Family Matters: A Systematic Study of Spatial vs. Frequency Masking for Continual Test-Time Adaptation

Chandler Timm C. Doloriel; Fadi Al Machot; Habib Ullah; Kristian Hovde Liland; Muhammad salman siddiqui; Taki Hasan Rafi; Tor Kristian Stevik; Yeonguk Yu; Yunbei Zhang

REVIEW 3 major objections 5 minor 10 references

Family Matters: A Systematic Study of Spatial vs. Frequency Masking for Continual Test-Time Adaptation

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The masking family, not the selection strategy, decides whether continual test-time adaptation compounds useful structure or compounds errors.

desk verdict Solid empirical study of masking families in CTTA, but the headline causal claim outruns the evidence because equal mask fractions don't equal equal information removal. read the letter →

arxiv 2512.08048 v3 pith:W7BWHKPS submitted 2025-12-08 cs.CV

Chandler Timm C. Doloriel , Yunbei Zhang , Yeonguk Yu , Taki Hasan Rafi , Muhammad salman siddiqui , Tor Kristian Stevik , Fadi Al Machot , Kristian Hovde Liland

show 1 more author

Habib Ullah

This is my paper · ORCID

classification cs.CV

keywords continualtest-timeadaptationmaskedimagemodelingspatialvsfrequencymaskingstructuralpreservationcorruptionrobustnessvisiontransformersdistributionshift

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 isolates a single design choice—what gets masked—in continual test-time adaptation, where a model keeps adapting to a stream of corrupted images. Using a controlled setup that changes only the masking family, it claims that spatial patch masking accumulates useful structure over long streams, while frequency masking collapses once its masked band overlaps a corruption's spectral damage zone. The same family that works on vision transformers matters far less on CNNs and can be outdone by frequency masking on global-cue tasks with large models. A structural-preservation principle is offered to predict when a masking family is 'spectrally safe.'

What carries the argument

M2A, a minimal adapter that fixes selection to random and uses standard consistency plus entropy losses, so the only varying component is the masking family. The family axis includes spatial (patch, pixel) and frequency (all-band, low-band, high-band) masks applied through an easy-to-hard masking schedule. The carrying account is the structural-preservation principle: stable adaptation needs perturbations that preserve spatially contiguous redundancy and avoid bands already damaged by the corruption. This principle does the work of predicting which spectral-overlap conditions cause gradient collapse.

What would settle it

Measure, at a fixed masked fraction, the mutual information or linear-probe accuracy of masked views against clean labels, for patch and low-frequency masks. If low-frequency masks are informationally comparable yet still collapse over passes, the family claim stands; if they remove markedly more information, the collapse is explained by asymmetric information removal, and a ratio-calibrated comparison should shrink the gap.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that the masking family, not the cleverness of mask selection, governs whether continual test-time adaptation builds robust representations or compounds errors. Across three corruption benchmarks and several backbones, random patch masking on ViTs steadily improves over ten passes, while random low-frequency masking degrades from moderate error to near-chance after a few passes. The authors attribute this to structural preservation: spatially coherent masks keep broad-spectrum signal, while zeroing frequency bins can terminally coincide with a corruption's power spectrum—for example blur—generating degenerate views that corrupt gradients. On CNNs the gap la

Load-bearing premise

The controlled comparison assumes equal masked-area fraction means comparable information removal across families; the paper concedes zeroing frequency coefficients redistributes energy across every pixel, so frequency masking may remove more task-relevant structure even at identical ratios.

Editorial extensions

If this is right

On ViT backbones, random patch masking can match or beat heuristic selection strategies, suggesting the family axis carries much of the benefit often attributed to selection.
Frequency-masking failures compound: error rises across passes rather than recovering, so the family choice determines whether adaptation is self-correcting or self-destructive.
The spectral-overlap diagnostic lets practitioners predict unsafe families for a given corruption, e.g., low-frequency masking on blur, without running the full benchmark.
Architecture-task alignment matters: CNNs show little family gap, and global-cue tasks with large ViTs can prefer frequency masking.
Very small batches reintroduce instability, so the stability claim is conditional on enough signal for the consistency loss.

