Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

SAMITE: Position Prompted SAM2 with Calibrated Memory for Visual Object Tracking

T0 review · 4 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 memory-scoring trick lifts zero-shot tracking to 74.9 AUC

desk verdict Solid zero-shot SAM2 tracking paper with real gains, but its 'position-wise' memory score is actually appearance-based — worth serious review, needs a fix and some reporting cleanup. read the letter →

arxiv 2507.21732 v1 pith:Y67HDBKL submitted 2025-07-29 cs.CV

Qianxiong Xu , Lanyun Zhu , Chenxi Liu , Guosheng Lin , Cheng Long , Ziyue Li , Rui Zhao This is my paper
classification cs.CV MSC 68T45
keywords visualobjecttrackingSAM2zero-shotprototypicalmemorybankselectionocclusiondistractorsuppressionpositionalprompt
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 usual way of adapting SAM2 to visual object tracking—feeding it the first frame's ground-truth box and the six most recent predictions as memory—lets occluded or distracted frames poison later predictions. SAMITE adds two training-free modules: a Prototypical Memory Bank that scores each past frame's usefulness by how its foreground prototype matches the first frame (feature accuracy) and the previous frame (positional proximity), and a Positional Prompt Generator that turns positional clues into a mask prompt to suppress lookalike distractors. On six benchmarks the authors report best or tied zero-shot results, including 74.9 AUC on LaSOT, a 2.0-point gain over the strongest SAM2-based baseline, and 60.7 AUC on LaSOText. If the claim holds, memory selection—not the segmentation model itself—is where tracking errors can be stopped.

What carries the argument

The load-bearing object is the Prototypical Memory Bank (PMB), which for each frame stores a memory plus a foreground prototype and a background prototype obtained by global average pooling of SAM2 features inside and outside the predicted mask. At frame $t$ it scores each candidate frame $\tau$ as $S_\tau = (1-\alpha)\,S^{\text{feat}}_\tau + \alpha\,S^{\text{pos}}_\tau$, where $S^{\text{feat}}_\tau$ is the cosine similarity of $\tau$'s foreground prototype to the first-frame anchor and $S^{\text{pos}}_\tau$ is the similarity to the previous-frame anchor, with $\alpha=0.3$ and a candidate window of $m=30$ recent frames; the top five candidates plus the two anchors form the seven memories given to SAM2's memory attention. The companion module, the Positional Prompt Generator (PPG), produces a discriminative prior mask from foreground-minus-background prototype similarities, multiplies it by a positional prior mask derived from the previous frame's predicted mask with 2D positional encodings, and feeds the result as a pseudo mask prompt—gated by a cycle-consistent mIoU check with threshold $\beta=0.7$. Together the two modules carry the argument: PMB intercepts error propagation at the memory level, and PPG supplies explicit positional evidence to reject distractors at the prompt level.

What would settle it

Run SAMITE on LaSOT's Fast Motion attribute (53 videos where the target moves more than its own box size between adjacent frames): if its AUC on that subset is statistically indistinguishable from simply using SAM2's default first-frame-plus-six-recent memories, then the position-wise anchor is not doing the claimed work and the method's core premise fails.

Watch

Extended reading notes

Core claim

SAMITE's central claim is that SAM2's fixed memory rule (first frame plus six most recent) is the real failure point in foundation-model tracking, and that error propagation can be intercepted by selecting which frames condition the current one. The paper treats each stored frame as having two independent kinds of correctness: feature-wise correctness, measured by cosine similarity between that frame's foreground prototype and the first frame's prototype, and position-wise correctness, measured by similarity to the previous frame's prototype. Frames that score low on either axis—occluded objects have incomplete features, distracted frames put the target in the wrong place—are dropped from the memory bank, while the two anchors are always kept. A second module builds a positional mask prompt by comparing current-frame features with stored foreground/background prototypes, reweighting the result with a position prior from the previous frame, and applies it only when a cycle-consistency check confirms it helps. The result, the paper claims, is a zero-shot tracker that surpasses trained VOT specialists on LaSOT and LaSOText without any VOT fine-tuning.

Load-bearing premise

Everything rests on the heuristic that cosine similarity between the current frame's foreground prototype and two fixed anchors—the first frame and the previous frame—measures both whether the target is correctly recognized and whether it is in the right place, which can fail under fast motion, abrupt appearance drift, or a corrupted previous frame.

