Pith. sign in

REVIEW 3 major objections 9 minor 38 references

SpecTrack: Spectral Prompt Guided Adaptive Experts for Multispectral Object Tracking

T0 review · 3 major / 9 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Spectral band variation routes tracking compute where ambiguity demands it

desk verdict SpecTrack is a solid engineering contribution to MSI/HSI tracking with thorough ablations, but the mechanistic claim about spectral channel-variation routing is weakly supported. read the letter →

arxiv 2607.05988 v1 pith:QFCL54T5 submitted 2026-07-07 cs.CV

Xingyu Tan , Yunrong Qin , Mengjie Hu This is my paper · ORCID
classification cs.CV
keywords multispectralobjecttrackingadaptivecomputationmixtureofexpertsspectral-spatialmodelingcapacityallocationsparseroutingvisiontransformer
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

SpecTrack argues that multispectral object tracking should not treat every search region with the same computational budget. Instead, it frames tracking as per-region adaptive capacity allocation: a router inspects semantic context, spatial boundary sharpness, and—critically—the variation across latent channels derived from multispectral patch embeddings to estimate how ambiguous a given search region is, then activates a sparse subset of capacity-ordered experts ranging from lightweight local processors to deep, wide-receptive-field transformers. A shared global expert runs in parallel to supply common context and prevent the sparse experts from making fragmented decisions. The core wager is that the learned latent-channel variation after multispectral embedding carries a signal correlated with spectral-spatial ambiguity, and that this signal can drive meaningful routing decisions—assigning cheap computation to clear frames and expensive computation to hard ones. The system achieves 65.2% AUC on MUST, 51.9% on MSITrack, 72.6% on HOTC20, and 79.3% AO on GOT-10k, with ablations isolating each component's contribution.

What carries the argument

Spectral Prompt Router + SAMoE (capacity-ordered expert pool) + Shared Global Expert

What would settle it

If the latent channel-variation cue were replaced by a random fixed vector per sample and the tracking accuracy remained within 1 AUC point of the full model, the spectral routing interpretation would be undermined.

Watch

Extended reading notes

Core claim

The paper's central object is the Spectral Prompt Router, which combines three cues—semantic context (global average pooled features), spatial boundary response (a fixed Laplacian high-pass filter), and latent channel-variation (first-order finite differences across learned latent channels after multispectral patch embedding)—to select experts from a capacity-ordered pool called SAMoE. The experts are arranged along a ladder of increasing latent rank (from r_min to r_max), receptive field (4×4 to 32×32 windows), convolution kernel size (3×3 to 9×9), and depth (1 to 4 mixer units). A Shared Global Expert provides a sigmoid-modulated common context vector that is broadcast across all selected稀

Load-bearing premise

The load-bearing premise is that the variation across learned latent channels after multispectral patch embedding provides a signal reliably correlated with spectral-spatial ambiguity. The perturbation experiment shows random channel permutations reduce AUC by only 1.2 points on average, with one seed showing almost no effect (-0.1), and the correlation between channel-variation strength and high-capacity allocation is 0.60 without a formal significance test. If this signalis

Editorial extensions

If this is right

  • If latent channel variation is a reliable ambiguity proxy, then multispectral trackers could adaptively skip expensive computation on easy frames, reducing inference cost for deployment on edge devices like UAVs.
  • The capacity-allocation formulation could extend to other modalities with band-wise structure (e.g., multispectral change detection, hyperspectral classification) where per-sample difficulty varies.
  • The spectral-aware channel expansion strategy for initializing patch embeddings from RGB-pretrained weights provides a practical recipe for adapting RGB trackers to arbitrary band counts without retraining from scratch.
  • The GOT-10k result (79.3% AO) suggests the capacity-ordered expert design transfers beyond multispectral to RGB tracking, implying the adaptive computation idea is not spectrally dependent at inference even if it is spectrally motivated in design.

Reading between the lines

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

  • The moderate Spearman correlation (0.60) between channel-variation strength and high-capacity allocation, combined with the variable seed results in the channel-order perturbation experiment, suggests the routing signal may be weaker and noisier than the paper's framing implies. If so, the system might function primarily as a learned gating network with the spectral interpretation being a post-hoc
  • The capacity-ordered expert ladder (rank, receptive field, depth scaling linearly from 1/N to N/N) is a specific structural choice whose optimality is not tested against alternative orderings (e.g., logarithmic, learned schedules). The linear schedule may not be the best way to partition the capacity space.
  • If calibrated reflectance data and sensor response functions were available, the latent channel-variation cue could be replaced or augmented with physically grounded material-distinctiveness measures, potentially strengthening the routing signal beyond what learned latent features provide.
  • The paper does not test whether the router's behavior generalizes to unseen spectral band configurations at inference time (e.g., training on 8-band and testing on 16-band), which would determine whether the routing mechanism learns a genuinely band-agnostic ambiguity estimator or a band-count-specific gating function.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 9 minor

