-C.; and Cheng, W.-H

    Sio, C. H.; Ma, Y.-J.; Shuai, H.-H.; Chen, J.-C.; and Cheng, W.-H. 2020. S2siamfc: Self-supervised fully convolutional siamese network for visual tracking. In Proceedings of the 28th ACM international conference on multimedia, 1948--1957

    2020

  53. [61]

    Voigtlaender, P.; Luiten, J.; Torr, P. H. S.; and Leibe, B. 2020. Siam R-CNN: Visual Tracking by Re-Detection. In CVPR , 6577--6587

    2020

  54. [62]

    Wang, N.; Zhou, W.; Wang, J.; and Li, H. 2021 a . Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 1571--1580

    2021

  55. [63]

    Wang, X.; Jabri, A.; and Efros, A. A. 2019. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2566--2576

    2019

  56. [64]

    Wang, X.; Shu, X.; Zhang, Z.; Jiang, B.; Wang, Y.; Tian, Y.; and Wu, F. 2021 b . Towards More Flexible and Accurate Object Tracking With Natural Language: Algorithms and Benchmark. In CVPR , 13763--13773

    2021

  57. [65]

    Wu, Q.; Yang, T.; Liu, Z.; Wu, B.; Shan, Y.; and Chan, A. B. 2023. DropMAE: Masked Autoencoders with Spatial-Attention Dropout for Tracking Tasks. CVPR, abs/2304.00571

    2023 arXiv

  58. [66]

    Wu, Y.; Lim, J.; and Yang, M. 2015. Object Tracking Benchmark. IEEE Trans. Pattern Anal. Mach. Intell. , 37(9): 1834--1848

    2015

  59. [67]

    Xiao, F.; Hu, S.; Shen, Y.; Fang, C.; Huang, J.; He, C.; Tang, L.; Yang, Z.; and Li, X. 2024. A Survey of Camouflaged Object Detection and Beyond. CAAI AIR

    2024

  60. [68]

    Xie, F.; Chu, L.; Li, J.; Lu, Y.; and Ma, C. 2023. VideoTrack: Learning to Track Objects via Video Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 22826--22835

    2023

  61. [69]

    Xie, F.; Wang, C.; Wang, G.; Cao, Y.; Yang, W.; and Zeng, W. 2022. Correlation-Aware Deep Tracking. In CVPR , 8741--8750

    2022

  62. [70]

    Xie, J.; Zhong, B.; Mo, Z.; Zhang, S.; Shi, L.; Song, S.; and Ji, R. 2024. Autoregressive Queries for Adaptive Tracking with Spatio-Temporal Transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19300--19309

    2024

  63. [71]

    Xing, W.; Yifan, B.; Yongchao, Z.; Dahu, S.; and Yihong, G. 2023. Autoregressive Visual Tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9697--9706

    2023

  64. [72]

    Xu, C.; Zhong, B.; Liang, Q.; Zheng, Y.; Li, G.; and Song, S. 2025. Less is more: Token context-aware learning for object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 8824--8832

    2025

  65. [73]

    Yan, B.; Peng, H.; Fu, J.; Wang, D.; and Lu, H. 2021. Learning Spatio-Temporal Transformer for Visual Tracking. In ICCV, 10428--10437

    2021

  66. [74]

    Ye, B.; Chang, H.; Ma, B.; Shan, S.; and Chen, X. 2022. Joint Feature Learning and Relation Modeling for Tracking: A One-Stream Framework. In ECCV (22) , 341--357

    2022

  67. [75]

    Yuan, D.; Chang, X.; Huang, P.-Y.; Liu, Q.; and He, Z. 2020. Self-supervised deep correlation tracking. IEEE Transactions on Image Processing, 30: 976--985

    2020

  68. [76]

    Y.; and Chen, Q

    Yuan, W.; Wang, M. Y.; and Chen, Q. 2020. Self-supervised object tracking with cycle-consistent siamese networks. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 10351--10358. IEEE

    2020

  69. [77]

    Zhang, Z.; and Peng, H. 2019. Deeper and Wider Siamese Networks for Real-Time Visual Tracking. In CVPR , 4591--4600

    2019

  70. [78]

    Zhang, Z.; Peng, H.; Fu, J.; Li, B.; and Hu, W. 2020. Ocean: Object-Aware Anchor-Free Tracking. In ECCV , 771--787

    2020

  71. [79]

    Zheng, Y.; Zhong, B.; Liang, Q.; Li, G.; Ji, R.; and Li, X. 2023. Toward unified token learning for vision-language tracking. IEEE Transactions on Circuits and Systems for Video Technology, 34(4): 2125--2135

    2023

  72. [80]

    Zheng, Y.; Zhong, B.; Liang, Q.; Mo, Z.; Zhang, S.; and Li, X. 2024. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 7588--7596

    2024

  73. [81]

    Zheng, Y.; Zhong, B.; Liang, Q.; Tang, Z.; Ji, R.; and Li, X. 2022. Leveraging local and global cues for visual tracking via parallel interaction network. IEEE Transactions on Circuits and Systems for Video Technology, 33(4): 1671--1683

    2022

Pith tools

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