Pith. sign in

REVIEW 4 major objections 6 minor 37 references

HiM2SAM: Enhancing SAM2 with Hierarchical Motion Estimation and Memory Optimization towards Long-term Tracking

T0 review · 4 major / 6 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 training-free combination of hierarchical motion estimation and a split long/short memory bank makes SAM2 materially more robust in long-term video tracking, with relative AUC gains of 9.6% on LaSOT and 7.2% on LaSOText using the large…

desk verdict A plausible training-free SAM2 upgrade with consistent gains, but the evaluation overstates SOTA and the key thresholds are tuned on the same test set, so the exact numbers need scrutiny. read the letter →

arxiv 2507.07603 v3 pith:TFSQG32W submitted 2025-07-10 cs.CV

Ruixiang Chen , Guolei Sun , Yawei Li , Jie Qin , Luca Benini This is my paper
classification cs.CV
keywords videoobjecttrackinglong-termSAM2hierarchicalmotionestimationmemorybankoptimizationtraining-freeKalmanfilterCoTracker3
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 tries to establish that SAM2's weaknesses in long-term video object tracking—occlusion, distractors, reappearance, fast motion—can be repaired without any training and with negligible runtime cost. Its proposal is to add a hierarchical motion estimator: a Kalman filter gives coarse box predictions on every frame, and a point tracker (CoTracker3) is invoked only on frames where confidence is low, providing pixel-level refinement. Simultaneously, the memory bank is split into short-term and long-term components, with motion-aware filtering for the short-term store and a directed-Hausdorff-distance heuristic selecting distractor-ambiguous frames for the long-term store. On LaSOT and LaSOText the large model improves over SAM2 by 9.6% and 7.2% relative AUC, and on VOT-LT2020 and VOT-LT2022 it reports the best zero-shot F-scores. If correct, the paper shows that the main bottlenecks of memory-based tracking lie in motion modeling and memory curation rather than in model capacity.

What carries the argument

The load-bearing machinery is a two-level scoring system for mask proposals. A coarse motion score $s_{\mathrm{coarse}}$ comes from IoU between the Kalman-filter predicted box and each SAM2 proposal, and is linearly combined with SAM2's IoU score via $\alpha$ (Eq. 1). When the combined score falls below threshold $\tau$, a fine motion score $s_{\mathrm{fine}}$ is computed by propagating farthest-point-sampled points backward through CoTracker3, reconstructing a soft mask with radial basis function kernels, and comparing to historical masks via the Dice coefficient (Eqs. 2 and 3). The same motion scores feed the memory filter: high-confidence frames go to short-term FIFO memory, while frames whose selected mask is far from alternative proposals under directed Hausdorff distance (Eq. 4) are stored as distinctive long-term memory to help later disambiguation.

What would settle it

On a held-out set of long LaSOT sequences with target disappearances, compare the full HiM2SAM-L against a variant whose long-term memory stores random high-confidence frames instead of Hausdorff-selected ones; if AUC does not drop, the distractor-selection heuristic is not the source of the gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that treating motion and memory as two coupled problems, rather than patching SAM2's components in isolation, yields consistent gains across all SAM2 model sizes. The authors demonstrate that coarse Kalman-based box prediction combined with selective fine-grained point tracking outperforms either alone, and that a memory bank separated into high-confidence short-term frames and distractor-selected long-term frames recovers targets after occlusion and appearance change better than SAM2's uniform FIFO memory. The improvements hold without retraining and with only a few milliseconds of added latency, and they are largest on the smallest models, where the motion and memory modules compensate for limited representation capacity.

Load-bearing premise

The load-bearing premise is that the hand-designed heuristics choose good memory: a large directed Hausdorff distance between SAM2's proposal masks reliably marks distractor ambiguity worth storing for later, and high IoU-plus-motion scores reliably mark frames safe to keep in short-term memory.

Editorial extensions

If this is right

  • If the reported gains replicate, SAM2-based trackers can be upgraded without retraining, preserving class-agnostic generalization.
  • Small SAM2 variants (Tiny and Small) benefit proportionally more, so lightweight edge deployments gain the most from the motion and memory modules.
  • The hierarchical trigger means the expensive point tracker runs on few frames, keeping real-time throughput; the reported +3.68 ms average latency on the large model is the cost of the full configuration.
  • The ablation shows the full model exceeds the sum of individual gains, suggesting motion estimation and memory curation reinforce each other rather than acting independently.

