Pith. sign in

REVIEW 5 major objections 8 minor 25 references

CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features

T0 review · 5 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CSTrack claims that RGB and auxiliary-modality streams can be fused into one compact feature space, achieving state-of-the-art tracking on depth, thermal, and event benchmarks.

desk verdict Genuinely new compact single-branch fusion design and solid ablations, but per-benchmark tuning and missing code leave the SOTA claims needing independent confirmation. read the letter →

arxiv 2505.19434 v1 pith:OLJE3334 submitted 2025-05-26 cs.CV cs.AI

X. Feng , D. Zhang , S. Hu , X. Li , M. Wu , J. Zhang , X. Chen , K. Huang This is my paper · ORCID
classification cs.CVcs.AI
keywords RGB-Xtrackingmultimodalcompactspatiotemporalfeaturesobjecttransformertrackerthermaldeptheventcamera
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

Object tracking normally benefits from pairing RGB with a second visual modality—depth, thermal, or event frames—but existing trackers process the two streams in parallel branches, doubling the feature space and consuming compute on within-modality modeling. CSTrack claims that this separation is unnecessary: because RGB and X images are spatially aligned, the two streams can be fused early into one compact feature space, and the compute saved can be redirected to temporal modeling. The paper reports that this single-branch design outperforms every compared prior tracker on DepthTrack, VOT-RGBD2022, LasHeR, RGBT234, and VisEvent, including settings where one modality is missing. The authors scope the claim to visual auxiliary modalities and explicitly list visual-language tracking as future work.

What carries the argument

The load-bearing mechanism is the Spatial Compact Module: it concatenates learnable modality queries with each modality's tokens, applies bidirectional cross-attention between RGB and X, and then adds the two refined feature streams element-wise, exploiting the premise that the modalities are spatially aligned. Because the resulting token sequence is no longer than a single modality's, a one-stream Transformer backbone can handle both intra- and inter-modality modeling at once. The Temporal Compact Module complements this by forming a parameter-free heatmap from a correlation with template features and a Gaussian centered on the predicted box, selecting the most target-like search tokens as a compact temporal memory.

What would settle it

Take a paired RGB-D or RGB-T sequence and artificially misalign the X stream by even a few pixels before feeding it to CSTrack: if the compact-addition fusion is what makes the method work, accuracy should drop sharply on slight misalignment while a two-branch tracker that keeps the streams separate stays robust, and if it does not drop, the alignment premise is not load-bearing.

Watch

Extended reading notes

Core claim

On its own terms, CSTrack's discovery is that a tracker need not maintain two separated modality feature spaces to exploit RGB-X complementarity; it can compress the RGB and X token sequences into a single compact spatial feature via bidirectional cross-attention, with four learnable queries per modality carrying modality-specific residue, then run one shared Transformer backbone over the merged tokens. A Temporal Compact Module builds a target-distribution heatmap from both intermediate and final predictions, selects the top tokens of the search feature as a compact per-frame memory, and aggregates the last four frames to guide the next prediction. The paper argues this yields more thorough inter-modality interaction as well as dense temporal modeling that sparse template updates alone cannot provide, and reports SOTA numbers across RGB-D, RGB-T, and RGB-E benchmarks.

Load-bearing premise

The load-bearing premise is that RGB and X images are aligned enough that element-wise addition of their processed features is a safe fusion, with just four learned query tokens per modality carrying what that addition would lose; if misalignment such as depth occlusion boundaries or thermal viewpoint differences is severe, the compacting step would discard the very cues fusion is meant to exploit.

Editorial extensions

If this is right

  • If the reported results hold, a single-branch multimodal tracker can match or beat two-branch architectures while using fewer parameters (73M versus 82-136M) and running at 35 FPS rather than 18-24 FPS.
  • Element-wise addition after cross-attention implies that RGB-X trackers can be built by modifying a standard one-stream RGB tracker, since the compact feature has the same length as one modality's tokens.
  • The heatmap-based temporal memory offers a parameter-free way to add dense temporal cues to existing trackers; the paper's ablations show that combining the intermediate and final heatmaps matters more than either alone.
  • Joint training across RGB, depth, thermal, and event data is part of the recipe: ablations show that removing RGB tracking datasets or training per-modality separately hurts performance, so the model's cross-modal generalization is tied to the joint training.
  • On modality-missing benchmarks, simply cloning the available image into the missing stream already beats dedicated missing-modality methods, indicating that the compact fusion is robust to absent input.