Summary. This paper proposes SpecTrack, a multispectral/hyperspectral object tracking framework that formulates tracking as search-region-level adaptive capacity allocation. The core architectural contribution is a Spectral Adaptive Mixture-of-Experts (SAMoE) module with capacity-ordered experts (varying in rank, receptive field, and depth) and a Spectral Prompt Router that uses semantic context, spatial boundary cues, and latent channel-variation cues to select a sparse subset of experts per search region. A Shared Global Expert provides common context. The system is evaluated on MUST, MSITrack, HOTC20, and GOT-10k (RGB generalization), with extensive mechanism-aligned ablations. The experimental design is notably thorough: the ablation tables (Tabs. 9–16) are organized as diagnostic questions, include same-backbone controls, and test channel-order sensitivity, routing diagnostics, and input modality contributions separately.

Significance. The paper makes a solid engineering contribution to MSI/HSI tracking by introducing capacity-ordered experts guided by spectral-spatial prompts, a design that is distinct from prior MoE trackers (MoETrack, SPMTrack, HotMoE) in its use of recorded-band-induced cues for routing. The ablation framework is well-constructed and addresses the right mechanistic questions. The GOT-10k evaluation (79.3% AO) demonstrates architectural generalization beyond multispectral settings. The promise of reproducible code is a strength. However, the interpretive claim that latent channel variation provides a spectrally meaningful routing signal is only weakly supported by the evidence, which tempers the mechanistic significance of the contribution even though the engineering results are competitive.

major comments (3)
  1. §4.4.4, Tab. 12: The channel-order permutation test is the central evidence for the claim that the latent channel-variation cue is spectrally meaningful rather than an arbitrary learned statistic. Across 5 random permutation seeds, AUC degradation ranges from -0.1 to -2.5 (mean -1.2, std 0.7). Seed 4 shows only -0.1 AUC, meaning a random permutation of channel order produces essentially no degradation. This is difficult to reconcile with a cue that genuinely depends on learned spectral-channel adjacency. The paper should either (a) increase the number of seeds and report a proper significance test (e.g., paired t-test or Wilcoxon against the identity baseline), or (b) acknowledge that the spectral specificity of this cue is weak and reframe the contribution accordingly. As it stands, the interpretive claim of 'recorded-band-induced ambiguity routing' is not adequately supported by this实验
  2. §4.4.7, Tab. 16: The Spearman correlation of 0.60 between channel-variation strength and high-capacity allocation (E3+E4 retained mass) is reported without a sample size, confidence interval, or significance test. Given that this correlation is the primary quantitative evidence linking the spectral cue to the routing mechanism's intended behavior, the absence of any statistical rigor is a gap. Please report the sample size, a p-value or bootstrap CI, and ideally a scatter plot or binned analysis showing the relationship. Without this, the reader cannot assess whether 0.60 reflects a genuine monotonic relationship or is an artifact of the logging procedure.
  3. §4.4.3, Tab. 11 and §4.4.5, Tab. 13: The contribution of the prompt routing mechanism is small in absolute terms. Tab. 11 shows the latent channel-variation cue adds 0.8 AUC over semantic-only routing (60.9→61.7), and Tab. 13 row 3 shows that semantic routing alone adds zero gain over capacity experts without routing (both 60.9). The entire prompt-routing contribution is 1.5 AUC (60.9→62.4), split between spatial and spectral cues. While small gains are not disqualifying, the paper's framing emphasizes the spectral routing mechanism as the central innovation. The authors should more explicitly contextualize these magnitudes—acknowledging that the primary gain comes from capacity-ordered experts (+1.8 AUC, Tab. 13 row 2) and that prompt routing provides a smaller incremental benefit—so that readers can calibrate expectations against the paper's framing.
minor comments (9)
  1. Abstract: 'object tracking object tracking' is duplicated in the opening sentence.
  2. §1, contributions list: 'The main contributions are summarized as follows: The main contributions are summarized as follows:' is duplicated.
  3. Tab. 2: SpecTrack-L384 reports 5.9 FPS, which is quite slow for practical tracking. The paper should discuss whether this operating point is intended purely as an accuracy upper bound or has practical deployment relevance.
  4. §4.2, Tab. 3: The HOTC20 entry notes '16 bands' but the text in §4.1 says HOTC20 covers '470–620 nm range.' Please clarify whether 16 bands within 150 nm is typical for this benchmark and whether the narrow spectral range limits the spectral routing cue's effectiveness on this dataset.
  5. Fig. 1 caption: The spectral curves are described as 'discrete digital-number or normalized-intensity traces' with a caveat about physical interpretation. This is appropriate, but the figure could benefit from error bars or shaded regions to indicate ROI selection uncertainty.
  6. §3.2, Eq. (9): The finite difference D_c operates on the normalized feature x-bar. It would help to clarify whether this is computed per-sample at inference or pre-computed, and whether the normalization in Eq. (10) is per-sample or uses running statistics.
  7. Tab. 12: The 'raw band' baseline (60.4 AUC, -2.0) is an important control but is mentioned only briefly. A sentence explaining why raw-band finite differences perform worse than latent-channel differences (e.g., because the learned embedding reorganizes spectral information) would strengthen the interpretation.
  8. References [5], [9]: The MUST and MSITrack datasets are co-authored by present paper authors. The paper should add an explicit note in the dataset section acknowledging this overlap, as is standard practice for self-authored benchmarks.
  9. §3.5: The loss weight λ_c is described as following 'the same code default in all datasets' but its value is not explicitly stated. Please report the numerical value for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee's three major comments all concern the strength of evidence supporting the interpretive claim that the latent channel-variation cue provides spectrally meaningful routing signal. We agree that this claim is currently over-stated relative to the evidence, and we will revise accordingly. Below we address each comment point by point.