Reading between the lines

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

  • The paper leaves implicit that the same hierarchical gating should transfer to multi-object tracking with SAM2MOT-style pipelines, where expensive point tracking can be reserved for the few frames with low proposal confidence.
  • A directly testable extension is to replace the directed Hausdorff criterion with a learned or simpler proxy such as mask IoU spread, and check whether the long-term memory gains persist, which would reveal whether exact distance computation matters.
  • If the thresholds in the released code are not robust across datasets, the reported gains may depend on per-dataset tuning; sweeping $\tau$ and the memory intervals on VOT-LT would make that dependency visible.
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

4 major / 6 minor

Summary. The paper proposes HiM2SAM, a training-free enhancement of SAM2 for long-term video object tracking. It adds a hierarchical motion estimator that combines a lightweight Kalman-filter-based coarse prediction with a selectively triggered CoTracker3-based fine point-tracking refinement, and it splits SAM2's memory into short-term and long-term banks with motion-aware and distractor-aware (directed Hausdorff distance) selection rules. Experiments on LaSOT, LaSOText, VOT-LT2020, and VOT-LT2022 report AUC and F-score improvements over SAM2 baselines and prior zero-shot SAM2-based trackers across Tiny, Small, Base, and Large model scales, with the claimed gains attributed to better occlusion and reappearance handling.

Significance. If the reported results hold, the paper provides a practical, low-overhead recipe for improving SAM2-based long-term tracking without additional training, which would be valuable given SAM2's wide adoption. The strengths of the manuscript include multi-benchmark evaluation, an ablation study (Table 4) showing that each proposed module contributes, and consistent improvements across model scales. The authors also provide a code link, which supports reproducibility in principle. However, the paper overstates its state-of-the-art claim on LaSOT, does not report the numerical values of several key hyperparameters, and performs its sensitivity analysis on the same test set used for the headline results, which weakens confidence that the gains are not partly due to benchmark-specific threshold selection.

major comments (4)
  1. [§5.4, Figure 3] The sensitivity analysis is performed on the LaSOT test set, which is also the source of the headline AUC numbers in Table 1, and the final parameter values (α, β, τ, point-tracker window, LM/SM intervals, Hausdorff threshold) are not reported in the paper, only said to be in the released code. This makes it impossible to verify that the improvements are robust rather than the result of tuning on the evaluation set. Please report the exact threshold values in the manuscript and evaluate sensitivity on a held-out validation split (e.g., LaSOT train sequences or a different benchmark) or provide sequence-level error bars and a statistical comparison.
  2. [Abstract and §5.2, Table 1] The abstract claims "state-of-the-art performance on LaSOT and LaSOText with the large model," but Table 1 shows MCITrack-B achieves AUC 75.3 on LaSOT and DAM4SAM-L ties HiM2SAM-L at 75.1. The claim is therefore inaccurate for LaSOT; the method is state-of-the-art among zero-shot methods, not absolute SOTA. Please revise the wording to "best zero-shot performance" and explicitly acknowledge the tie with DAM4SAM-L and the higher supervised result.
  3. [§5.3, Table 4] The ablation reports latency averaged over a "representative LaSOT subset" and claims the full method adds only 3.68 ms, but the table also shows that applying the point tracker (PT) alone adds 165.95 ms. The average overhead therefore depends entirely on how often the fine-motion stage is triggered by the threshold τ. Without reporting the trigger rate or the distribution of per-frame latencies, the "low-overhead" claim is not fully supported. Please report the fraction of frames on which CoTracker3 is run and the variance of the latency.
  4. [§4.2, Eq. (4)] The long-term memory selection relies on the heuristic that a large directed Hausdorff distance between SAM2 proposal masks indicates distractor-induced ambiguity, but no evidence is provided that this specific criterion identifies useful memory frames. The paper also does not state the threshold or how it is normalized (e.g., relative to object size), which is important because the raw pixel distance will scale with object dimensions. Please provide an ablation that isolates the Hausdorff-distance selection rule and report the chosen threshold and normalization scheme.
minor comments (6)
  1. [General] Several table captions contain the typo "T able" instead of "Table" (Tables 1, 2, and 3).
  2. [Abstract] The term "trainless" is informal; consider using "training-free" for consistency with the rest of the paper.
  3. [§4.1, Eq. (3)] The parameters α and β are introduced but the text does not state their feasible range; please state that α, β ≥ 0 and α + β ≤ 1, and indicate where their values are set.
  4. [Figure 3] The right panel of Figure 3 is difficult to interpret because the color scale and circle-size legend for the AUC values are not explained in the caption; please make the axes and legend self-contained.
  5. [Table 2] SAM2Long is not included in the VOT-LT2020/VOT-LT2022 comparison; if its results are not available, please state this explicitly, or include them if they are reported in the literature.
  6. [§5.4] The sentence "The final selection of parameters is provided in our released code" is not sufficient for a self-contained paper; at least the numerical values of the main thresholds should appear in a table in the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the claimed gains are empirical measurements against independent benchmarks, and the hyperparameter-tuning concern is an evidential limitation, not a definitional reduction.