Reading between the lines

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

One extension the paper leaves open: a corruption-aware selection strategy that avoids the damaging band could plausibly rescue frequency masking, since the instability is characterized under random selection only.
If masking were calibrated so that equal masked fractions removed equal information, part of the family gap might shrink; the 'family determines stability' claim could then be refined to 'family plus information removal.'
The same structural-preservation principle could plausibly transfer to masking in other modalities or to pre-training, where reconstruction currently masks the spatial-frequency distinction.

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.

Desk Editor's Note

read the letter

Worth reading before you next pick a masking scheme for CTTA. The paper fills a real gap: it fixes selection strategy to random and varies only the masking family, across seven architectures and several benchmarks. No one had done that cleanly. The empirical breadth is genuine — multiple seeds with standard deviations, a lifelong ten-pass protocol, hyperparameter ablations, and a real-world aquaculture task. The consistent result, patch masking is stable on ViTs while frequency masking degrades and sometimes collapses, survives diverse protocols. The architecture/task conditioning is also useful: the family gap narrows on CNNs, and frequency masking becomes competitive on large ViTs with global-cue tasks. Credit where due: the authors are unusually honest about scope, and the system-level comparisons are labeled suggestive rather than overclaimed.

The soft spot is exactly the one A.7 concedes: equal masked fractions do not mean equal information removal. Frequency masking alters every pixel; patch masking leaves most pixels intact. So the comparison is confounded: frequency masking might collapse not because it is "frequency" but because it removes more task-relevant information per unit masked ratio. The hyperparameter sweeps vary the ratio but don't independently control information content. That means the causal claim in the abstract — "the masking family determines whether adaptation compounds useful structure or compounds errors" — is too strong. What is well-supported is the practical guidance: for ViTs with local cues, patch masking is a safe default; for large ViTs with global cues, frequency masking can be competitive. Those recommendations survive the confound. The structural-preservation principle is plausible but post-hoc; calling it predictive without a held-out test overstates it. No code or data artifacts either, which makes independent verification harder.

I'd send this to referees. The confound is fixable in principle — match information removal via mutual-information estimates or calibrate ratios per family — and the empirical core is broad enough that a revision can meaningfully strengthen it. The paper is honest, the experiments are extensive, and the design guidance is actionable. If I were refereeing, I'd ask for the information-calibration analysis and a softer causal claim, but I would not desk-reject.

Referee Report

3 major / 5 minor

Summary. The paper introduces M2A, a controlled continual test-time adaptation (CTTA) instantiation that fixes the selection strategy S=random and standard consistency/entropy losses, in order to isolate the masking family F (patch, pixel, all/low/high-frequency) as the sole experimental variable. Across CIFAR-10/100-C, ImageNet-C, CarlaTTA, and an aquaculture benchmark, with seven backbones and multiple streaming protocols, the authors report that spatial patch masking is the most stable family on ViTs, improving monotonically over ten lifelong passes (e.g., ImageNet-C mean error 37.4→33.3), while low-frequency masking degrades sharply (43.6→91.7). They propose a 'structural-preservation principle' to explain the instability as spectral overlap between the masking band and the corruption's damage zone, and they derive architecture- and task-dependent design guidance. The paper explicitly labels its system-level baseline comparisons as suggestive rather than controlled.