read point-by-point responses

  1. Referee: §4.4.4, Tab. 12: Channel-order permutation test with 5 seeds shows AUC degradation from -0.1 to -2.5 (mean -1.2, std 0.7). Seed 4 shows only -0.1, difficult to reconcile with spectrally meaningful cue. Request more seeds with significance test or reframing.

    Authors: The referee is correct that the current evidence is insufficient to support the strong interpretive claim of 'recorded-band-induced ambiguity routing.' We acknowledge two issues: (1) five seeds is too few for a reliable significance test, and (2) the presence of seed 4 (-0.1 AUC) demonstrates that at least one random permutation produces negligible degradation, which is inconsistent with a cue that strictly depends on learned spectral-channel adjacency. In the revision we will take the following steps. First, we will increase the number of random permutation seeds to at least 20 and report a paired Wilcoxon signed-rank test against the identity baseline, along with the full distribution of per-seed degradations (histogram or strip plot). Second, we will reframe the interpretive claim. Rather than asserting that the latent channel-variation cue is 'spectrally meaningful' in the strong sense of depending on learned spectral-channel adjacency, we will state more precisely that the cue provides a weak but consistent routing signal: random permutations degrade performance on average, but the effect is modest and variable across seeds, indicating partial rather than strict order sensitivity. The contribution of the channel-variation cue will be described as an incremental routing signal that complements semantic and spatial cues, not as the central mechanistic innovation. We agree this reframing is necessary and will adjust the abstract, introduction, and §4.4.4 accordingly. revision: yes

  2. Referee: §4.4.7, Tab. 16: Spearman correlation of 0.60 reported without sample size, CI, or significance test. Request sample size, p-value or bootstrap CI, and scatter plot or binned analysis.

    Authors: The referee is correct that reporting a correlation coefficient without sample size, confidence interval, or significance test is inadequate. We will add the following in the revision: (1) the sample size (number of logged search-region samples from the MUST test split used to compute the correlation), (2) a bootstrap 95% confidence interval for the Spearman correlation, (3) a p-value from the corresponding permutation test, and (4) a binned scatter plot showing the relationship between latent channel-variation strength and pre-renormalization retained mass of E3+E4. We will also report the correlation for the ablated model without the latent channel prompt as a control, so the reader can compare the relationship under the intended routing signal versus its absence. If the confidence interval or significance test reveals that the correlation is not statistically robust, we will state this explicitly and downgrade the claim accordingly. revision: yes

  3. Referee: §4.4.3, Tab. 11 and §4.4.5, Tab. 13: Prompt routing contribution is small in absolute terms (0.8 AUC for latent cue, 1.5 AUC total for prompt routing vs. +1.8 from capacity experts). Paper framing over-emphasizes spectral routing. Request explicit contextualization of magnitudes.