full rationale

The paper's contribution is an engineering modification of SAM2, not a derivation from first principles whose output is defined by its input. The score-fusion equations (1) and (3) linearly combine motion confidence, SAM2 IoU, and point-tracker Dice scores; the selected mask is the one maximizing the combined score, and the reported AUC/F-score is measured against ground truth, so the result is not equivalent to the input by construction. The memory filter (Eq. 4) uses directed Hausdorff distance to select frames for long-term memory, but its effect is tested empirically in the ablation (Table 4), not assumed by definition. The main legitimate concern is evidential: Section 5.4 says 'The final selection of parameters is provided in our released code,' and the sensitivity analysis is run on the same LaSOT dataset used for the headline 9.6% gain, so the LaSOT number is not fully out-of-sample. However, this is a reproducibility and overfitting risk, not circularity: no fitted parameter is renamed as a prediction, and the method also improves on LaSOText, VOT-LT2020, and VOT-LT2022, which were not used for the reported parameter sensitivity. Self-citations ([28], [36]) appear only in related-work context and are not load-bearing; no uniqueness theorem from the authors is invoked to forbid alternatives. The derivation chain is therefore self-contained as an empirical claim, with no step reducing to its own inputs.

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

The central claim is empirical and engineering-level: it rests on the pretrained SAM2.1 and CoTracker3 models, on a set of tuned hyperparameters, and on standard VOT benchmarks. There are no fitted constants derived from first principles; instead, the method's generalization depends on the heuristics above being robust outside the datasets used for tuning.

free parameters (6)
  • Alpha (alpha) in Eq. (1) and Eq. (3) = Not reported in paper; final value in code
    Weights the coarse Kalman motion score against SAM2's IoU score, and later against the fine point-tracker score. It directly controls proposal selection and is tuned on LaSOT.
  • Beta (beta) in Eq. (3)
    Weights the fine point-tracker similarity score in the confidence combination. Its value is not given in the paper.
  • Threshold tau for fine motion trigger
    When the confidence falls below this IoU threshold, the expensive point tracker is invoked. Sensitivity is shown in Fig. 3 middle, but the final value is only in code.
  • Point-tracker temporal window
    Number of historical frames used for backward propagation and mask reconstruction; sensitivity is in Fig. 3 left, final value in code.
  • Long-term and short-term memory frame intervals
    How often frames are considered for the short and long memory banks; Fig. 3 right evaluates combinations, final selection in code.
  • Hausdorff distance threshold for long-term memory inclusion
    Frames whose proposal separation exceeds a threshold are stored as distractor-aware long-term memory; the threshold is not quantified in the paper.
assumptions (6)
  • domain assumption SAM2.1 with its pre-trained weights provides the base tracker and memory attention as described in Section 3.
    The whole method operates as a wrapper around SAM2.1; if the base model or its released weights behave differently, the reported results change.
  • domain assumption CoTracker3 provides sufficiently reliable pixel correspondences for backward point propagation.
    Fine motion estimation reconstructs soft masks from tracked points using RBF (Eq. 2); if CoTracker3 loses the target, the fine score becomes noise.
  • domain assumption The Kalman filter's constant-velocity box model is a reasonable coarse motion prior.
    Coarse estimation treats adjacent-frame motion as smooth and linear; fast or non-linear motion is assumed to be caught by the confidence trigger.
  • domain assumption Benchmark AUC, P, Pnorm and F-score on LaSOT, LaSOText, VOT-LT2020, and VOT-LT2022 measure long-term tracking quality.
    All claims of improvement rest on these public benchmarks; they are standard in the field but are single-run numbers.
  • standard math Farthest point sampling plus Gaussian RBF reconstruction yields a usable soft mask from visible tracked points.
    Eq. (2) is a standard interpolation choice; its adequacy for mask reconstruction is assumed without validation.
  • ad hoc to paper Large directed Hausdorff distance between SAM2 proposals indicates distractor-induced ambiguity.
    This heuristic (Eq. 4) is introduced specifically for long-term memory selection and is not validated independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HiM2SAM: Enhancing SAM2 with Hierarchical Motion Estimation and Memory Optimization towards Long-term Tracking." pith.science (2026). https://pith.science/paper/TFSQG32W

