Pith. sign in

REVIEW 5 major objections 5 minor 74 references

Towards Compact Unified Multimodal Tracking: Synergizing Knowledge Distillation with Structural Pruning

T0 review · 5 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 A pruned-head multimodal tracker runs 54 FPS at 91.5% MPR on RGBT234 by distilling spatial and semantic knowledge separately from a heavy teacher.

desk verdict Useful 17-way distillation comparison for multimodal tracking, but the headline 'superior accuracy' claim is contradicted by the paper's own tables and the head-pruning ablation is missing. read the letter →

arxiv 2608.01488 v1 pith:TIAYBQAO submitted 2026-08-02 cs.CV

Yuqi Li , Yuedong Tan , Huiran Duan , Weilun Feng , Chuanguang Yang , Zhulin An , Zongwei Wu , Shiping Wen
show 2 more authors
Tingwen Huang Yingli Tian
This is my paper · ORCID
classification cs.CV
keywords multimodalobjecttrackingknowledgedistillationstructuralpruningpredictionheadreal-timeinferenceRGB-thermalRGB-depthRGB-event
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 argues that the prediction head, not the backbone, is the overlooked bottleneck in unified multimodal tracking, and that deleting two of its five convolutional layers unlocks real-time speed without sacrificing accuracy. To repair the capacity gap the pruning creates, it introduces Dual-Alignment Distillation, which separates what the student should learn into a spatial stream (L2-normalized feature matching, 'where to track') and a semantic stream (standardized logits matched by KL divergence, 'what to track'). In a systematic study of 17 distillation strategies across five benchmarks spanning RGB-thermal, RGB-depth, and RGB-event data, the two streams together beat either alone. The result is DTrack, an 18M-parameter tracker that runs at 54 FPS on an RTX 4090 and reaches 91.5% MPR on RGBT234, about a 5x speedup over its SUTrack-B224 teacher while the paper claims accuracy stays on par with it.

What carries the argument

The load-bearing object is the pair of distillation losses in Eqs. (1)-(5). Spatial Structure Alignment projects student and teacher encoder features onto a unit hypersphere and minimizes MSE between them, so the student learns the teacher's directional attention independently of activation magnitude. Semantic Distribution Alignment standardizes features channel-wise, applies softmax with temperature T over spatial positions, and minimizes KL divergence, so the student inherits the teacher's 'dark knowledge' about target-distractor relationships. The second mechanism is structural pruning of the prediction head: removing layers 2 and 4 from the standard 5-layer branch (Eq. 7 becomes Eq. 8),

What would settle it

Take the DTrack student trained with the unpruned 5-layer head and the same dual-alignment distillation, run it on RGBT234, and compare MPR with the pruned 3-layer version. If the pruned head's MPR falls by more than the run-to-run noise, the claim that head pruning is loss-free fails; the paper does not report this comparison, so it is the decisive missing measurement.

Watch

Extended reading notes

Core claim

DTrack's central claim is that a unified multimodal tracker can be made real-time by pruning the prediction head from five convolutional layers to three (dropping layers 2 and 4 in every task branch) without losing accuracy, provided the student is distilled with two complementary objectives. The spatial objective normalizes encoder feature maps to unit length and minimizes the mean squared error between student and teacher, teaching the student where the foreground is. The semantic objective standardizes the same features, replaces magnitude with relative confidence, and aligns the resulting spatial softmax distributions via KL divergence with temperature scaling, transferring the teacher's

Load-bearing premise

The paper assumes that deleting the second and fourth convolutional layers of the prediction head costs no meaningful accuracy, but it never reports the pruned head vs unpruned head accuracy under otherwise identical training; if the deletion does cost accuracy, the advertised 54 FPS includes an unmeasured penalty.

Editorial extensions

If this is right

  • The prediction head should become a first-class target in efficient tracker design; head pruning yields 7 FPS of the total 54 FPS in this work.
  • Decoupling spatial and semantic distillation sets a new baseline: on RGBT234, DTrack's 91.5% MPR beats the best spatial-only (CAD, 89.5%) and best semantic-only (LOGIT, 90.8%) strategies.
  • A single encoder jointly trained on RGB-T, RGB-E, and RGB-D can serve all three modalities at real-time speed, enabling deployment on drones and robots.
  • The deeper the distilled feature layer (up to the final encoder layer), the better the student's precision, with no added inference latency.
  • The 17-strategy comparison provides a map of which distillation families help tracking and which do not, useful to future work.