    Authors: The referee's reading of the ablation magnitudes is accurate. The capacity-ordered expert design contributes +1.8 AUC over the uniform baseline (Tab. 13, row 2), while the full prompt-routing mechanism adds a further +1.5 AUC (60.9 to 62.4), of which the latent channel-variation cue accounts for +0.8 AUC (Tab. 11). These are modest incremental gains, and the current framing over-emphasizes the spectral routing mechanism relative to its actual quantitative contribution. In the revision we will: (1) explicitly state in the introduction and contributions list that the primary architectural gain comes from capacity-ordered experts, and that the Spectral Prompt Router provides a smaller but consistent incremental benefit; (2) re-order the contribution bullets to reflect this hierarchy, listing the SAMoE capacity-ordered expert pool as the primary contribution and the Spectral Prompt Router as a complementary mechanism; (3) add a sentence in §4.4.5 noting that semantic-only routing adds zero gain over capacity experts alone, and that the prompt routing benefit is concentrated in the spatial and spectral cues rather than in generic semantic routing. We agree that calibrating the framing to match the observed magnitudes is important for reader expectations. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity found; the architecture is defined by equations and validated empirically, not by self-referential definitions.

full rationale

The paper's derivation chain is self-contained. The Spectral Prompt Router computes its cues from input features via explicitly defined operations (Eqs. 6–10: GAP pooling, fixed Laplacian high-pass, finite-difference channel variation), combines them into routing logits (Eq. 12), and selects experts via TopK softmax (Eq. 13). The SAMoE experts are defined by rank projections, FFT mixers, and MLPs (Eqs. 17–24), and the Shared Global Expert provides modulation (Eqs. 25–27). No step reduces to its own inputs by construction. The ablations (Tabs. 9–16) test each component by removal or perturbation, not by re-deriving a fitted quantity. The self-citation of the MUST dataset [5] (co-authored by a present author with surname overlap) is a benchmark-familiarity concern but does not make the architectural claims circular: the routing mechanism is defined independently of benchmark results, and the system is also evaluated on MSITrack [9], HOTC20 [10], and GOT-10k [11], which are external. The Spearman 0.60 correlation (Tab. 16) is reported as a post-hoc diagnostic on logged samples, not as a derivation step or a fitted parameter renamed as a prediction. No uniqueness theorem is invoked, no ansatz is smuggled through self-citation, and no prediction is statistically forced by a fit. The one point is assigned for the minor self-citation of the MUST benchmark, which is not load-bearing for the architectural derivation.

Assumptions & free parameters 7 free parameters · 5 assumptions · 4 invented entities

The paper introduces several architectural entities (SAMoE, Spectral Prompt Router, Shared Global Expert, latent channel-variation prompt) with ablation support. Free parameters are mostly backbone-determined or code-defaulted, with several values not explicitly stated. The key domain assumption (latent channel variation as ambiguity proxy) receives only moderate experimental support.

free parameters (7)
  • r_min, r_max (expert rank bounds) = fixed by backbone width
    Eqs. 15-16 define nested rank and depth schedules; r_min and r_max are set by backbone width and kept unchanged across datasets, but their specific values are not stated in the paper.
  • t_max (max expert depth) = 4
    Eq. 16; t_max=4 for N=4 experts, producing depths 1,1,2,3 per Tab. 1 (though Tab. 1 shows 1,2,3,4, suggesting ceiling rather than floor).
  • alpha_l (residual scale) = learnable
    Eq. 4; alpha_l is a learnable per-layer residual scale controlling SAMoE contribution.
  • lambda_G, lambda_L1, lambda_moe, lambda_c (loss weights) = 2, 5, 0.01, code default
    Eq. 34 and surrounding text; lambda_c is stated as 'code default' without explicit value.
  • sigma_epsilon (routing noise std) = code default
    Eq. 12; Gaussian routing jitter with fixed code-default std, value not stated.
  • tau (normalization constant) = small constant
    Eqs. 8, 10, 30, 31; numerical stability constant, value not specified.
  • eta_c (channel expansion scale) = 1 or 1/2
    Eq. 35; eta_c=1 for insert, eta_c=1/2 for insert halfcopy, chosen per dataset.
assumptions (5)
  • domain assumption Latent channel variation after multispectral patch embedding correlates with spectral-spatial ambiguity
    Sec. 3.2; the entire routing mechanism depends on this assumption. Tab. 12 provides partial support but with high variance.
  • domain assumption Capacity-ordered experts (increasing rank, receptive field, depth) provide a meaningful representational ladder for tracking difficulty
    Sec. 3.3; assumed but not independently verified outside the ablation in Tab. 13.
  • domain assumption A Shared Global Expert reduces fragmented sparse-routing decisions
    Sec. 3.4; supported by Tab. 13 row 5 (removal drops AUC from 62.4 to 61.5).
  • domain assumption RGB-pretrained weights can be transferred to MSI via wavelength-proximity channel expansion
    Sec. 4.1, Eqs. 35-36; supported by Tab. 10 ablation.
  • domain assumption Standard MoE load-balancing loss (coefficient of variation) applies to heterogeneous-capacity experts
    Sec. 3.5, Eqs. 29-33; the complexity bias term (Eq. 32) is added to handle this, but the suitability of CV-based balancing for non-homogeneous experts is assumed.
invented entities (4)
  • SAMoE (Spectral Adaptive Mixture-of-Experts) module independent evidence
    purpose: Capacity-ordered expert pool for adaptive spectral-spatial feature transformation
    Ablated in Tab. 13 with same-backbone controls; routing behavior diagnosed in Tab. 16.
  • Spectral Prompt Router independent evidence
    purpose: Routes search regions to experts using semantic, spatial, and latent channel-variation cues
    Ablated in Tab. 11 (prompt type) and Tab. 12 (channel order sensitivity); provides falsifiable predictions about routing behavior.
  • Shared Global Expert independent evidence
    purpose: Provides common latent context to reduce fragmented sparse routing
    Ablated in Tab. 13 row 5; removal reduces AUC by 0.9.
  • Latent channel-variation prompt independent evidence
    purpose: First-order finite difference along learned latent channels as routing cue
    Tested in Tab. 12 with order perturbation and raw-band replacement; the evidence is mixed (high variance across seeds).