@misc{pith2026250707603,
  author       = {Pith},
  title        = {Pith review of: HiM2SAM: Enhancing SAM2 with Hierarchical Motion Estimation and Memory Optimization towards Long-term Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TFSQG32W}},
  note         = {Machine review of arXiv:2507.07603}
}
read the original abstract

This paper presents enhancements to the SAM2 framework for video object tracking task, addressing challenges such as occlusions, background clutter, and target reappearance. We introduce a hierarchical motion estimation strategy, combining lightweight linear prediction with selective non-linear refinement to improve tracking accuracy without requiring additional training. In addition, we optimize the memory bank by distinguishing long-term and short-term memory frames, enabling more reliable tracking under long-term occlusions and appearance changes. Experimental results show consistent improvements across different model scales. Our method achieves state-of-the-art performance on LaSOT and LaSOText with the large model, achieving 9.6% and 7.2% relative improvements in AUC over the original SAM2, and demonstrates even larger relative gains on smaller models, highlighting the effectiveness of our trainless, low-overhead improvements for boosting long-term tracking performance. The code is available at https://github.com/LouisFinner/HiM2SAM.

Figures

Figures reproduced from arXiv: 2507.07603 by the authors.

Figure 1
Figure 1. HiM2SAM with hierarchical motion estimation and long-short memory. Left: Blue box: bounding box of detected target object; Orange box: bounding box of distrac￾tive object; Green box: ground truth; Red box: inaccurate tracking result; White box: coarse motion estimation; White mask: fine motion estimation. Right: Comparison with state-of-the-art trackers on LaSOT series datasets. Circle size indicates FPS; axes show … view at source ↗
Figure 2
Figure 2. The structure of HiM2SAM. The upper part illustrates the overall architecture, where the yellow blocks represent the original SAM2 components, while the red and purple blocks indicate the modules introduced by HiM2SAM. The bottom-left shows the hierarchical motion estimation module, while the bottom-right details the memory filter module. M(t) (x) = X N i=1 ϕ(∥x − xi∥), (2) where ϕ(·) is a Gaussian kernel, xi are th… view at source ↗
Figure 3
Figure 3. Sensitivity analysis of key parameters on the LaSOT dataset using HiM2sam￾large. Left: AUC vs. the number of frames tracked by the point tracker. Middle: AUC vs. the IoU threshold τ in fine motion estimation. Right: Performance under different combinations of long-term memory (LM) and short-term memory (SM) frame intervals. Circle size indicates the corresponding AUC score. est runtime overhead. Runtime is measured … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of tracking results on long video sequences. Yellow for SAMURAI, red masks represent DAM4SAM, and blue for HiM2SAM (ours). Green bounding boxes denote ground truth. HiM2SAM provides more stable and accurate tracking results under long-term challe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    Cambridge University Press (2003)

    Buhmann, M.: Radial Basis Functions: Theory and Implementations. Cambridge University Press (2003)

    work page 2003

  2. [2]

    Cai, W., Liu, Q., Wang, Y.: Hiptrack: Visual tracking with historical prompts (2024), https://arxiv.org/abs/2311.02072

    work page Pith review arXiv 2024

  3. [3]

    In: arXiv (2023)

    Cheng, H.K., Oh, S.W., Price, B., Lee, J.Y., Schwing, A.: Putting the object back into video object segmentation. In: arXiv (2023)

    work page 2023

  4. [4]

    In: European Conference on Computer Vision (ECCV) (2022)

    Cheng, H.K., Schwing, A.G.: XMem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision (ECCV) (2022)

    work page 2022

  5. [5]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Cuttano, C., Trivigno, G., Rosi, G., Masone, C., Averta, G.: Samwise: Infusing wis- dom in sam2 for text-driven video segmentation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 3395–3405 (2025) 14 R. Chen et al

    work page 2025

  6. [6]

    arXiv preprint arXiv:2410.16268 (2024)