Reading between the lines

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

  • A natural extension is to apply the same spatial/semantic decoupling to RGB-only tracking or to video-object segmentation, where head-heavy decoders dominate inference cost; the head-pruning recipe may transfer directly.
  • The paper reports no edge-hardware measurements. The 5x speedup is on an RTX 4090; on memory-bandwidth-limited edge devices, removing layers 2 and 4 may save a different fraction, so the 'edge deployment' conclusion is an extrapolation.
  • Because the layer-depth ablation shows no latency cost for any distillation target, an adaptive policy that picks the distillation layer per input difficulty could squeeze accuracy further without hurting speed.
  • The claim that spatial-only and semantic-only streams are 'mutually reinforcing' is only shown additively; a factorial analysis with different weighting hyperparameters would reveal whether there is a sweet spot or a plateau.
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

5 major / 5 minor

Summary. The paper proposes DTrack, a compact unified multimodal tracker obtained by compressing SUTrack-B224 through two means: (1) structural pruning of the prediction head (removing layers 2 and 4 from each task branch, Eqs. 7–8) and (2) a Dual-Alignment Distillation framework with spatial feature alignment (Eq. 2) and semantic logit alignment (Eq. 5). The authors report a systematic comparison of 17 distillation strategies on RGB-Event, RGB-Thermal, and RGB-Depth benchmarks, claiming state-of-the-art accuracy among distillation methods at 54 FPS, a 5× speedup over the teacher.

Significance. The paper addresses a real deployment problem for unified multimodal trackers and provides one of the first systematic studies of knowledge-distillation mechanisms in this setting. The dual-alignment idea is credible and the efficiency measurements are reported under a consistent protocol. If the central claim—that head pruning yields 'negligible accuracy loss' and that the two distillation streams synergize—were supported by a head-pruning ablation, the contribution would be solid. However, that ablation is missing, and the abstract overstates the accuracy relationship to the teacher, so the significance is currently conditional on new experiments.

major comments (5)
  1. [§III-B, §IV-B, Table I] The central efficiency claim is that removing layers 2 and 4 from the prediction head increases FPS from 47 to 54 with 'negligible accuracy loss' (Sec. IV-B). Table I reports parameters/FLOPs/FPS for 'Ours (w/o Head Pruning)' and 'Ours', but no accuracy is reported for the unpruned-head student on any benchmark. Without this ablation, the speed gain cannot be separated from an unmeasured accuracy cost, and the claimed synergy between structural pruning and dual-alignment distillation is not established. Please add per-benchmark accuracy for 'Ours (w/o Head Pruning)' and 'Ours' (at least RGBT234, VisEvent, LasHeR, DepthTrack, VOT-RGBD22).
  2. [Abstract, Table VI] The abstract and contributions state that DTrack achieves a 5× speedup over the teacher 'while maintaining superior accuracy.' Table VI contradicts this: SUTrack-B224 outperforms DTrack on every headline benchmark (VisEvent PR 79.9 vs 79.7; LasHeR PR 74.5 vs 72.4; RGBT234 MPR 92.2 vs 91.5; DepthTrack F-score 65.1 vs 57.1; VOT-RGBD22 EAO 76.5 vs 74.7). At best, DTrack is competitive on some RGB-T and RGB-E metrics, and it is substantially worse on DepthTrack. Please correct the claim and characterize the accuracy-efficiency trade-off precisely.
  3. [§IV-C, Table IV] The RGB-D results are overclaimed. The text says DTrack 'maintains competitive performance' on DepthTrack, but Table IV shows DTrack's F-score (57.1) is below OneTracker (60.9), SDSTrack (61.4), UnTrack (61.0), ViPT (59.4), and ProTrack (57.8). Similarly, Table VI lists DepthTrack F-score 57.1 versus the teacher's 65.1. Given that the paper claims state-of-the-art across five benchmarks, this substantial deficit on a core benchmark should be acknowledged and discussed, not downplayed.
  4. [§V, Table V] The distillation target layer is selected by an ablation on RGBT234 (Table V), and then the final model with the chosen Layer 12 is evaluated and reported as a main result on the same RGBT234 benchmark (Tables III and VI). This is a form of test-set tuning and inflates the reported performance. Please either use a held-out validation split for layer selection or disclose the tuning and treat the RGBT234 numbers as partially in-sample.
  5. [§V, Table VI] The paper states '17 distillation strategies' multiple times, but Table VI lists 16 named distillation baselines (CWD, L2, FNKD, SP, CUSTOMKD, CBAM, CAD, StrucKD, DSD as Spatial; KD, DHO, DKD, GAN, LOGIT, MTKD, RKD as Semantic) plus DTrack. Clarify whether the count includes the proposed method or if one baseline is missing. This affects the claimed comprehensiveness of the study.
minor comments (5)
  1. [Abstract / §IV-B] The abstract and contributions say '5× speedup', while §IV-B correctly reports '4.9× end-to-end speedup' (11 FPS to 54 FPS). Please make the numbers consistent.
  2. [Table IV] The DTrack row reads '57.1 56.9 57.074.7 82.0 90.5'; this appears to lack a space between the PR value (57.0) and the EAO (74.7). Fix the formatting.
  3. [Table VII] The table caption says 'OUR METHOD (DTTRACK)'—typo: should be DTrack.
  4. [References] References [12] and [14] are the same paper; [27] and [28] are also duplicates. Please deduplicate.
  5. [§III-A, Eq. (5)] The softmax in Eq. (5) is stated to be applied 'over the spatial dimensions', but the notation softmax(·) typically implies channel-wise application, and the dimensionality of the resulting probability distribution is ambiguous. Specify the axis explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: losses and pruning equations are externally defined and benchmark numbers are empirical; the main weaknesses are missing ablations and test-set tuning, not circular reasoning.