Editorial extensions

If this is right

  • Frames whose targets are occluded lose feature similarity to the first-frame anchor and are filtered out before their incomplete features can condition later frames.
  • Frames where the target jumps to a distractor lose positional similarity to the previous-frame anchor, so that incorrect position is not propagated.
  • Because the selection is training-free and prototype-based, the gains transfer across SAM2 sizes T, S, B and L, with the largest relative gains on the smaller models.
  • On benchmarks with unseen object categories (LaSOText, GOT-10k), the zero-shot SAMITE surpasses supervised trackers trained on VOT data, e.g., 73.1 vs. 65.5 normalized precision on LaSOText.
  • The added cost is linear—$O(2m)$ for memory scoring and $16HW$ for prompt generation—so the method runs at 9.2 FPS on the B variant, comparable to the distractor-aware baseline SAM2.1++.

Reading between the lines

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

  • The same prototype-anchor trick could be lifted to other memory-conditioned video models: any autoregressive segmentation or tracking system could score stored states by consistency with a trusted anchor, without retraining.
  • Because the paper's own hyperparameter study shows that enlarging the candidate window hurts on fast-motion videos, an adaptive window sized by estimated motion could be a natural extension beyond the fixed $m=30$.
  • The cycle-consistent checking is a generic quality gate: it could be reused to decide when to trust any promptable model's self-generated prompts in other dense-prediction tasks, not just tracking.
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 / 5 minor

Summary. The paper proposes SAMITE, a zero-shot visual object tracker built on SAM2, which adds two modules: a Prototypical Memory Bank (PMB) and a Positional Prompt Generator (PPG). PMB scores candidate memory frames by cosine similarity between global-average-pooled foreground prototypes and two anchors (frame 1 and frame t-1), selects the top five candidates in addition to the two anchors, and uses these seven memories for memory attention. PPG constructs a discriminative prior mask from foreground/background prototypes, multiplies it by a positional prior derived from the previous frame's mask, averages the resulting prompts over the selected memories, and uses a cycle-consistent check to decide whether to apply the prompt. Experiments are reported on LaSOT, LaSOText, GOT-10k, TrackingNet, NFS, and OTB, with ablations, attribute-level results, model-size comparisons, efficiency numbers, and hyperparameter studies. The headline results include 74.9 AUC on LaSOT and 60.7 AUC on LaSOText, which the paper interprets as evidence that memory calibration and positional prompting intercept occlusion- and distraction-induced error propagation without VOT training.

Significance. If the reported results hold, SAMITE would be a valuable training-free enhancement of SAM2 for visual object tracking, and the memory-calibration idea is worth studying beyond this specific instantiation. The empirical package is extensive: six benchmarks, attribute-wise analysis on LaSOT, scaling across four SAM2 sizes, efficiency measurements, and parameter studies for all three introduced hyperparameters. The robustness of the gains across datasets and across nearby hyperparameter values is a genuine strength, and the paper is not circular in the sense of deriving the result from the result. However, the central conceptual contribution as written is not fully supported: the claimed 'position-wise' correctness score is computed from global-average-pooled appearance features, and the ablation uses an incompletely specified 'tailored' baseline. These issues affect the interpretation of the mechanism and the attribution of the gains, so the current version does not yet establish the paper's conceptual claims at the level of confidence the abstract suggests.