    Ding, S., Qian, R., Dong, X., Zhang, P., Zang, Y., Cao, Y., Guo, Y., Lin, D., Wang, J.: Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree. arXiv preprint arXiv:2410.16268 (2024)

    arXiv 2024

  7. [7]

    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., Ling, H.: Lasot: A high-quality large-scale single object tracking benchmark (2020),https://arxiv.org/abs/2009.03465

    arXiv 2020

  8. [8]

    Fan, H., Lin, L., Yang, F., Chu, P., Deng, G., Yu, S., Bai, H., Xu, Y., Liao, C., Ling, H.: Lasot: A high-quality benchmark for large-scale single object tracking (2019), https://arxiv.org/abs/1809.07845

    work page Pith review arXiv 2019

Show all 37 references
  1. [9]

    IEEE Transactions on Pattern Analysis and Machine Intelli- gence 15(9), 850–863 (1993).https://doi.org/10.1109/34.232073

    Huttenlocher, D., Klanderman, G., Rucklidge, W.: Comparing images using the hausdorff distance. IEEE Transactions on Pattern Analysis and Machine Intelli- gence 15(9), 850–863 (1993).https://doi.org/10.1109/34.232073

    1993 doi

  2. [10]

    arXiv preprint arXiv:2504.04519 (2025)

    Jiang, J., Wang, Z., Zhao, M., Li, Y., Jiang, D.: Sam2mot: A novel paradigm of multi-object tracking by segmentation. arXiv preprint arXiv:2504.04519 (2025)

    2025

  3. [11]

    Trans- actions of the ASME–Journal of Basic Engineering82(Series D), 35–45 (1960)

    Kalman, R.E.: A new approach to linear filtering and prediction problems. Trans- actions of the ASME–Journal of Basic Engineering82(Series D), 35–45 (1960)

    1960

  4. [12]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2025)

    Kang, B., Chen, X., Lai, S., Liu, Y., Liu, Y., Wang, D.: Exploring enhanced con- textual information for video-level object tracking. In: Proceedings of the AAAI Conference on Artificial Intelligence (2025)

    2025

  5. [13]

    Karaev, N., Makarov, I., Wang, J., Neverova, N., Vedaldi, A., Rupprecht, C.: Co- Tracker3: Simpler and better point tracking by pseudo-labelling real videos (2024)

    2024

  6. [14]

    In: European Conference on Computer Vision (ECCV) (2024)

    Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. In: European Conference on Computer Vision (ECCV) (2024)

    2024

  7. [15]

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything (2023), https://arxiv.org/abs/2304.02643

    2023 arXiv

  8. [16]

    In: European Conference on Com- puter Vision (ECCV)

    Kristan, M., Leonardis, A., Matas, J., Felsberg, M., Pflugfelder, R., Kämäräinen, J.K., Chang, H.J., Danelljan, M., Čehovin Zajc, L., Lukežič, A., et al.: The tenth visual object tracking vot2022 challenge results. In: European Conference on Com- puter Vision (ECCV). Springer (2022)

    2022

  9. [17]

    In: European Conference on Com- puter Vision (ECCV Workshops)

    Kristan, M., Leonardis, A., Matas, J., Felsberg, M., Pflugfelder, R., Kämäräinen, J.K., Danelljan, M., Čehovin Zajc, L., Lukežič, A., Drbohlav, O., et al.: The eighth visual object tracking vot2020 challenge results. In: European Conference on Com- puter Vision (ECCV Workshops...

    2020

  10. [18]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Li, J., Cantu, F.J.P., Yu, E., Wong, A., Cui, Y., Chen, Y.: Samjam: Zero-shot video scene graph generation for egocentric kitchen videos. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 467–473 (2025)

    2025

  11. [19]

    In: European Conference on Computer Vision (ECCV) (2024)

    Lin, L., Fan, H., Zhang, Z., Wang, Y., Xu, Y., Ling, H.: Tracking meets lora: Faster training, larger model, stronger performance. In: European Conference on Computer Vision (ECCV) (2024)

    2024

  12. [20]

    Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks forvolumetricmedicalimagesegmentation(2016), https://arxiv.org/abs/1606. 04797

    2016

  13. [21]

    arXiv preprint arXiv:1706.02413 (2017)

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. arXiv preprint arXiv:1706.02413 (2017)

    2017 arXiv

  14. [22]

    arXiv preprint arXiv:2408.00714 (2024),https://arxiv.org/ abs/2408.00714 Enhancing SAM2 with Motion and Memory Optimization 15

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv...