full rationale

I find no circular step in this paper. The dual-alignment losses (Eqs. 1-5) and the structural pruning formulas (Eqs. 7-8) are externally defined objectives/architectural transformations, and the reported accuracy (91.5% MPR, etc.) and speed (54 FPS) are empirical measurements, not quantities derived from the equations by construction. There is no load-bearing self-citation: the cited works by the authors (e.g., [2], [12]/[14]) are contextual, not used to justify the central claim. The most defensible concerns are correctness/selection issues rather than circularity: (i) Sec. IV-B asserts head pruning gives 'negligible accuracy loss' without reporting a pruned-vs-unpruned accuracy ablation; (ii) the distillation layer is selected via ablation on RGBT234 (Table V), the same benchmark used for final RGBT234 reporting; and (iii) the abstract's 'superior accuracy' is contradicted by Table VI, where DTrack is below SUTrack-B224 on all five headline benchmarks. These are legitimate methodological and factual concerns, but none involves a prediction reducing to its input by definition or by self-citation.

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

No new physical or architectural entities are introduced. The method relies on established distillation losses (L2-normalized feature MSE and KL divergence) plus a hand-chosen head pruning pattern. The free parameters are the loss weights, temperature, pruning depth, and distillation layer, none of which are reported as values or sensitivity studies.

free parameters (5)
  • lambda_spa = not reported
    Balancing weight for spatial alignment loss in Eq. (6); no value or sensitivity study given.
  • lambda_sem = not reported
    Balancing weight for semantic alignment loss in Eq. (6); no value or sensitivity study given.
  • temperature T = not reported
    Softening temperature in Eq. (5); affects the semantic KL distribution and is not specified.
  • pruned head layers (2 and 4) = layers 2 and 4 removed
    Structural pruning choice in Section III-B; chosen by design, not by an empirical search, and no pruning-depth ablation is shown.
  • distillation target layer (Layer 12) = Layer 12 (final encoder layer)
    Selected via ablations on RGBT234 (Table V); this tunes the method on the same benchmark used for final results.
assumptions (5)
  • domain assumption L2 normalization along the channel dimension preserves spatial structural attention while removing magnitude differences
    Section III-A.1, Eq. (1); the paper assumes this normalization makes teacher and student features comparable.
  • domain assumption Channel-wise standardization followed by softmax over spatial dimensions yields a meaningful semantic probability distribution
    Section III-A.2, Eq. (3)-(5); used to define semantic alignment via KL divergence.
  • domain assumption The teacher's encoder features are a reliable target for distillation
    Section III-A; the whole method depends on SUTrack-B224 features being worth mimicking.
  • ad hoc to paper Removing intermediate convolutional layers 2 and 4 from the prediction head retains essential task-specific capabilities
    Section III-B, Eq. (7)-(8); no accuracy ablation supports this.
  • domain assumption Joint training on LasHeR, VisEvent, and DepthTrack with masking simulations transfers to a unified model
    Section IV-A; the training recipe is stated but its effect is not isolated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Compact Unified Multimodal Tracking: Synergizing Knowledge Distillation with Structural Pruning." pith.science (2026). https://pith.science/paper/TIAYBQAO