major comments (4)
  1. [Section 3.2.1, Eq. (4)-(5)] The score called 'position-wise' is the cosine similarity between global-average-pooled foreground prototypes P^FG_tau and P^FG_{t-1}. Global average pooling removes spatial layout, and the paper does not state that the feature map F_t carries positional encodings in this module; hence S^pos_tau measures appearance/temporal consistency, not whether the candidate target lies near the current target. A distractor whose appearance is similar to the previous target can therefore score high, which is exactly the failure mode in Figure 1(b). The t-1 anchor is also used unconditionally even when frame t-1 itself is wrong, with no validity check; Appendix C.4 acknowledges the fast-motion limitation of the candidate window but does not validate the anchor. The claim that PMB quantifies position-wise correctness and intercepts distraction errors is therefore not established by the computation as written. Please reformulate the score using explicit position information, or add a direct evaluation (for example, score distributions conditioned on known position errors or on distractor presence) that demonstrates the position-error-filtering behavior.
  2. [Section 4.3, Table 3] The 'tailored model' used as the no-PMB/no-PPG baseline is never defined, yet it already achieves 72.1 AUC on LaSOT, compared with 66.0 for the vanilla SAM2.1-B reported in Table 1 and 72.9 for SAM2.1++-B. Without knowing what the tailored model contains (for example, bounding-box prompting details, multi-mask handling, or other implementation choices), the ablation cannot attribute the subsequent gains to PMB and PPG. Please specify the exact configuration of the tailored model and, ideally, include the vanilla SAM2.1-B and SAM2.1++-B rows in the same ablation table so the incremental contributions are interpretable.
  3. [Table 2] The comparisons on TrackingNet, NFS, and OTB omit SAMURAI-B and SAM2.1++-B, both of which appear in Table 1. The claim in Section 4.2 that SAMITE achieves 'best or tied' results on these benchmarks is therefore not supported by the table as shown. Add the missing baselines so the cross-benchmark claim can be checked directly against the two most relevant SAM2-based competitors.
  4. [Appendix C.4] The hyperparameters alpha, beta, and m are all tuned on LaSOT, and the reported 74.9 AUC is the best value over the grids (alpha=0.3, m=30, beta=0.7). This is a selection-on-the-test-set issue for the headline LaSOT number. The performance is robust across the grid and the gains persist on other datasets, so the central conclusion remains plausible, but the paper should either validate on a held-out split or explicitly state that the LaSOT margin is partly a grid-selected result.
minor comments (5)
  1. [Section 3.2.1, Eq. (4)] The normalization operation Norm(.) applied to cosine similarities is not defined; specify the mapping (for example, min-max normalization over the candidate set), because it affects the relative contribution of the feature-wise and position-wise terms in Eq. (5).
  2. [Section 3.2.2, Eq. (10)-(11)] The positional prior mask uses the previous frame's mask prediction without a validity check; clarify whether the cycle-consistent check gates only the final prompt or also the positional prior input, since a wrong mask at frame t-1 could reinforce an incorrect positional bias.
  3. [Section 4.3 and Table 5] The attribute-wise results show that SAMITE is worse than SAM2.1++ on View Change (-2.3 AUC) and only slightly better on Fast Motion (+0.8 AUC); the text's characterization of 'particularly good' performance over Table 5 should be qualified to acknowledge these attributes.
  4. [Appendix A and Abstract] Appendix A states that the code 'will be made public after paper acceptance,' while the abstract and introduction state that code is available at a URL; make the availability status consistent and indicate what the URL currently contains.
  5. [Section 4.1] There are small typos, including 'm = 30as' (missing space) and 'Appendix C.4..' (doubled period), and the reference to SAM2 using 'multi-head branches [50]' cites SAMURAI rather than the original SAM2 architecture; please correct these editorial issues.

Circularity Check

1 steps flagged · score 4.0 of 10

LaSOT-tuned hyperparameters partly select the headline AUC; no construction-level circularity in the derivation

  1. fitted input called prediction [Section 4.1 Implementation Details; Appendix C.4 Tables 8-10; Table 1]
    "we set α = 0.3 in Eq. 5 and m = 30 as the size of reduced candidate set for PMB, and set β = 0.7 as the threshold to use positional mask prompt in PPG, the parameter studies are included in Appendix C.4. ... when α = 0.3, the best performance can be achieved, where the AUC score can be as high as 74.9%"

    The hyperparameters α, m, β are selected on the LaSOT test set by argmax: Tables 8-10 show the chosen values each give the best LaSOT AUC (74.9). That same 74.9 is then reported in Table 1 as SAMITE's headline LaSOT result and used to claim it 'can achieve the best performance in most situations.' The headline number is thus partly a product of selection on the same benchmark, making the best-on-LaSOT claim statistically forced rather than an unbiased evaluation of a fixed configuration. The forcing is only partial: the AUC surface is smooth (α=0.1-0.5 gives 74.2-74.9), and untuned gains appear on LaSOText and other datasets, so the central claim retains independent content.