    2024 arXiv

  15. [23]

    In: Pro- ceedings of the 40th International Conference on Machine Learning (ICML) (2023)

    Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: Pro- ceedings of the 40th International Co...

    2023

  16. [24]

    arXiv preprint arXiv:2410.01806 (2024)

    Segu, M., Piccinelli, L., Li, S., Yang, Y.H., Van Gool, L., Schiele, B.: Samba: Synchronized set-of-sequences modeling for end-to-end multiple object tracking. arXiv preprint arXiv:2410.01806 (2024)

    2024 arXiv

  17. [25]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)

    Shi, L., Zhong, B., Liang, Q., Li, N., Zhang, S., Li, X.: Explicit visual prompts for visual object tracking. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)

    2024

  18. [26]

    2024 IEEE International Conference on Image Processing (ICIP) pp

    Shim, K., Ko, K., Hwang, J., Kim, C.: Adaptrack: Adaptive thresholding-based matching for multi-object tracking. 2024 IEEE International Conference on Image Processing (ICIP) pp. 2222–2228 (2024)

    2024

  19. [27]

    Machine Vision and Applications 35(3) (2024)

    Stanojevic, V.D., Todorovic, B.T.: Boosttrack: boosting the similarity measure and detection confidence for improved multiple object tracking. Machine Vision and Applications 35(3) (2024)

    2024

  20. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Sun, G., Liu, Y., Ding, H., Probst, T., Van Gool, L.: Coarse-to-fine feature mining for video semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3126–3137 (2022)

    2022

  21. [29]

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention is all you need (2023),https://arxiv.org/abs/1706. 03762

    2023

  22. [30]

    In: Comp

    Videnovic, J., Lukezic, A., Kristan, M.: A distractor-aware memory for visual ob- ject tracking with SAM2. In: Comp. Vis. Patt. Recognition (2025)

    2025

  23. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Xie, J., Zhong, B., Mo, Z., Zhang, S., Shi, L., Song, S., Ji, R.: Autoregressive queries for adaptive tracking with spatio-temporal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19300–19309 (2024)

    2024

  24. [32]

    Yang, C.Y., Huang, H.W., Chai, W., Jiang, Z., Hwang, J.N.: Samurai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory (2024), https://arxiv.org/abs/2411.11922

    2024 arXiv

  25. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence38(7), 6702–6710 (Mar 2024).https: //doi.org/10.1609/aaai.v38i7.28493

    Yi, K., Luo, K., Luo, X., Huang, J., Wu, H., Hu, R., Hao, W.: Ucmctrack: Multi- object tracking with uniform camera motion compensation. Proceedings of the AAAI Conference on Artificial Intelligence38(7), 6702–6710 (Mar 2024).https: //doi.org/10.1609/aaai.v38i7.28493

    2024 doi

  26. [34]

    In: European Conference on Computer Vision (ECCV) (2022)

    Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision (ECCV) (2022)

    2022

  27. [35]

    In: Proceedings of the AAAI Confer- ence on Artificial Intelligence (2024)

    Zheng, Y., Zhong, B., Liang, Q., Mo, Z., Zhang, S., Li, X.: Odtrack: Online dense temporal token learning for visual tracking. In: Proceedings of the AAAI Confer- ence on Artificial Intelligence (2024)

    2024

  28. [36]

    Visual Intelligence3(1), 10 (2025)

    Zhou, Y., Sun, G., Li, Y., Xie, G.S., Benini, L., Konukoglu, E.: When sam2 meets video camouflaged object segmentation: A comprehensive evaluation and adapta- tion. Visual Intelligence3(1), 10 (2025)

    2025

  29. [37]

    IEEE Transactions on Circuits and Sys- tems for Video Technology (2024)

    Zhu, W., Cao, J., Xie, J., Yang, S., Pang, Y.: Clip-vis: Adapting clip for open- vocabulary video instance segmentation. IEEE Transactions on Circuits and Sys- tems for Video Technology (2024)

    2024

Pith tools

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