@misc{pith2026260801488,
  author       = {Pith},
  title        = {Pith review of: Towards Compact Unified Multimodal Tracking: Synergizing Knowledge Distillation with Structural Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TIAYBQAO}},
  note         = {Machine review of arXiv:2608.01488}
}
read the original abstract

Unified multimodal object tracking has achieved remarkable robustness by leveraging complementary sensor data (e.g., RGB, Thermal, Depth), yet the heavy computational burden of state-of-the-art models hinders their deployment on resource-constrained edge devices. In this work, we identify the prediction head as a critical but often overlooked efficiency bottleneck. By strategically streamlining the decoder architecture, we unlock the potential for real-time inference but simultaneously introduce a capacity gap between the lightweight student and the heavy teacher. To resolve this, we conduct a systematic analysis of 17 distillation strategies and introduce a Dual-Alignment Distillation framework. Our key insight is that effective compression requires decoupling knowledge transfer into two complementary streams: (1) Spatial Representation Alignment, which employs feature distillation to sharpen the student's spatial focus on foreground targets ("Where to track"); and (2) Semantic Distribution Alignment, which utilizes logit-based distillation to align decision boundaries and transfer discriminative dark knowledge ("What to track"). Extensive experiments across five benchmarks demonstrate that our approach significantly outperforms complex state-of-the-art methods. Notably, our distilled model achieves 91.5% MPR on RGBT234 and operates at 54 FPS on a single RTX 4090, representing a 5x speedup over the teacher model while maintaining superior accuracy.

Figures

Figures reproduced from arXiv: 2608.01488 by the authors.

Figure 1
Figure 1. Speed & Performance comparison with state-of-the-art methods on RGBT234 dataset: Our method (red bubble) achieves an optimal trade-off between accuracy and efficiency, attaining state-of-the-art precision comparable to the heavyweight teacher model (SUTrack-B224 [5]) while operating at real-time speeds (54 FPS). Compared to other trackers (e.g., UnTrack [6], SDSTrack [7]) and high-performance baselines (e.g., CAForm… view at source ↗
Figure 2
Figure 2. Overview of our proposed DTrack framework. The architecture is designed for unified multimodal tracking (handling RGB-D, RGB-E, and RGB-T inputs). We introduce a Lightweight Prediction Head by structurally pruning the intermediate convolutional layers (visualized as dashed blocks) from the standard heavy decoder. This design significantly reduces computational redundancy, unlocking real-time inference speeds. To com… view at source ↗
Figure 3
Figure 3. Attribute-based comparison on the RGBT234 dataset (MPR). The radar chart illustrates the performance of different distillation strategies across 12 challenging attributes (e.g., Fast Motion, Occlusion, Low Illumination). Our proposed DTrack consistently encompasses other methods, demonstrating superior robustness across diverse tracking scenarios. (R), and F-score (F). VOT-RGBD22 is evaluated using the Expected Aver… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 68 canonical work pages

  1. [1]

    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,” inECCV, 2022

    work page 2022

  2. [2]

    Hhtrack: Hyperspectral object tracking using hybrid attention,

    Y . Tan, “Hhtrack: Hyperspectral object tracking using hybrid attention,” 2023

    work page 2023

  3. [3]

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

    Y . Cui, C. Jiang, L. Wang, and G. Wu, “Mixformer: End-to-end tracking with iterative mixed attention,” inCVPR, 2022

    work page 2022

  4. [4]

    Cafuser: Condition- aware multimodal fusion for robust semantic perception of driving scenes,

    T. Br¨odermann, C. Sakaridis, Y . Fu, and L. Van Gool, “Cafuser: Condition- aware multimodal fusion for robust semantic perception of driving scenes,” IEEE Robotics and Automation Letters, 2025

    work page 2025

  5. [5]

    Sutrack: Towards simple and unified single object tracking,

    X. Chen, B. Kang, W. Geng, J. Zhu, Y . Liu, D. Wang, and H. Lu, “Sutrack: Towards simple and unified single object tracking,”arXiv preprint arXiv:2412.19138, 2024

    arXiv 2024

  6. [6]