Reading between the lines

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

  • A natural next test is whether the four learnable queries per modality scale with modality count or content; text-based modalities do not form a spatially aligned image stream, so the compact-addition step would likely need replacement for visual-language tracking.
  • The alignment premise predicts a sharp failure mode: on sequences where the X stream is misregistered or has occlusion boundaries that do not line up with RGB edges, CSTrack should lose its advantage, and a deliberately misaligned test set would separate the value of compact fusion from the value of the backbone and training data.
  • Because the temporal memory stores tokens selected by the current predicted box, a wrong or drifting box could poison future selections; this suggests an error-feedback loop that could be measured by comparing CSTrack's failure slope after a short occlusion with that of trackers that update templates only sparsely.
  • The performance gain pattern across benchmarks might be explained by how well each X modality aligns with RGB, so testing on modalities with known misalignment would probe whether the fusion rule is tuned to alignment quality.
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 / 8 minor

Summary. The paper proposes CSTrack, a unified RGB-X tracker for depth, thermal, and event modalities. The method consists of a Spatial Compact Module (SCM) that merges the RGB and X token streams through bidirectional cross-attention and element-wise addition into a single compact branch, and a Temporal Compact Module (TCM) that builds a parameter-free target-distribution heatmap from intermediate and final predictions, selects the top-Nm search tokens as a compact temporal feature, and stores the last L such features in a memory buffer. The model is trained jointly on RGB and RGB-X datasets with a HiViT backbone and evaluated on DepthTrack, VOT-RGBD2022, LasHeR, RGBT234, and VisEvent. The paper reports new state-of-the-art results on all five benchmarks, with large margins such as 75.6 vs 69.4 PR on LasHeR, 66.4 vs 60.6 recall on DepthTrack, and 82.4 vs 76.7 PR on VisEvent, supported by ablations of the spatial, temporal, and training choices.

Significance. If the reported results hold under a single fixed protocol, this is a strong empirical contribution: the compact single-branch design reduces parameter count and FLOPs compared to symmetric dual-branch alternatives (73M/36G vs 136M/59G in Table 5) while improving accuracy, and the parameter-free temporal heatmap selection is a clean alternative to learned temporal queries. The paper is thorough in ablating the spatial module, temporal module, training strategy, and memory length, and the trends are internally consistent across modalities. However, the central SOTA claim is not fully verifiable from the manuscript alone: several evaluation choices are benchmark-specific with undisclosed selection procedures, and the joint-training ablation (Table 8) shows that part of the gain may come from additional RGB training data rather than from the proposed architecture. The authors should be credited for making the architecture description and ablation details sufficiently explicit to allow the protocol concerns to be tested.