Significance. If the central comparison were clean, this would be a valuable empirical contribution: it gives actionable design guidance, covers a broad protocol space (three seed runs, seven architectures, multiple tasks), and is unusually honest about its limitations. The self-imposed controlled instantiation and the explicit separation of F from S are useful methodological steps for a field that has mostly confounded the two axes. However, the central causal attribution—that the masking family per se determines stability—is not yet cleanly established because the controlled comparison does not equalize information removal across families, and at least one architecture-level conclusion is contradicted by the paper's own table. The structural-preservation principle is a plausible post-hoc account, not yet a validated predictive theory.

major comments (3)

[A.7; Sections 3.2–3.3] The controlling variable in the family comparison is masked-area fraction mt, but Appendix A.7 concedes that 'equal masked ratios do not imply equal information removal': zeroing frequency coefficients redistributes energy globally and alters every pixel, whereas spatial patch masking leaves non-masked pixels intact. Therefore the observed gaps (e.g., Table 3: patch 37.4→33.3 vs low-freq 43.6→91.7) could reflect the amount of perturbation severity rather than the family identity. The low-frequency hyperparameter sweeps in Figure 16 vary mt but do not independently control information content. Please add a calibration condition that matches information removal across families (e.g., via mutual-information estimates or a task-relevant probe), or soften the headline claim to 'with equal masked-area fraction, patch masking is more stable than frequency masking' and treat the stronger causal
[Table 3] The claim that low-frequency masking 'collapses catastrophically' rests on mean errors with very large seed variance: from pass 3 onward the standard deviations are 25–34 percentage points (e.g., pass 3: 55.8±25.3; pass 4: 60.9±33.8). This indicates that the collapse is not uniform across seeds, and 'irreversible compounding' is not robustly demonstrated by the reported averages. Report per-seed trajectories, state how many of the three seeds actually enter the 80–90% regime, and characterize the distribution (e.g., is it bimodal?). Without this, the central stability claim is overstated.
[Section 4.6, Table 5] The text says that on traditional CNNs 'the family gap vanishes and the choice is less consequential,' but Table 5 shows F=low-freq outperforming F=patch on all three traditional CNNs by about 3 points (ResNext-50: 58.7 vs 61.7; WideResNet-50: 58.6 vs 61.5; ResNet-50: 62.2 vs 64.9). This is a systematic reversal, not a vanished gap, and it is hard to reconcile with the stated explanation that overlapping convolutional receptive fields dilute patch occlusion. Either provide statistical evidence that a ~3-point difference is within noise, or revise the architecture-scoping conclusion to acknowledge that on these CNNs low-frequency masking is actually better.

minor comments (5)

[Table 2(a)] F=high-freq reports exactly 90.0% error with zero standard deviation on every unseen corruption in CIFAR-10-C. This is consistent with a degenerate one-class collapse (10-class chance level). Please state this explicitly or verify that the constant is not an implementation artifact.
[Section 4.1; References] Typographical and formatting issues: 'Detailes' in Section 4.1; 'Internetional Conference' in the Niu et al. reference; inconsistent capitalization of CoTTA/Cotta in Table 4; and some table cells appear misaligned (e.g., Table 1b, VDP row).
[Appendix A.5.6] Cross-reference error: 'Figures 11 visualize...' should presumably be 'Figures 11a/11b' or similar. Several figure references in the appendix (e.g., Figures 8, 10) need consistency checks.
[Section 3 / A.7] The information-removal caveat is placed only in Appendix A.7, but it conditions the interpretation of every family comparison in Sections 4. Consider moving the caveat to Section 3 where the equal-mt assumption is first introduced, so readers can assess the confound from the start.
[Section 5] The 'structural-preservation principle' is presented as a predictive account, but the paper does not state a falsifiable prediction made before observing the results. Clarify whether the principle is intended as retrospective interpretation or as a forward-looking diagnostic, and if the latter, specify a concrete testable prediction.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the family axis is varied experimentally against fixed external benchmarks, and the only flagged caveat is a validity confound, not a constructional equivalence.