    Single-model and any-modality for video object tracking,

    Z. Wu, J. Zheng, X. Ren, F.-A. Vasluianu, C. Ma, D. P. Paudel, L. Van Gool, and R. Timofte, “Single-model and any-modality for video object tracking,” inCVPR, 2024

    work page 2024

  7. [7]

    Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking,

    X. Hou, J. Xing, Y . Qian, Y . Guo, S. Xin, J. Chen, K. Tang, M. Wang, Z. Jiang, L. Liuet al., “Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking,” inCVPR, 2024

    work page 2024

  8. [8]

    Cross-modality distillation for multi-modal tracking,

    T. Zhang, Q. Zhang, K. Debattista, and J. Han, “Cross-modality distillation for multi-modal tracking,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025

Show all 74 references
  1. [9]

    Bi-directional adapter for multimodal tracking,

    B. Cao, J. Guo, P. Zhu, and Q. Hu, “Bi-directional adapter for multimodal tracking,” inAAAI, vol. 38, no. 2, 2024

    2024

  2. [10]

    Visual prompt multi-modal tracking,

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

    2023

  3. [11]

    Onetracker: Unifying visual object tracking with foundation models and efficient tuning,

    L. Hong, S. Yan, R. Zhang, W. Li, X. Zhou, P. Guo, K. Jiang, Y . Chen, J. Li, Z. Chenet al., “Onetracker: Unifying visual object tracking with foundation models and efficient tuning,” inCVPR, 2024

    2024

  4. [13]

    Xtrack: Multimodal training boosts rgb-x video object trackers,

    Y . Tan, Z. Wu, Y . Fu, Z. Zhou, G. Sun, E. Zamfir, C. Ma, D. Paudel, L. Van Gool, and R. Timofte, “Xtrack: Multimodal training boosts rgb-x video object trackers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 5734–5744

    2025

  5. [14]

    What you have is what you track: Adaptive and robust multimodal tracking,

    Y . Tan, J. Shao, E. Zamfir, R. Li, Z. An, C. Ma, D. Paudel, L. Van Gool, R. Timofte, and Z. Wu, “What you have is what you track: Adaptive and robust multimodal tracking,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 3455–3465

    2025

  6. [15]

    Exploiting multimodal spatial-temporal patterns for video object tracking,

    X. Hu, Y . Tai, X. Zhao, C. Zhao, Z. Zhang, J. Li, B. Zhong, and J. Yang, “Exploiting multimodal spatial-temporal patterns for video object tracking,” arXiv preprint arXiv:2412.15691, 2024

    2024 arXiv

  7. [16]

    Mambaevt: Event stream-based visual object tracking using state space model,

    X. Wang, C. Wang, S. Wang, X. Wang, Z.-C. Zhao, L. Zhu, and B. Jiang, “Mambaevt: Event stream-based visual object tracking using state space model,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 36, no. 1, pp. 278–291, 2026

    2026

  8. [17]

    Self-supervised learning for rgb-d object tracking,

    X.-F. Zhu, T. Xu, S. Atito, M. Awais, X.-J. Wu, Z. Feng, and J. Kittler, “Self-supervised learning for rgb-d object tracking,”Pattern Recognition, vol. 155, p. 110543, 2024

    2024

  9. [18]

    Fmtrack: Frequency-aware interaction and multi-expert fusion for rgb- t tracking,

    Y . Xue, G. Jin, B. Zhong, T. Shen, L. Tan, C. Xue, and Y . Zheng, “Fmtrack: Frequency-aware interaction and multi-expert fusion for rgb- t tracking,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 36, no. 2, pp. 1655–1667, 2026

    2026

  10. [19]

    Cross-modal orthogonal high-rank augmentation for rgb-event transformer-trackers,

    Z. Zhu, J. Hou, and D. O. Wu, “Cross-modal orthogonal high-rank augmentation for rgb-event transformer-trackers,” inICCV, 2023

    2023

  11. [20]

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

    B. Yan, H. Peng, K. Wu, D. Wang, J. Fu, and H. Lu, “Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15 180–15 189

    2021

  12. [21]

    Litetrack: Layer pruning with asynchronous feature extraction for lightweight and efficient visual tracking,