major comments (5)
  1. [Sec. B / Tables 1-3] The headline SOTA claim is compromised by benchmark-specific evaluation settings whose selection procedure is not disclosed. Appendix B states that the dynamic-template update threshold is 0.45 for LasHeR and RGBT234 and 0.7 for all other datasets, and Table 9 shows that input normalization statistics are computed per dataset. No validation split or criterion for choosing these values is given, and the code and models are not yet available. Because these choices interact with test-set statistics, the reported margins (e.g., +6.2 PR on LasHeR and +5.7 PR on VisEvent) cannot currently be attributed to the architecture under a single fixed protocol. Please report results with one fixed protocol, or justify the thresholds from a validation set and provide sensitivity analyses.
  2. [Sec. 4.3.2 / Fig. 5] The temporal memory length L=4 is selected from Fig. 5, which reports performance on the test benchmarks (RGBT234, DepthTrack, VisEvent), not on a held-out validation set. Similarly, Nq=4, Nm=16, the 0.5 fusion weight in Eq. (12), and the w/3 and h/3 Gaussian scales in Eq. (11) are fixed constants without sensitivity analyses. This constitutes test-set model selection, which can inflate the reported improvements. Provide validation-based selection or a sensitivity table over these hyperparameters, and state explicitly whether any hyperparameter was chosen by looking at test results.
  3. [Sec. 4.3.3 / Table 8] Table 8 shows that removing the large RGB training datasets (LaSOT, GOT-10K, COCO, TrackingNet, VastTrack, TNL2k) from joint training lowers RGBT234 MSR/MPR by 7.7/7.1 points and DepthTrack recall/precision by about 4 points, while removing joint training entirely costs 2-7 points depending on the benchmark. These magnitudes are comparable to the reported SOTA margins over prior trackers, so the headline gains may reflect additional training data rather than the compact spatiotemporal design. The comparison to prior trackers should be controlled by training competing methods on the same data, or by reporting a version of CSTrack trained with the same protocol as the baselines.
  4. [Sec. 3.1.2 / Eq. (6)] Eq. (6) adds f''r and f''x element-wise after cross-attention, which is valid only if the two modality features are spatially aligned at token level. The paper asserts this alignment by citing Yang et al. (2022) and Wu et al. (2024), but no experiment tests robustness to realistic misalignments such as depth occlusion boundaries, thermal perspective differences, or event temporal offsets. The ablation without modality queries (Table 6 #2) shows only a 1-2 point drop, so the safety margin for misalignment is unknown. Please add a diagnostic or analysis (e.g., controlled translation or occlusion of the X input) to show that Eq. (6) does not silently discard modality-specific cues.
  5. [Tables 1-4] No error bars or repeated-seed statistics are reported for any benchmark table. Given the protocol concerns above and the fact that some margins are moderate relative to typical run-to-run variance in deep trackers, the reader cannot assess whether the improvements are statistically stable. Report mean and standard deviation over at least three runs for the main comparisons, or provide significance information.
minor comments (8)
  1. [Sec. 4.1] The paragraph contains a duplicated and grammatically broken sentence: 'The CSTrack is initialized with Fast-iTPN (Tian et al., 2024) pre-trained weights, and the token dimension D is set to 512. which are initialized with the Fast-iTPN pre-trained parameters and the token dimension D set to 512.' Please fix.
  2. [Table 5] Table 5 #3 reports 73M parameters and 35 FPS, while Sec. 4.1 reports 75M parameters and 33 FPS for the full model; state explicitly that Table 5's one-compact-branch row excludes TCM, or reconcile the numbers.
  3. [Eqs. (9) and (13)] The notation s'_c is used for two different quantities: the self-correlation output in Eq. (9) and the cross-attention output in Eq. (13). Please rename one of them.
  4. [Eq. (12)] The Norm operation in Eq. (12) is not specified; clarify whether it is L2 normalization, layer normalization, or min-max normalization over the spatial dimension.
  5. [Fig. 5] Figure 5 does not identify which dataset or metric the 'Delta Precision' values refer to; add this information and, if possible, error bars.
  6. [Appendix C.2] In the first sentence of the 'w/o queries' paragraph, 'removes the modality-specific queries, qr and qr' should read 'qr and qx'.
  7. [Sec. 3] The methods overview says 'shared Path Embedding' but should be 'Patch Embedding'.
  8. [Tables 1-3] The red/blue top-two highlighting is not explicitly explained in the caption text and may be ambiguous in black-and-white printing; consider adding bold or symbols.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SOTA claim is an external benchmark ranking, and the architecture is derived from stated design choices rather than from the evaluated metrics.

full rationale

The paper's central claim is an empirical ranking on independent public benchmarks (DepthTrack, VOT-RGBD2022, LasHeR, RGBT234, VisEvent), so the reported numbers are measurements against fixed external labels rather than predictions that reduce to fitted values. The architecture components (Spatial Compact Module, Eqs. 2-7; Temporal Compact Module, Eqs. 9-14) are stated design operations on the input streams and intermediate features, and the ablation studies in Tables 5-8 compare controlled variants without defining the headline metrics. The per-dataset dynamic-template thresholds (0.45 for LasHeR/RGBT234 and 0.7 otherwise, Sec. B) and input-normalization statistics (Tab. 9, computed on training videos) are hyperparameters and preprocessing choices; they may raise reproducibility concerns if selected using test-set knowledge, but they are not circular because they do not define or derive the benchmark scores. Self-citations appear only as related work, motivation, or prior modules (e.g., Feng et al. 2024, MemVLT; Zhang et al. 2024a) and are not load-bearing for the SOTA claim. The temporal feature selection uses the current frame's predicted box, but the box is produced from temporal memory of past steps, not from the current temporal feature being stored, so there is no self-referential feedback that makes the prediction equivalent to its input by construction. Overall, no step in the derivation chain reduces to its own inputs.