full rationale

The paper's central claim is empirical rather than analytic: M2A fixes S=random, the consistency/entropy losses, the masking schedule, and the one-step update protocol, so the only manipulated variable is the masking family. The headline numbers (e.g., Table 3: F=patch 37.4 to 33.3 vs F=low-freq 43.6 to 91.7 over ten ImageNet-C passes) are measured against fixed corruption benchmarks, not generated from any fitted parameter or from the definition of F. The 'structural-preservation principle' is presented as a predictive/explanatory account after the experiments, not as an equation whose output is the observed ranking; it is a post-hoc mechanistic interpretation that can be right or wrong but is not circular. The paper explicitly labels its confounded system-level comparisons with heuristic baselines as 'suggestive context rather than a controlled quantification of S', so no fitted input is being renamed as a prediction. Self-citations (DPCore, OT-VP) appear only as related-work baselines and are not load-bearing for the family claim; there is no uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result as a new derivation. The one passage that could be seen as undermining the central attribution is Appendix A.7: 'Equal masked ratios do not imply equal information removal... frequency masking may remove more task-relevant structure even at identical ratios.' This is an experimental-validity limitation about whether the effect is driven by F or by the amount of information removed, not a case where a prediction is identical to its input by construction. It is a confound to weigh in correctness, not circularity. No circular step meets the required evidentiary bar, so the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim is empirical rather than derivational, so there are no fitted constants in the usual sense. The free parameters are hyperparameters and the band cutoff; the main additional load is the ad-hoc structural-preservation principle and the representativeness of spectral footprints.

free parameters (4)

mask_step alpha = 0.1 (default)
Controls the masked-area fraction per view; ablated but hand-chosen.
number of views n = 3 (default)
Sweet spot from ablation; still a chosen hyperparameter.
entropy loss weight lambda = not stated (positive)
lambda=0 causes collapse; exact value not critical per ablation, but the central loss depends on it.
band cutoff r_c = 0.5
Defines low vs high frequency bands; not ablated, and it affects which corruptions spectrally overlap.

assumptions (5)

standard math Orthonormal 2D DFT and Hermitian conjugate-pair symmetry are preserved in frequency masking.
Section 3.2; standard properties used to construct real-valued masked views.
domain assumption ImageNet-C/CIFAR-C corruptions at severity 5 represent the CTTA distribution-shift regime.
Section 4.1; standard in the field, but the centrality of these benchmarks is an empirical scope choice.
domain assumption Average power spectra over 5,000 images per corruption capture the corruption's spectral signature.
Appendix A.4, Figure 6; if unrepresentative, the spectral-overlap explanation loses its grounding.
ad hoc to paper The structural-preservation principle — spatially coherent masking preserves broad-spectrum redundancy and avoids spectral overlap — is a valid explanatory account.
Introduced after observing the results; illustrated with power spectra and CAMs but not independently validated as a predictive rule.
domain assumption Updating only layer-norm parameters (following REM) is sufficient to expose family-dependent behavior.
Section 4.2; restricts scope to the parameter-update regime common in CTTA baselines.

invented entities (1)

Structural-preservation principle
purpose: Explains why spatial masking is stable: spatially coherent views preserve broad-spectrum redundancy and avoid terminal spectral overlap with corruption.
Post-hoc explanatory framework; supported only by averaged power spectra, GradCAM images, and t-SNE plots generated for this study, with no independent falsifiable prediction outside the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Family Matters: A Systematic Study of Spatial vs. Frequency Masking for Continual Test-Time Adaptation." pith.science (2026). https://pith.science/paper/W7BWHKPS

@misc{pith2026251208048,
  author       = {Pith},
  title        = {Pith review of: Family Matters: A Systematic Study of Spatial vs. Frequency Masking for Continual Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7BWHKPS}},
  note         = {Machine review of arXiv:2512.08048}
}