    Q. Wei, B. Zeng, J. Liu, L. He, and G. Zeng, “Litetrack: Layer pruning with asynchronous feature extraction for lightweight and efficient visual tracking,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 4968–4975

    2024

  13. [22]

    Mixformerv2: Efficient fully transformer tracking,

    Y . Cui, T. Song, G. Wu, and L. Wang, “Mixformerv2: Efficient fully transformer tracking,”Advances in neural information processing systems, vol. 36, pp. 58 736–58 751, 2023

    2023

  14. [23]

    Emtrack: Efficient multimodal object tracking,

    C. Liu, Z. Guan, S. Lai, L. Feng, H. Lu, and D. Wang, “Emtrack: Efficient multimodal object tracking,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 3, pp. 2202–2214, 2025

    2025

  15. [24]

    Exploring pruning-based efficient object tracking via hybrid knowledge distillation,

    Y . Song, S. Wang, Z. Zeng, J. Zheng, Z. Wang, and J. Ning, “Exploring pruning-based efficient object tracking via hybrid knowledge distillation,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 36, no. 2, pp. 2433–2448, 2026

    2026

  16. [25]

    Distilling the knowledge in a neural network,

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

    2015 arXiv

  17. [26]

    Decoupled knowledge distillation,

    B. Zhao, Q. Cui, R. Song, Y . Qiu, and J. Liang, “Decoupled knowledge distillation,” inProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2022, pp. 11 953–11 962

    2022

  18. [27]

    Relational knowledge distillation,

    W. Park, D. Kim, Y . Lu, and M. Cho, “Relational knowledge distillation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3967–3976

    2019

  19. [28]

    Relational knowledge distillation,

    ——, “Relational knowledge distillation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3967–3976

    2019

  20. [29]

    Learning from multiple teacher networks,

    S. You, C. Xu, C. Xu, and D. Tao, “Learning from multiple teacher networks,” inProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2017, pp. 1285– 1294

    2017

  21. [30]

    Structured knowledge distillation for semantic segmentation,

    Y . Liu, K. Chen, C. Liu, Z. Qin, Z. Luo, and J. Wang, “Structured knowledge distillation for semantic segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2604–2613

    2019

  22. [31]

    Double similarity distillation for semantic image segmentation,

    Y . Feng, X. Sun, W. Diao, J. Li, and X. Gao, “Double similarity distillation for semantic image segmentation,”IEEE Transactions on Image Processing, vol. 30, pp. 5363–5376, 2021

    2021

  23. [32]

    Prompting for multi-modal tracking,

    J. Yang, Z. Li, F. Zheng, A. Leonardis, and J. Song, “Prompting for multi-modal tracking,” inACMMM, 2022

    2022

  24. [33]

    Siam R-CNN: Visual tracking by re-detection,

    P. V oigtlaender, J. Luiten, P. H. S. Torr, and B. Leibe, “Siam R-CNN: Visual tracking by re-detection,” inCVPR, 2020

    2020

  25. [34]

    Transformer tracking,

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

    2021

  26. [35]

    High-performance long-term tracking with meta-updater,

    K. Dai, Y . Zhang, D. Wang, J. Li, H. Lu, and X. Yang, “High-performance long-term tracking with meta-updater,” inCVPR, 2020

    2020

  27. [36]

    Probabilistic regression for visual tracking,

    M. Danelljan, L. V . Gool, and R. Timofte, “Probabilistic regression for visual tracking,” inCVPR, 2020

    2020

  28. [37]

    Vital: Visual tracking via adversarial learning,

    Y . Song, C. Ma, X. Wu, L. Gong, L. Bao, W. Zuo, C. Shen, R. W. Lau, and M.-H. Yang, “Vital: Visual tracking via adversarial learning,” in CVPR, 2018

    2018

  29. [38]

    Learning multi-domain convolutional neural networks for visual tracking,

    H. Nam and B. Han, “Learning multi-domain convolutional neural networks for visual tracking,” inCVPR, 2016

    2016

  30. [39]

    Atom: Accurate tracking by overlap maximization,

    M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg, “Atom: Accurate tracking by overlap maximization,” inCVPR, 2019

    2019

  31. [40]

    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

    2021

  32. [41]