how reviews work

0 comments
Cite this review

Pith. "Pith review of SpecTrack: Spectral Prompt Guided Adaptive Experts for Multispectral Object Tracking." pith.science (2026). https://pith.science/paper/QFCL54T5

@misc{pith2026260705988,
  author       = {Pith},
  title        = {Pith review of: SpecTrack: Spectral Prompt Guided Adaptive Experts for Multispectral Object Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QFCL54T5}},
  note         = {Machine review of arXiv:2607.05988}
}
read the original abstract

Multispectral image(MSI) and hyperspectral image(HSI) object tracking object tracking exploits recorded band-wise observations to improve target--background discrimination under similar RGB appearance, mixed pixels, illumination variation, occlusion, and clutter. However, existing trackers commonly process all search regions through a fixed capacity spectral--spatial path, ignoring that tracking difficulty varies substantially across frames and target states. Clear regions may require only lightweight local discrimination, whereas ambiguous boundaries and spectrally similar distractors often demand stronger contextual reasoning. To address this limitation, we propose SpecTrack, a spectral--spatial complexity-aware tracker that formulates MSI tracking as search-region-level adaptive capacity allocation. Its core component, the Spectral Adaptive Mixture-of-Experts (SAMoE) module, provides a capacity-ordered expert pool with progressively increasing latent rank, receptive field, and depth. Expert selection is guided by a Spectral Prompt Router, which fuses semantic context, spatial boundary cues, and a latent channel-variation cue computed after multispectral patch embedding to activate a sparse subset of SAMoE experts for each search region. In parallel, a Shared Global Expert supplies common latent spectral--spatial context to reduce fragmented sparse-routing decisions. Experiments on MUST, MSITrack, and HOTC20 demonstrate a favorable accuracy--efficiency trade-off. The accuracy-oriented SpecTrack-L384 achieves state-of-the-art or highly competitive AUCs of 65.2\%, 51.9\%, and 72.6\% on the three benchmarks, while the balanced SpecTrack-B224 reaches 62.4\% AUC at 43.7 FPS on MUST. An additional GOT-10k evaluation indicates RGB-domain architectural generalization, with SpecTrack-L384 achieving 79.3\% AO.

Figures

Figures reproduced from arXiv: 2607.05988 by the authors.