read the original abstract

Recent continual test-time adaptation (CTTA) methods adopt masked image modeling to stabilize learning under distribution shift, yet each treats its masking family F as a fixed design choice and innovates exclusively along the selection strategy S, leaving the family axis underexplored. We present a systematic empirical study that isolates this axis. Using a controlled CTTA instantiation -- Mask to Adapt (M2A) -- that fixes S = random and standard losses, we vary only F across spatial (patch, pixel) and frequency (all-band, low-band, high-band) families while keeping every other component identical. The study's contributions are the design guidance it extracts for the CTTA settings we evaluated: (1) the masking family determines whether adaptation compounds useful structure or compounds errors -- on patch-tokenized architectures, spatial masking accumulates stable representations over long streams while frequency masking collapses catastrophically. We characterize this instability through a structural-preservation account, where spatial coherence maintains the broad-spectrum redundancy needed to avoid terminally overlapping with a corruption's spectral signature; (2) the optimal family depends on architecture-task alignment -- on CNNs, whose overlapping receptive fields dilute patch occlusion, the family gap vanishes, whereas on fine-grained tasks with global cues and large-capacity ViTs, frequency masking becomes competitive. In confounded system-level comparisons -- where baselines also differ in losses and auxiliary components -- M2A's random selection performs comparably to heuristic strategies, though we treat this observation as suggestive context rather than a controlled quantification of S's relative importance.

Figures

Figures reproduced from arXiv: 2512.08048 by the authors.

**Figure 1.** Motivation: The masking family axis is underexplored in CTTA. Recent masking-based CTTA methods couple a specific masking family F with a specific selection strategy S, innovating exclusively along the strategy axis while treating the family as a fixed design choice. Left: Standard CTTA pipeline showing error accumulation over time under distribution shift. Center: The two orthogonal design axes— masking family (spa… view at source ↗

**Figure 2.** Masking types in classification error rate (%) under CTTA using ViT-B/16 backbone. Each [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Class activation maps using GradCAM on ImageNet-C with severity 5 (highest). [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (14 more)

**Figure 4.** Figure 4: t-SNE visualizations of forward features from the adapted ViT-B/16 backbone on ImageNet-C [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: M2A(F=patch) ablations on ImageNet-C under CTTA. Each panel shows the effect of a single hyperparameter: (a) Lambda λ, (b) mask views, (c) mask steps, (d) gradient steps, and (e) batch size. diminishing returns beyond two; batch sizes from 4 to 128 keep performance in …

**Figure 6.** Figure 6: Average power spectra for 15 ImageNet-C corruption types (severity 5). Each plot shows the [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

**Figure 7.** Figure 7: Visualization of the segmentation results in CarlaTTA under CTTA. We use DeepLabV2 as the [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]

**Figure 8.** Figure 8: Per-corruption error profiles across seven architectures on ImageNet-C under CTTA, extending [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

**Figure 9.** Figure 9: Samples of MRSFFIA-C dataset at severity level 5 (highest) with different corruption domains [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]

**Figure 10.** Figure 10: Feature visualizations of CIFAR10-C at severity level 1 and 5 obtained from ViT-B/16 (source) [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]

**Figure 11.** Figure 11: Feature visualizations of MRSFFIA-C at severity level 1 and 5 obtained from ViT-B/16 (source) [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]

**Figure 12.** Figure 12: Ablations on loss functions: Consistency Loss (CL), Entropy Loss (EL), and Ranking Loss (RL) [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]

**Figure 13.** Figure 13: M2A(F=patch) classification error under CTTA as the number of masked patches increases. Performance is stable at low counts but catastrophically collapses once patches exceed a dataset-dependent threshold (eight on CIFAR-100-C, four on ImageNet-C), consistent with fin…

**Figure 14.** Figure 14: M2A(F=patch) ablations on CIFAR10-C under CTTA. Each panel shows the effect of a single hyperparameter: (a) Lambda λ, (b) mask views, (c) mask steps, (d) gradient steps, and (e) batch size. A.6.3 Domain Generalization Details [PITH_FULL_IMAGE:figures/full_fig_p026_14.png]

**Figure 15.** Figure 15: Effect of very small batch sizes on M2A(F=patch) under CTTA. Each curve sweeps batch sizes {1, 2, 3, 4, 8} over the 15 corruptions. While moderate batches (3, 4, 8) stay in a similar performance range, batches of size 1–2 can trigger corruption-specific collapse, with…

**Figure 16.** Figure 16: M2A(F=low-freq) ablations on ImageNet-C under CTTA. Each panel shows the effect of a single hyperparameter: (a) Lambda λ, (b) mask views, (c) mask steps, (d) gradient steps, and (e) batch size. Compare with [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]