    Siamban: Target-aware tracking with siamese box adaptive network,

    Z. Chen, B. Zhong, G. Li, S. Zhang, R. Ji, Z. Tang, and X. Li, “Siamban: Target-aware tracking with siamese box adaptive network,”TPAMI, vol. 45, no. 4, pp. 5158–5173, 2022

    2022

  33. [42]

    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,” inCVPR, 2019

    2019

  34. [43]

    Attribute-based progressive fusion network for rgbt tracking,

    Y . Xiao, M. Yang, C. Li, L. Liu, and J. Tang, “Attribute-based progressive fusion network for rgbt tracking,” inAAAI, vol. 36, no. 3, 2022

    2022

  35. [44]

    Cross-modal pattern-propagation for rgb-t tracking,

    C. Wang, C. Xu, Z. Cui, L. Zhou, T. Zhang, X. Zhang, and J. Yang, “Cross-modal pattern-propagation for rgb-t tracking,” inCVPR, 2020

    2020

  36. [45]

    Jointly modeling motion and appearance cues for robust RGB-T tracking,

    P. Zhang, J. Zhao, C. Bo, D. Wang, H. Lu, and X. Yang, “Jointly modeling motion and appearance cues for robust RGB-T tracking,”TIP, pp. 3335–3347, 2021

    2021

  37. [46]

    Challenge-aware RGBT tracking,

    C. Li, L. Liu, A. Lu, Q. Ji, and J. Tang, “Challenge-aware RGBT tracking,” inECCV, 2020

    2020

  38. [47]

    Quality-aware feature aggregation network for robust RGBT tracking,

    Y . Zhu, C. Li, J. Tang, and B. Luo, “Quality-aware feature aggregation network for robust RGBT tracking,”TIV, pp. 121–130, 2020

    2020

  39. [48]

    Dense feature aggregation and pruning for RGBT tracking,

    Y . Zhu, C. Li, B. Luo, J. Tang, and X. Wang, “Dense feature aggregation and pruning for RGBT tracking,” inACMMM, 2019

    2019

  40. [49]

    Deep adaptive fusion network for high performance RGBT tracking,

    Y . Gao, C. Li, Y . Zhu, J. Tang, T. He, and F. Wang, “Deep adaptive fusion network for high performance RGBT tracking,” inICCVW, 2019

    2019

  41. [50]

    Object tracking in RGB-T videos using modal-aware attention network and competitive learning,

    H. Zhang, L. Zhang, L. Zhuo, and J. Zhang, “Object tracking in RGB-T videos using modal-aware attention network and competitive learning,” Sensors, p. 393, 2020

    2020

  42. [51]

    Exploring enhanced contextual information for video-level object tracking,

    B. Kang, X. Chen, S. Lai, Y . Liu, Y . Liu, and D. Wang, “Exploring enhanced contextual information for video-level object tracking,”arXiv preprint arXiv:2412.11023, 2024

    2024 arXiv

  43. [52]

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

    Y . Tian, L. Xie, J. Qiu, J. Jiao, Y . Wang, Q. Tian, and Q. Ye, “Fast-itpn: Integrally pre-trained transformer pyramid network with token migration,” TPAMI, 2024

    2024

  44. [53]

    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,”TPAMI, vol. 43, no. 5, pp. 1562–1577, 2019

    2019

  45. [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. 10

    2018

  46. [55]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” inECCV, 2014

    2014

  47. [56]

    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

    2019

  48. [57]

    Lasher: A large-scale high-diversity benchmark for rgbt tracking,

    C. Li, W. Xue, Y . Jia, Z. Qu, B. Luo, J. Tang, and D. Sun, “Lasher: A large-scale high-diversity benchmark for rgbt tracking,”TIP, vol. 31, pp. 392–404, 2021

    2021

  49. [58]

    Visevent: Reliable object tracking via collaboration of frame and event flows,

    X. Wang, J. Li, L. Zhu, Z. Zhang, Z. Chen, X. Li, Y . Wang, Y . Tian, and F. Wu, “Visevent: Reliable object tracking via collaboration of frame and event flows,”TCYB, pp. 1–14, 2023

    2023

  50. [59]

    Depthtrack: Unveiling the power of rgbd tracking,

    S. Yan, J. Yang, J. K¨apyl¨a, F. Zheng, A. Leonardis, and J.-K. K ¨am¨ar¨ainen, “Depthtrack: Unveiling the power of rgbd tracking,” inICCV, 2021

    2021

  51. [60]

    Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking,

    X.-F. Zhu, T. Xu, Z. Tang, Z. Wu, H. Liu, X. Yang, X.-J. Wu, and J. Kittler, “Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking,” inAAAI, vol. 37, no. 3, 2023

    2023

  52. [61]

    Correlation- aware deep tracking,

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

    2022

  53. [62]

    The tenth visual object tracking vot2022 challenge results,

    M. Kristan, A. Leonardis, J. Matas, M. Felsberg, R. Pflugfelder, J.-K. K¨am¨ar¨ainen, H. J. Chang, M. Danelljan, L. ˇC. Zajc, A. Luke ˇziˇcet al., “The tenth visual object tracking vot2022 challenge results,” inECCVW, 2023

    2023

  54. [63]

    The eighth visual object tracking VOT2020 challenge results,

    M. Kristan, A. Leonardis, J. Matas, M. Felsberg, R. Pflugfelder, J.- K. K ¨am¨ar¨ainen, M. Danelljan, L. ˇC. Zajc, A. Luke ˇziˇc, O. Drbohlav et al., “The eighth visual object tracking VOT2020 challenge results,” in ECCVW, 2020

    2020

  55. [64]

    Dal: A deep depth-aware long-term tracker,

    Y . Qian, S. Yan, A. Lukeˇziˇc, M. Kristan, J.-K. K ¨am¨ar¨ainen, and J. Matas, “Dal: A deep depth-aware long-term tracker,” inICPR, 2021

    2021

  56. [65]

    The seventh visual object tracking VOT2019 challenge results,

    M. Kristan, J. Matas, A. Leonardis, M. Felsberg, R. Pflugfelder, J.- K. Kamarainen, L. Cehovin Zajc, O. Drbohlav, A. Lukezic, A. Berg et al., “The seventh visual object tracking VOT2019 challenge results,” inICCVW, 2019

    2019

  57. [66]

    Context-aware three-dimensional mean-shift with occlusion handling for robust object tracking in RGB-D videos,

    Y . Liu, X.-Y . Jing, J. Nie, H. Gao, J. Liu, and G.-P. Jiang, “Context-aware three-dimensional mean-shift with occlusion handling for robust object tracking in RGB-D videos,”TMM, pp. 664–677, 2018

    2018

  58. [67]

    Learning discrimi- native model prediction for tracking,

    G. Bhat, M. Danelljan, L. V . Gool, and R. Timofte, “Learning discrimi- native model prediction for tracking,” inICCV, 2019

    2019

  59. [68]

    Channel-wise knowledge distillation for dense prediction,

    C. Shu, Y . Liu, J. Gao, Z. Yan, and C. Shen, “Channel-wise knowledge distillation for dense prediction,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 5311–5320

    2021

  60. [69]

    Fitnets: Hints for thin deep nets,

    A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y . Bengio, “Fitnets: Hints for thin deep nets,”arXiv preprint arXiv:1412.6550, 2014

    2014 arXiv

  61. [70]

    Similarity-preserving knowledge distillation,

    F. Tung and G. Mori, “Similarity-preserving knowledge distillation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1365–1374

    2019

  62. [71]

    Customkd: Customizing large vision foundation for edge model improvement via knowledge distillation,

    J. Lee, D. Das, M. Hayat, S. Choi, K. Hwang, and F. Porikli, “Customkd: Customizing large vision foundation for edge model improvement via knowledge distillation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 25 176–25 186

    2025

  63. [72]

    Attention- guided feature distillation for semantic segmentation,

    A. M. Mansourian, A. Jalali, R. Ahmadi, and S. Kasaei, “Attention- guided feature distillation for semantic segmentation,”arXiv preprint arXiv:2403.05451, 2024

    2024 arXiv

  64. [73]

    Rethinking knowledge distillation with raw features for semantic segmentation,

    T. Liu, C. Chen, X. Yang, and W. Tan, “Rethinking knowledge distillation with raw features for semantic segmentation,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 1155–1164

    2024

  65. [74]

    Simple semi-supervised knowledge distillation from vision-language models via dual-head optimization,

    S. Kang, D. B. Lee, H. Jang, and S. J. Hwang, “Simple semi-supervised knowledge distillation from vision-language models via dual-head optimization,”arXiv preprint arXiv:2505.07675, 2025

    2025

  66. [75]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

    2020

Pith tools

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