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

The paper's empirical claim rests mainly on two domain assumptions: that RGB and X images are spatially aligned so they can be merged by addition, and that the heatmap built from the model's own prediction is a reliable guide for selecting temporal tokens. The free parameters listed are all hand-set or test-set-tuned. No new physical or ontological entities are introduced; the learnable query tokens are standard transformer parameters.

free parameters (6)
  • Nq (modality-specific query count) = 4
    Number of learnable query tokens per modality retained in the compact feature; set by hand in Sec. 4.1.
  • Nm (temporal feature token count) = 16
    Number of target tokens selected per time step for the temporal memory; set by hand in Sec. 4.1.
  • L (temporal memory length) = 4
    Selected from an ablation on benchmark test sets (Fig. 5), so it is tuned against evaluation data.
  • Heatmap fusion weight = 0.5 / 0.5
    Weighted sum of the intermediate and final heatmaps in Eq. 12; chosen by hand, no sensitivity study.
  • Gaussian scale divisor = 3
    In Eq. 11 the predicted box width and height are divided by 3 to set the heatmap standard deviation; chosen by hand.
  • Dynamic template update thresholds = 0.45 (RGB-T), 0.7 (others)
    Per-benchmark thresholds adopted from TATrack and tuned separately for RGB-T versus other modalities (Appendix B).
assumptions (4)
  • domain assumption RGB and X images are well aligned in spatial semantic content, so element-wise addition after cross-attention is a safe fusion.
    Invoked in Sec. 3.1.2 before Eq. 6, citing Yang et al. 2022 and Wu et al. 2024; the paper offers no quantitative alignment analysis.
  • domain assumption A one-stream transformer backbone can perform both intra- and inter-modality modeling on the compact fused feature.
    Invoked in Sec. 3.1.3; relies on the general token-mixing power of transformers without a dedicated proof.
  • domain assumption The target distribution heatmap built from similarity and the predicted box reflects the true target position well enough to select useful temporal tokens.
    Invoked in Sec. 3.2.1; if the tracker drifts, the heatmap and stored tokens could compound the error.
  • domain assumption Attenuation Theory motivates coarse-to-fine compression of multiple information sources.
    Cited from Treisman 1964 in Sec. 1 as biological motivation; it is not mathematically load-bearing for the architecture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features." pith.science (2026). https://pith.science/paper/OLJE3334

@misc{pith2026250519434,
  author       = {Pith},
  title        = {Pith review of: CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OLJE3334}},
  note         = {Machine review of arXiv:2505.19434}
}
read the original abstract

Effectively modeling and utilizing spatiotemporal features from RGB and other modalities (\eg, depth, thermal, and event data, denoted as X) is the core of RGB-X tracker design. Existing methods often employ two parallel branches to separately process the RGB and X input streams, requiring the model to simultaneously handle two dispersed feature spaces, which complicates both the model structure and computation process. More critically, intra-modality spatial modeling within each dispersed space incurs substantial computational overhead, limiting resources for inter-modality spatial modeling and temporal modeling. To address this, we propose a novel tracker, CSTrack, which focuses on modeling Compact Spatiotemporal features to achieve simple yet effective tracking. Specifically, we first introduce an innovative Spatial Compact Module that integrates the RGB-X dual input streams into a compact spatial feature, enabling thorough intra- and inter-modality spatial modeling. Additionally, we design an efficient Temporal Compact Module that compactly represents temporal features by constructing the refined target distribution heatmap. Extensive experiments validate the effectiveness of our compact spatiotemporal modeling method, with CSTrack achieving new SOTA results on mainstream RGB-X benchmarks. The code and models will be released at: https://github.com/XiaokunFeng/CSTrack.

Figures

Figures reproduced from arXiv: 2505.19434 by the authors.