full rationale

The core derivation is not circular: the memory selection and prompt generation are specified by explicit equations (Eqs. 1-13) and the reported gains are benchmark scores against external baselines, not values recovered from the inputs by construction. The one partial circularity is the hyperparameter selection: α, m, β are tuned on the LaSOT test set (Appendix C.4, Tables 8-10), and the resulting 74.9 AUC is presented in Table 1 as the method's headline LaSOT performance, so part of the claimed superiority on LaSOT is a selected maximum rather than an unbiased prediction. This is mitigated by the smoothness of the AUC surface and by untuned gains on LaSOText and other datasets. The PPG follows AENet [46], a self-citation (shared first author and co-authors), but AENet is an externally published, falsifiable method and is a component rather than the load-bearing justification, so it does not raise the score. A separate correctness risk should not be confused with circularity: the 'position-wise' score in Eq. 4 is a cosine similarity between GAP-pooled prototypes (Eq. 2), and global average pooling discards spatial layout, so the claim that it quantifies position-wise accuracy is not established by the computation; this is a mechanistic/falsifiability concern, as echoed by the paper's own Appendix C.4 fast-motion caveat and Appendix E limitations. Verdict: no construction-level circularity; score 4 for the partial test-set-tuning circularity of the headline LaSOT claim.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method is an inference-time modification of SAM2. Its three hyperparameters are tuned on LaSOT, and several score-based heuristics (prototype cosine similarity, adjacent-frame anchor, positional re-weighting) are introduced ad hoc to make memory selection and prompting work. These are not derived from first principles.

free parameters (3)
  • α = 0.3
    Balances feature-wise and position-wise scores in memory calibration (Eq. 5). Tuned on LaSOT (Appendix C.4) where it gives best AUC.
  • m = 30
    Number of recent frames considered as candidates in the reduced candidate set. Tuned on LaSOT; best AUC at 30.
  • β = 0.7
    Threshold for cycle-consistent checking to decide whether to use the positional mask prompt. Tuned on LaSOT; best AUC at 0.7.
assumptions (5)
  • domain assumption The first frame's mask, prompted by ground-truth bounding box, is an accurate feature-wise anchor (Section 3.2.1)
    The method relies on first-frame GT being free of tracking error; in zero-shot evaluation it is provided.
  • domain assumption The target object in adjacent frames is spatially close, so frame t-1 is a reliable position-wise anchor (Section 3.2.1, Eq. 4-5)
    Assumes smooth motion; can fail for fast motion or abrupt cuts, as acknowledged in their own candidate set discussion.
  • ad hoc to paper Cosine similarity between foreground prototypes measures feature-wise and position-wise correctness of a tracked frame (Eq. 4)
    This is the core heuristic; no theoretical justification or learned calibration is given.
  • ad hoc to paper Global average pooling over predicted mask yields prototypes representing the target (Eq. 2)
    Design choice; assumes the predicted mask is accurate enough for pooling to be informative.
  • domain assumption Pretrained SAM2 features transfer to VOT without training (Section 3.1)
    The method is zero-shot and inherits SAM2's learned representations as-is.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAMITE: Position Prompted SAM2 with Calibrated Memory for Visual Object Tracking." pith.science (2026). https://pith.science/paper/Y67HDBKL

@misc{pith2026250721732,
  author       = {Pith},
  title        = {Pith review of: SAMITE: Position Prompted SAM2 with Calibrated Memory for Visual Object Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y67HDBKL}},
  note         = {Machine review of arXiv:2507.21732}
}
read the original abstract

Visual Object Tracking (VOT) is widely used in applications like autonomous driving to continuously track targets in videos. Existing methods can be roughly categorized into template matching and autoregressive methods, where the former usually neglects the temporal dependencies across frames and the latter tends to get biased towards the object categories during training, showing weak generalizability to unseen classes. To address these issues, some methods propose to adapt the video foundation model SAM2 for VOT, where the tracking results of each frame would be encoded as memory for conditioning the rest of frames in an autoregressive manner. Nevertheless, existing methods fail to overcome the challenges of object occlusions and distractions, and do not have any measures to intercept the propagation of tracking errors. To tackle them, we present a SAMITE model, built upon SAM2 with additional modules, including: (1) Prototypical Memory Bank: We propose to quantify the feature-wise and position-wise correctness of each frame's tracking results, and select the best frames to condition subsequent frames. As the features of occluded and distracting objects are feature-wise and position-wise inaccurate, their scores would naturally be lower and thus can be filtered to intercept error propagation; (2) Positional Prompt Generator: To further reduce the impacts of distractors, we propose to generate positional mask prompts to provide explicit positional clues for the target, leading to more accurate tracking. Extensive experiments have been conducted on six benchmarks, showing the superiority of SAMITE. The code is available at https://github.com/Sam1224/SAMITE.