Figure 1
Figure 1. Motivating examples for multispectral tracking. Target and background regions with similar RGB appearance can show separabil￾ity in recorded band-wise responses. The curves are discrete digital￾number or normalized-intensity traces; their physical interpretation depends on sensor response, bandpass, radiometric calibration and ac￾quisition conditions. time. This observation motivates a different question: How can a … view at source ↗
Figure 2
Figure 2. Mechanism level comparison of MoE-style trackers. MoETrack focuses on RGB-T modality validity and confidence-aware fusion; SPM￾Track uses tracking-specific MoE blocks for RGB spatio-temporal relation modeling; HotMoE performs sparse HSI expert activation. SpecTrack differs by using recorded-band-induced ambiguity cues to route each MSI/HSI search region to capacity ordered experts inside a shared one-stream backbone… view at source ↗
Figure 3
Figure 3. Overall framework of SpecTrack. Template and search MSI inputs are patch embedded and interact in a one-stream Transformer encoder. Standard blocks process the early encoder stages, while later blocks use SAMoE, which contains an adaptive expert set, a Spectral Prompt Router and a Shared Global Expert. The center head consumes the final search tokens and predicts the score map, target size, offset and bounding box. … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Attribute-level AUC comparison between SpecTrack and published trackers on the MUST dataset. The values in parentheses denote the strongest listed baseline AUC and SpecTrack AUC, respec￾tively. This naming follows the dataset source in Xiong et al. [10] and avoids conf…
Figure 5
Figure 5. Figure 5: Qualitative tracking examples on two representative MUST sequences with multiple challenge attributes. The predicted boxes illustrate localization behavior under similar color appearance, clut￾tered background and challenging target background separation. 4.4. Ablation…
Figure 6
Figure 6. Figure 6: Attention map comparison between the uniform process￾ing MSA baseline and the adaptive SAMoE block. The SAMoE￾enhanced block denotes the adaptive expert block used by SpecTrack, whereas MSA denotes the uniform baseline. The adaptive block pro￾duces more target-concentr…
Figure 7
Figure 7. Figure 7: Failure-oriented attribute analysis of SpecTrack-B224 on MUST. Attributes are sorted by AUC in ascending order. Red bars mark the four weakest attributes, and the dashed vertical line denotes the unweighted attribute mean. The right panel reports the percentage of fram…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [1]

    Z. Li, F. Xiong, J. Zhou, J. Wang, J. Lu, Y . Qian, Bae-net: A band attention aware ensemble network for hyperspectral object tracking, in: 2020 IEEE International Con- ference on Image Processing (ICIP), IEEE, 2020, pp. 2106–2110, iSSN: 2381-8549. doi:10.1109/ICIP40778.2020.9191105

    work page doi:10.1109/icip40778.2020.9191105 2020

  2. [4]

    Y . Chen, Q. Y uan, Y . Tang, Y . Xiao, J. He, L. Zhang, Spirit: Spectral awareness interaction network with dynamic template for hyperspec- tral object tracking, IEEE Transactions on Geo- science and Remote Sensing 62 (2024) 1–16. doi:10.1109/TGRS.2023.3347950

    work page doi:10.1109/tgrs.2023.3347950 2024

  3. [5]

    H. Qin, T. Xu, T. Li, Z. Chen, T. Feng, J. Li, Must: The first dataset and unified framework for multispectral uav single ob- ject tracking, in: 2025 IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 16882–16891, iSSN: 2575-

    work page 2025

  4. [6]

    doi:10.1109 /CVPR52734.2025.01573

    work page arXiv 2025

  5. [7]

    Z. Tang, T. Xu, X.-J. Wu, X. Zhu, C. Cheng, Z. Feng, J. Kittler, Revisiting rgbt tracking bench- marks from the perspective of modality valid- ity: A new benchmark, problem, and solution, IEEE Transactions on Image Processing 34 (2025) 7235–7249. doi:10.1109 /TIP .2025.3611687

    work page arXiv 2025

  6. [8]

    W. Cai, Q. Liu, Y . Wang, Spmtrack: Spatio- temporal parameter-efficient fine-tuning with mix- ture of experts for scalable visual tracking, in: 2025 IEEE /CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2025, pp. 16871–16881, iSSN: 2575-7075. doi:10.1109/CVPR52734.2025.01572

    work page doi:10.1109/cvpr52734.2025.01572 2025

  7. [9]

    W. Sun, Y . Tan, J. Li, S. Hou, X. Li, Y . Shao, Z. Wang, B. Song, Hotmoe: Exploring sparse mixture-of-experts for hyperspectral object track- ing, IEEE Transactions on Multimedia 27 (2025) 4072–4083. doi:10.1109 /TMM.2025.3535339

    work page arXiv 2025

  8. [10]

    T. Feng, T. Xu, H. Qin, T. Li, S. Han, X. Zou, Z. Lv, J. Li, Msitrack: A chal- lenging benchmark for multispectral single ob- ject tracking, in: Proceedings of the 33rd ACM International Conference on Multimedia, MM ’25, Association for Computing Machinery, New Y ork, NY , USA, 2025, pp. 12614–12620. doi:10.1145/3746027.3758197

    work page doi:10.1145/3746027.3758197 2025

Show all 38 references
  1. [11]

    Xiong, J

    F. Xiong, J. Zhou, Y . Qian, Material based object tracking in hyperspectral videos, IEEE Transac- tions on Image Processing 29 (2020) 3719–3733. doi:10.1109/TIP .2020.2965302