Figure 1
Figure 1. Framework of our proposed CSTrack. Given the RGB and X (e.g., thermal data) input streams at time t (t ≥ 1), the shared Patch Embedding initially transforms them into token sequences. Then, the Spatial Compact Module integrates them into a compact feature space, which is subsequently fed into a One-stream Backbone for comprehensive spatial modeling. Next, the Temporal Guidance Module uses the previously stored tempo… view at source ↗
Figure 2
Figure 2. Illustration of different target heatmaps (using an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison results of our tracker with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Tracking results of the model under different input [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Model performance variation (∆ precision) with different temporal lengths. 5. Conclusions To address the limitations of existing methods in handling RGB-X dual dispersed feature spaces, we propose a novel tracker named CSTrack, focusing on compact spatiotempo￾ral featu…
Figure 6
Figure 6. Figure 6: Processing workflows of symmetric and asymmetric dual-branch [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison results of our tracker with other two trackers ( [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages

  1. [1]

    The storage and update methods will be described later. w ht i. This setting selects the top-Nm search tokens from the search features based solely on ht i, which are then used to represent the compact temporal feature mt. w ht f . This setting selects the top-Nm search tokens from the search features based solely on ht f , which are then used to represen...

    work page 2024

  2. [4]

    Enhancing vision-language tracking by effectively converting textual cues into visual cues

    Feng, X., Zhang, D., Hu, S., Li, X., Wu, M., Zhang, J., Chen, X., and Huang, K. Enhancing vision-language tracking by effectively converting textual cues into visual cues. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5. IEEE,

    work page 2025

  3. [10]

    Dal: A deep depth-aware long-term tracker

    Qian, Y ., Yan, S., Lukeˇziˇc, A., Kristan, M., K ¨am¨ar¨ainen, J.-K., and Matas, J. Dal: A deep depth-aware long-term tracker. In 2020 25th International conference on pattern recognition (ICPR), pp. 7825–7832. IEEE,

    work page 2020

  4. [12]

    Transformer RGBT Tracking with Spatio-Temporal Multimodal Tokens

    Sun, D., Pan, Y ., Lu, A., Li, C., and Luo, B. Transformer rgbt tracking with spatio-temporal multimodal tokens. arXiv preprint arXiv:2401.01674,

    work page Pith review arXiv

  5. [13]

    Revisiting color-event based tracking: A unified network, dataset, and metric

    Tang, C., Wang, X., Huang, J., Jiang, B., Zhu, L., Zhang, J., Wang, Y ., and Tian, Y . Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010,

    arXiv

  6. [15]

    Visible- thermal uav tracking: A large-scale benchmark and new baseline

    Zhang, P., Zhao, J., Wang, D., Lu, H., and Ruan, X. Visible- thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, 2022a. Zhang, P., Wang, D., and Lu, H. Multi-modal visual track- ing: Review and experimental comparison. Computa- tional Visual Media, 10(2):193–214, 2024b. Zhang, X., Tian, Y ., Xie, L., Huang, W., Dai, Q., Ye, Q., and Ti...

    arXiv

  7. [16]

    Removal and selection: Improving rgb-infrared ob- ject detection via coarse-to-fine fusion

    Zhao, T., Yuan, M., Jiang, F., Wang, N., and Wei, X. Removal and selection: Improving rgb-infrared ob- ject detection via coarse-to-fine fusion. arXiv preprint arXiv:2401.10731,

    arXiv

  8. [17]

    ODTrack: Online Dense Temporal Token Learning for Visual Tracking

    Zheng, Y ., Zhong, B., Liang, Q., Mo, Z., Zhang, S., and Li, X. Odtrack: Online dense temporal token learning for visual tracking. arXiv preprint arXiv:2401.01686,

    work page Pith review arXiv

Show all 25 references
  1. [18]

    More Details on the RGB-X Benchmarks As discussed in Sec

    13 CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features A. More Details on the RGB-X Benchmarks As discussed in Sec. 4.2.1, we perform a comprehensive evaluation of CSTrack on mainstream RGB-X benchmarks. Specifically, the model’s performance on the RGB-Depth ...

    2022

  2. [19]

    In this section, we provide an overview of these benchmarks and their respective evaluation metrics

    and RGBT234 (Li et al., 2019), while the RGB-Event task is assessed based on VisEvent (Wang et al., 2023). In this section, we provide an overview of these benchmarks and their respective evaluation metrics. DepthTrack. DepthTrack (Yan et al., 2021b) is a comprehensive and lon...

    2019

  3. [20]

    This dataset employs an anchor-based short-term evaluation protocol (Kristan et al., 2020), which requires trackers to restart multiple times from different initialization points

    is the latest benchmark in RGB-D tracking, consisting of 127 short-term RGB-D sequences designed to explore the role of depth in RGB-D tracking. This dataset employs an anchor-based short-term evaluation protocol (Kristan et al., 2020), which requires trackers to restart multi...