Figures

Figures reproduced from arXiv: 2507.21732 by the authors.

Figure 1
Figure 1. Two failure cases of occlusion and distraction. (a) The occluded frames shall not be selected to condition subsequent frames, as the tracking target may be wrongly reduced from the last train carriage to its louver; (b) The spatial positions of target objects in adjacent frames should be close. To address this, a few recent advances [38, 50] propose to leverage the robust video foundation model SAM2 [35] for VOT. SA… view at source ↗
Figure 2
Figure 2. Overview of SAMITE, which is built upon SAM2, including: (1) Prototypical Memory Bank (PMB) is responsible for selecting calibrated memories with accurate and nearby target objects; (2) Positional Prompt Generator (PPG) generates pseudo mask prompt that can activates the target object, while suppressing other distracting objects, acting as positional clues. 3.2 SAMITE As shown in [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 3
Figure 3. Details about Positional Prompt Generator (PPG). Take bird-15 of LaSOT as an example, where frame #501 is the current frame, and frame #471 is one of the selected frames (i.e., memories). As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons with SAM2.1++ [38]. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Attribute-wise Performance. 𝑀𝑀�𝜏𝜏 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 (Eq. 9) 𝑀𝑀�𝜏𝜏 (Eq. 12) Predictions and GT w/o Prior w/ Prior GT Activate Suppress #217 #264 #252 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Illustration of using SAM2 for Visual Object Tracking. The first frame [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Visual impacts of Memory Calibration. The example is taken from basketball-11 of LaSOT, [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: More visualizations of positional mask prompt. [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: More qualitative results. More qualitative comparisons between the designed SAMITE and SAM2.1++ [38] are displayed in [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: More qualitative results. E Limitation Although the proposed zero-shot SAMITE model have already achieved appealing performance on multiple VOT benchmarks, there exist 2 potential limitations or future directions, including: (1) Efficiency: The designed Prototypical M…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Models as Tools: An Agentic Coordination Framework for Unified Multimodal Visual Tracking

    cs.CV 2026-08 conditional novelty 6.0 of 10

    ACTrack coordinates a tracker, a segmentation model, and a VLM through event-triggered conflicts, reporting state-of-the-art RGB and multimodal tracking with 30% trainable parameters.

  2. SENTRY: SAM2-Enhanced Neighbor-Aware and Temporally Reasoned Memory for Visual Tracking

    cs.CV 2026-06 conditional novelty 6.0 of 10

    SENTRY is a plug-and-play module that replaces confidence-based memory writes with neighbor-aware cycle-consistent validation in SAM2 trackers, yielding new zero-shot SOTA results on LaSOT, GOT-10k and other benchmarks.

  3. Segment Anything with Motion, Geometry, and Semantic Adaptation for Complex Nonlinear Visual Object Tracking

    cs.CV 2026-05 unverdicted novelty 5.0 of 10

    SAMOSA adapts SAM 2 for complex visual object tracking by integrating explicit nonlinear motion prediction, semantic cues for failure recovery, and geometric constraints for stability, outperforming prior SAM 2-based ...

Reference graph

Works this paper leans on

56 extracted references · 49 canonical work pages · cited by 3 Pith papers

  1. [1]

    Artrackv2: Prompting autoregressive tracker where to look and how to describe

    Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19048–19057, 2024

    work page 2024

  2. [2]

    Fully- convolutional siamese networks for object tracking

    Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully- convolutional siamese networks for object tracking. InComputer vision–ECCV 2016 workshops: Amsterdam, the Netherlands, October 8-10 and 15-16, 2016, proceedings, part II 14 , pages 850–865. Springer, 2016

    work page 2016

  3. [3]

    Learning discriminative model prediction for tracking

    Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6182–6191, 2019

    work page 2019

  4. [4]

    One-shot video object segmentation

    Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixé, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 221–230, 2017

    work page 2017

  5. [5]

    Hiptrack: Visual tracking with historical prompts

    Wenrui Cai, Qingjie Liu, and Yunhong Wang. Hiptrack: Visual tracking with historical prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 19258–19267, 2024

    work page 2024

  6. [6]

    Robust object modeling for visual tracking

    Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9589–9600, 2023

    work page 2023

  7. [7]