    2020 doi

  2. [12]

    Huang, X

    L. Huang, X. Zhao, K. Huang, Got-10k: A large high-diversity benchmark for generic object track- ing in the wild, IEEE Transactions on Pattern Analysis and Machine Intelligence 43 (5) (2021) 1562–1577. doi:10.1109 /TPAMI.2019.2957464

    2021

  3. [13]

    Uzkent, A

    B. Uzkent, A. Rangnekar, M. J. Ho ffman, Aerial vehicle tracking by adaptive fusion of hyperspec- tral likelihood maps, in: 2017 IEEE Conference on Computer Vision and Pattern Recognition Work- shops (CVPRW), 2017, pp. 233–242, iSSN: 2160-

    2017

  4. [14]

    doi:10.1109 /CVPRW.2017.35

    2017

  5. [15]

    Z. Li, X. Y e, F. Xiong, J. Lu, J. Zhou, Y . Qian, Spectral-spatial-temporal attention network for hyperspectral tracking, in: 2021 11th Work- shop on Hyperspectral Imaging and Signal Pro- cessing: Evolution in Remote Sensing (WHIS- PERS), IEEE, 2021, pp. 1–5, iSSN: 2158-6276....

    2021 doi

  6. [16]

    L. Gao, P . Liu, Y . Jiang, W. Xie, J. Lei, Y . Li, Q. Du, Cb ff-net: A new frame- work for e fficient and accurate hyperspectral object tracking, IEEE Transactions on Geo- science and Remote Sensing 61 (2023) 1–14. doi:10.1109/TGRS.2023.3253173

    2023 doi

  7. [17]

    Y . Chen, Q. Y uan, Y . Tang, Y . Xiao, J. He, Z. Liu, Sense: Hyperspectral video ob- ject tracker via fusing material and motion cues, Information Fusion 109 (2024) 102395. doi:10.1016/j.inffus.2024.102395

    2024 doi

  8. [18]

    L. Zhao, S. Xie, J. Li, P . Tan, W. Hu, Mvp-hot: A moderate visual prompt for hyperspectral ob- ject tracking, Journal of Visual Communication and Image Representation 105 (2024) 104326. doi:10.1016/j.jvcir.2024.104326

    2024 doi

  9. [19]

    Y . Chen, Q. Y uan, Y . Tang, X. Wang, Y . Xiao, J. He, Profit: A prompt-guided frequency- aware filtering and template-enhanced in- teraction framework for hyperspectral video tracking, ISPRS Journal of Photogrammetry and Remote Sensing 226 (2025) 164–186. doi:10.1016/j.isprsjp...

    2025 doi

  10. [20]

    Aljundi, P

    R. Aljundi, P . Chakravarty, T. Tuytelaars, Ex- pert gate: Lifelong learning with a net- work of experts, in: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 7120–7129, iSSN: 1063-6919. doi:10.1109/CVPR.2017.753

    2017 doi

  11. [21]

    Riquelme, J

    C. Riquelme, J. Puigcerver, B. Mustafa, M. Neu- mann, R. Jenatton, A. Susano Pinto, D. Keysers, N. Houlsby, Scaling vision with sparse mixture of experts, in: Advances in Neural Information Pro- cessing Systems, V ol. 34, Curran Associates, Inc., 2021, pp. 8583–8595

    2021

  12. [22]

    Puigcerver, C

    J. Puigcerver, C. Riquelme Ruiz, B. Mustafa, N. Houlsby, From sparse to soft mixtures of ex- perts, V ol. 2024, 2024, pp. 28435–28445

    2024

  13. [23]

    B. Y e, H. Chang, B. Ma, S. Shan, X. Chen, Joint feature learning and relation modeling for track- ing: A one-stream framework, in: S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, T. Hassner (Eds.), Computer Vision ECCV 2022, Springer Nature Switzerland, Cham, 2022, pp. 341–...

    2022 doi

  14. [24]

    H. Law, J. Deng, Cornernet: Detecting objects as paired keypoints, in: V . Ferrari, M. Hebert, C. Sminchisescu, Y . Weiss (Eds.), Computer Vi- sion ECCV 2018, Springer International Publish- ing, Cham, 2018, pp. 765–781. doi:10.1007 /978- 3-030-01264-9_45

    2018

  15. [25]

    Rezatofighi, N

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, S. Savarese, Generalized intersection over union: A metric and a loss for bound- ing box regression, in: 2019 IEEE /CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 658–666, iSSN: 2575-7075. d...