    2020

  4. [21]

    As shown in Tab

    as a reference representing natural image distributions. As shown in Tab. 9, for the RGB modality, LaSOT and DepthTrack—both collected in typical natural environments—exhibit mean values that are similar to those of ImageNet. In contrast, Lasher, which often focuses on dark sc...

    2024

  5. [22]

    Our compact spatial modeling method, through the pro- posed Spatial Compact Module, integrates the features of both RGB and X modalities into a compact feature space

    models using the same backbone network and training strategy as CSTrack. Our compact spatial modeling method, through the pro- posed Spatial Compact Module, integrates the features of both RGB and X modalities into a compact feature space. This allows sub- sequent intra- and i...

    2024

  6. [24]

    These methods compress the temporal interaction between search and cue features into a small set of temporal queries, providing temporal guidance for the tracker

    and AQATrack (Xie et al., 2024). These methods compress the temporal interaction between search and cue features into a small set of temporal queries, providing temporal guidance for the tracker. Since these temporal queries can be seen as a compact way to construct temporal f...

    2024

  7. [2009]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    9 CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image reco...

    2010 arXiv

  8. [2014]

    Vmbench: A benchmark for perception-aligned video motion generation

    Ling, X., Zhu, C., Wu, M., Li, H., Feng, X., Yang, C., Hao, A., Zhu, J., Wu, J., and Chu, X. Vmbench: A benchmark for perception-aligned video motion generation. arXiv preprint arXiv:2503.10076,

    arXiv

  9. [2015]

    This approach has been widely adopted in the tracking field, such as in TrDiMP (Wang et al., 2021a) and JointNLT (Zhou et al., 2023)

    to represent compact temporal features. This approach has been widely adopted in the tracking field, such as in TrDiMP (Wang et al., 2021a) and JointNLT (Zhou et al., 2023). Specifically, we apply RoI processing to the search features st c using the predicted bounding box scal...

    2023

  10. [2016]

    Vasttrack: Vast category visual object tracking

    Peng, L., Gao, J., Liu, X., Li, W., Dong, S., Zhang, Z., Fan, H., and Zhang, L. Vasttrack: Vast category visual object tracking. arXiv preprint arXiv:2403.03493,

    arXiv

  11. [2018]

    The power of scale for parameter-efficient prompt tuning

    Lester, B., Al-Rfou, R., and Constant, N. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691,

    arXiv

  12. [2019]

    Explicit visual prompts for visual object tracking

    Shi, L., Zhong, B., Liang, Q., Li, N., Zhang, S., and Li, X. Explicit visual prompts for visual object tracking. arXiv preprint arXiv:2401.03142,

    arXiv

  13. [2020]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

    2009

  14. [2021]

    Dtllm-vlt: Diverse text generation for visual language tracking based on llm

    Li, X., Feng, X., Hu, S., Wu, M., Zhang, D., Zhang, J., and Huang, K. Dtllm-vlt: Diverse text generation for visual language tracking based on llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7283–7292, 2024a. Li, X., Hu, S., Feng,...

    arXiv

  15. [2022]

    Autoregressive queries for adaptive track- ing with spatio-temporaltransformers

    Xie, J., Zhong, B., Mo, Z., Zhang, S., Shi, L., Song, S., and Ji, R. Autoregressive queries for adaptive track- ing with spatio-temporaltransformers. arXiv preprint arXiv:2403.10574,

    arXiv

  16. [2023]

    ˇC., Lukeˇziˇc, A., Drbohlav, O., et al

    Kristan, M., Leonardis, A., Matas, J., Felsberg, M., Pflugfelder, R., K¨am¨ar¨ainen, J.-K., Danelljan, M., Zajc, L. ˇC., Lukeˇziˇc, A., Drbohlav, O., et al. The eighth visual object tracking vot2020 challenge results. In Computer Vision–ECCV 2020 Workshops: Glasgow, UK, August...

    2020

  17. [2024]

    Revealing the dark secrets of extremely large kernel con- vnets on robustness

    Chen, H., Zhang, Y ., Feng, X., Chu, X., and Huang, K. Revealing the dark secrets of extremely large kernel con- vnets on robustness. arXiv preprint arXiv:2407.08972,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.