    Seqtrack: Sequence to sequence learning for visual object tracking

    Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14572–14581, 2023

    work page 2023

  8. [8]

    Transformer tracking

    Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 8126–8135, 2021

    work page 2021

Show all 56 references
  1. [9]

    Siamese box adaptive network for visual tracking

    Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, and Rongrong Ji. Siamese box adaptive network for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668–6677, 2020

    2020

  2. [10]

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

    Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13608–13618, 2022

    2022

  3. [11]

    Atom: Accurate tracking by overlap maximization

    Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4660–4669, 2019

    2019

  4. [12]

    Probabilistic regression for visual tracking

    Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7183–7192, 2020

    2020

  5. [13]

    Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree

    Shuangrui Ding, Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Yuwei Guo, Dahua Lin, and Jiaqi Wang. Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree. arXiv preprint arXiv:2410.16268, 2024

    arXiv

  6. [14]

    Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset

    Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF Internat...

    2021

  7. [15]

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

    Heng Fan, Hexin Bai, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Harshit, Mingzhen Huang, Juehuan Liu, et al. Lasot: A high-quality large-scale single object tracking benchmark. International Journal of Computer Vision, 129:439–461, 2021. 10

    2021

  8. [16]

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

    Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages ...

    2019

  9. [17]

    People tracking in surveillance applications

    Luis M Fuentes and Sergio A Velastin. People tracking in surveillance applications. Image and Vision Computing, 24(11):1165–1171, 2006

    2006

  10. [18]

    Aiatrack: Attention in attention for transformer visual tracking

    Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In European conference on computer vision, pages 146–164. Springer, 2022

    2022

  11. [19]

    Generalized relation modeling for trans- former tracking

    Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for trans- former tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18686–18695, 2023

    2023

  12. [20]

    Graph attention tracking

    Dongyan Guo, Yanyan Shao, Ying Cui, Zhenhua Wang, Liyan Zhang, and Chunhua Shen. Graph attention tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9543–9552, 2021

    2021

  13. [21]

    Siamcar: Siamese fully convolutional classification and regression for visual tracking

    Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. Siamcar: Siamese fully convolutional classification and regression for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6269–6277, 2020

    2020

  14. [22]

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

    Lingyi Hong, Shilin Yan, Renrui Zhang, Wanyun Li, Xinyu Zhou, Pinxue Guo, Kaixun Jiang, Yiting Chen, Jinglun Li, Zhaoyu Chen, et al. Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In Proceedings of the IEEE/CVF conference on computer v...

    2024

  15. [23]

    Automatic traffic surveil- lance system for vehicle tracking and classification

    Jun-Wei Hsieh, Shih-Hao Yu, Yung-Sheng Chen, and Wen-Fong Hu. Automatic traffic surveil- lance system for vehicle tracking and classification. IEEE Transactions on intelligent trans- portation systems, 7(2):175–187, 2006

    2006

  16. [24]

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

    Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence, 43(5):1562–1577, 2019

    2019

  17. [25]

    A new approach to linear filtering and prediction problems

    Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. 1960

    1960

  18. [26]

    Need for speed: A benchmark for higher frame rate object tracking

    Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In Proceedings of the IEEE international conference on computer vision, pages 1125–1134, 2017

    2017

  19. [27]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  20. [28]

    A review of tracking and trajectory prediction methods for autonomous driving

    Florin Leon and Marius Gavrilescu. A review of tracking and trajectory prediction methods for autonomous driving. Mathematics, 9(6):660, 2021

    2021

  21. [29]

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

    Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4282–4291, 2019

    2019

  22. [30]

    High performance visual tracking with siamese region proposal network

    Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8971–8980, 2018

    2018

  23. [31]

    Tracking meets lora: Faster training, larger model, stronger performance

    Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. In European Conference on Computer Vision, pages 300–318. Springer, 2024. 11

    2024

  24. [32]

    Swintrack: A simple and strong baseline for transformer tracking

    Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. Advances in Neural Information Processing Systems, 35:16743–16754, 2022

    2022

  25. [33]

    Learning target candidate association to keep track of what not to track

    Christoph Mayer, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Learning target candidate association to keep track of what not to track. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13444–13454, 2021

    2021

  26. [34]

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

    Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Track- ingnet: A large-scale dataset and benchmark for object tracking in the wild. In Proceedings of the European conference on computer vision (ECCV), pages 300–317, 2018

    2018

  27. [35]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  28. [36]