    2019 doi

  16. [26]

    B. Y an, H. Peng, J. Fu, D. Wang, H. Lu, Learning spatio-temporal transformer for vi- sual tracking, in: 2021 IEEE /CVF Interna- tional Conference on Computer Vision (ICCV), 2021, pp. 10428–10437, iSSN: 2380-7504. doi:10.1109/ICCV48922.2021.01028

    2021 doi

  17. [27]

    Zhang, Y

    X. Zhang, Y . Tian, L. Xie, W. Huang, Q. Dai, Q. Y e, Q. Tian, Hivit: A simpler and more e ffi- cient design of hierarchical vision transformer, in: International Conference on Learning Representa- tions, 2022

    2022

  18. [28]

    Y . Tian, L. Xie, J. Qiu, J. Jiao, Y . Wang, Q. Tian, Q. Y e, Fast-itpn: Integrally pre- trained transformer pyramid network with token migration, IEEE Trans. Pattern Anal. Mach. Intell. 46 (12) (2024) 9766–9779. doi:10.1109/TPAMI.2024.3429508

    2024 doi

  19. [29]

    Zheng, B

    Y . Zheng, B. Zhong, Q. Liang, Z. Mo, S. Zhang, X. Li, Odtrack: Online dense temporal to- ken learning for visual tracking, in: Pro- ceedings of the AAAI Conference on Artifi- cial Intelligence, V ol. 38, 2024, pp. 7588–7596. doi:10.1609/aaai.v38i7.28591

    2024 doi

  20. [30]

    Y . Kou, J. Gao, B. Li, G. Wang, W. Hu, Y . Wang, L. Li, Zoomtrack: Target-aware non-uniform re- sizing for efficient visual tracking, in: Advances in Neural Information Processing Systems, V ol. 36, 2023, pp. 50959–50977

    2023

  21. [31]

    C. Xu, B. Zhong, Q. Liang, Y . Zheng, G. Li, S. Song, Less is more: Token context- aware learning for object tracking, in: Pro- ceedings of the AAAI Conference on Artifi- cial Intelligence, V ol. 39, 2025, pp. 8824–8832. doi:10.1609/aaai.v39i8.32954

    2025 doi

  22. [32]

    Chen, C.-Y

    Y .-H. Chen, C.-Y . Wang, C.-Y . Y ang, H.-S. Chang, Y .-L. Lin, Y .-Y . Chuang, H.-Y . M. Liao, Neighbor- track: Single object tracking by bipartite match- ing with neighbor tracklets and its applications to sports, in: 2023 IEEE /CVF Conference on Com- puter Vision and Patte...

    2023

  23. [33]

    doi:10.1109 /CVPRW59228.2023.00542

    2023

  24. [34]

    L. Shi, B. Zhong, Q. Liang, N. Li, S. Zhang, X. Li, Explicit visual prompts for visual object tracking, in: Proceedings of the AAAI Conference on Arti- ficial Intelligence, V ol. 38, 2024, pp. 4838–4846. doi:10.1609/aaai.v38i5.28286

    2024 doi

  25. [35]

    J. Xie, B. Zhong, Z. Mo, S. Zhang, L. Shi, S. Song, R. Ji, Autoregressive queries for adap- tive tracking with spatio-temporal transform- ers, in: 2024 IEEE /CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024, pp. 19300–19309, iSSN: 2575-7075. doi:10.1109...

    2024 doi

  26. [36]

    S. Gao, C. Zhou, J. Zhang, Generalized re- lation modeling for transformer tracking, in: 2023 IEEE /CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 15 2023, pp. 18686–18695, iSSN: 2575-7075. doi:10.1109/CVPR52729.2023.01792

    2023 doi

  27. [37]

    S. Y u, J. Ni, S. Fu, T. Qu, Hyperspectral atten- tion network for object tracking, Sensors 24 (19) (2024) 6178. doi:10.3390 /s24196178

    2024

  28. [38]

    L. Lin, H. Fan, Z. Zhang, Y . Wang, Y . Xu, H. Ling, Tracking meets lora: Faster training, larger model, stronger performance, in: A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, G. V arol (Eds.), Computer Vision ECCV 2024, Springer, Springer Nature Switzerland, C...

    2024

  29. [39]

    X. Chen, H. Peng, D. Wang, H. Lu, H. Hu, Se- qtrack: Sequence to sequence learning for vi- sual object tracking, in: 2023 IEEE /CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 14572–14581, iSSN: 2575-

    2023

  30. [40]

    doi:10.1109 /CVPR52729.2023.01400. 16

    2023

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.