    Explicit visual prompts for visual object tracking

    Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Shengping Zhang, and Xianxian Li. Explicit visual prompts for visual object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4838–4846, 2024

    2024

  29. [37]

    Prior guided feature enrichment network for few-shot segmentation

    Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050–1065, 2020

    2020

  30. [38]

    A distractor-aware memory for visual object tracking with sam2

    Jovana Videnovic, Alan Lukezic, and Matej Kristan. A distractor-aware memory for visual object tracking with sam2. arXiv preprint arXiv:2411.17576, 2024

    2024 arXiv

  31. [39]

    Siam r-cnn: Visual tracking by re-detection

    Paul V oigtlaender, Jonathon Luiten, Philip HS Torr, and Bastian Leibe. Siam r-cnn: Visual tracking by re-detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6578–6588, 2020

    2020

  32. [40]

    Translating math formula images to latex sequences using deep neural networks with sequence-level training, 2019

    Zelun Wang and Jyh-Charn Liu. Translating math formula images to latex sequences using deep neural networks with sequence-level training, 2019

    2019

  33. [41]

    Autoregressive vi- sual tracking

    Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive vi- sual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9697–9706, 2023

    2023

  34. [42]

    Online object tracking: A benchmark

    Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Online object tracking: A benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2411– 2418, 2013

    2013

  35. [43]

    Object tracking benchmark

    Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Object tracking benchmark. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(9):1834–1848, 2015

    2015

  36. [44]

    Correlation-aware deep tracking

    Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8751–8760, 2022

    2022

  37. [45]

    Autoregressive queries for adaptive tracking with spatio-temporal transformers

    Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19300–...

    2024

  38. [46]

    Eliminating feature ambiguity for few-shot segmentation

    Qianxiong Xu, Guosheng Lin, Chen Change Loy, Cheng Long, Ziyue Li, and Rui Zhao. Eliminating feature ambiguity for few-shot segmentation. InEuropean Conference on Computer Vision, pages 416–433. Springer, 2024

    2024

  39. [47]

    Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines

    Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 12549–12556, 2020

    2020

  40. [48]

    Learning spatio-temporal transformer for visual tracking

    Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10448–10457, 2021. 12

    2021

  41. [49]

    Alpha-refine: Boosting tracking performance by precise bounding box estimation

    Bin Yan, Xinyu Zhang, Dong Wang, Huchuan Lu, and Xiaoyun Yang. Alpha-refine: Boosting tracking performance by precise bounding box estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5289–5298, 2021

    2021

  42. [50]

    Samurai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory

    Cheng-Yen Yang, Hsiang-Wei Huang, Wenhao Chai, Zhongyu Jiang, and Jenq-Neng Hwang. Samurai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory. arXiv preprint arXiv:2411.11922, 2024

    2024 arXiv

  43. [51]

    Mosam: Motion-guided segment anything model with spatial-temporal memory selection

    Qiushi Yang, Yuan Yao, Miaomiao Cui, and Liefeng Bo. Mosam: Motion-guided segment anything model with spatial-temporal memory selection. arXiv preprint arXiv:2505.00739, 2025

    2025 arXiv

  44. [52]

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

    Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European conference on computer vision, pages 341–357. Springer, 2022

    2022

  45. [53]

    Learn to match: Automatic matching network design for visual tracking

    Zhipeng Zhang, Yihao Liu, Xiao Wang, Bing Li, and Weiming Hu. Learn to match: Automatic matching network design for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13339–13348, 2021

    2021

  46. [54]

    Ocean: Object-aware anchor-free tracking

    Zhipeng Zhang, Houwen Peng, Jianlong Fu, Bing Li, and Weiming Hu. Ocean: Object-aware anchor-free tracking. In European conference on computer vision, pages 771–787. Springer, 2020

    2020

  47. [55]

    Odtrack: Online dense temporal token learning for visual tracking

    Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 7588–7596, 2024

    2024

  48. [56]

    Detrack: In-model latent denoising learning for visual object tracking

    Xinyu Zhou, Jinglun Li, Lingyi Hong, Kaixun Jiang, Pinxue Guo, Weifeng Ge, and Wenqiang Zhang. Detrack: In-model latent denoising learning for visual object tracking. Advances in Neural Information Processing Systems, 37:90579–90599, 2024. 13 A Code The source code is provided...

    2024

Pith tools

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