**Figure 17.** Figure 17: M2A(SM:Patch) predicted samples from ImageNet-C (severity 5) under CTTA. We save the images [PITH_FULL_IMAGE:figures/full_fig_p028_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 2 linked inside Pith

[3]

All notation follows the main text:x∈ [0, 1]C×H×W is the input image,mt = tα is the masked-area fraction for viewt, andM (t)∈{ 0, 1}H×W is the binary mask broadcast across channels as ˜M (t) = broadcastC(M (t)). A.3.1 Spatial Masking Patch-based masking.Given the target masked fraction mt and a patch countPt, each square patch has side length s= ⌊√ mtHW P...

2000
[5]

Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy

URLhttps: //arxiv.org/abs/2411.03687. Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training.ICLR,

arXiv
[6]

Simmim: a simple framework for masked image modeling.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: a simple framework for masked image modeling.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

2022
[7]

Robust test-time adaptation in dynamic scenarios.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

Longhui Yuan, Binhui Xie, and Shuangliang Li. Robust test-time adaptation in dynamic scenarios.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

2023
[8]

Ot-vp: Optimal transport-guided visual prompting for test-time adaptation

Yunbei Zhang, Akshay Mehra, and Jihun Hamm. Ot-vp: Optimal transport-guided visual prompting for test-time adaptation. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1122–1132. IEEE, 2025a. Yunbei Zhang, Akshay Mehra, Shuaicheng Niu, and Jihun Hamm. DPCore: Dynamic prompt coreset for continual test-time adaptation. InFort...

2025
[10]

consists of top-down video frames ofPunctatus Oplegnathus cultured in a recirculating tank (diameter 3m, depth 0.75m) in Yantai City, Shandong Province, China. The task—feeding-intensity recognition—targets a core aquaculture monitoring problem: estimating feeding activity from overhead cameras to support operational decisions (e.g., feeding control) with...

2018
[2022]

Fremim: Fourier transform meets masked image modeling for medical image segmentation.2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV),

Wenxuan Wang, Jing Wang, Chen Chen, Jianbo Jiao, Lichao Sun, Yuanxiu Cai, Shanshan Song, and Jiangyun Li. Fremim: Fourier transform meets masked image modeling for medical image segmentation.2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV),

2024
[2023]

Is less more? exploring token condensation as training-free test-time adaptation

Zixin Wang, Dong Gong, Sen Wang, Zi Huang, and Yadan Luo. Is less more? exploring token condensation as training-free test-time adaptation. InInternational Conference on Computer Vision (ICCV), 3 2025b. URLhttp://arxiv.org/abs/2410.14729. Chen Wei, Haoqi Fan, Saining Xie, Chaoxia Wu, Alan Loddon Yuille, and Christoph Feichtenhofer. Masked feature predicti...

arXiv 2022

Show all 10 references

[2024]

doi: 10.1007/ s11263-024-02181-w

ISSN 1573-1405. doi: 10.1007/ s11263-024-02181-w. URLhttp://dx.doi.org/10.1007/s11263-024-02181-w. Jiaming Liu, Ran Xu, Senqiao Yang, Renrui Zhang, Qizhe Zhang, Zehui Chen, Yandong Guo, and Shanghang Zhang. Continual-mae: Adaptive distribution masked autoencoders for continual...

2024 doi
[2025]

Girshick

Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll’ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [3]

All notation follows the main text:x∈ [0, 1]C×H×W is the input image,mt = tα is the masked-area fraction for viewt, andM (t)∈{ 0, 1}H×W is the binary mask broadcast across channels as ˜M (t) = broadcastC(M (t)). A.3.1 Spatial Masking Patch-based masking.Given the target masked fraction mt and a patch countPt, each square patch has side length s= ⌊√ mtHW P...

2000

[2] [5]

Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy

URLhttps: //arxiv.org/abs/2411.03687. Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training.ICLR,

arXiv

[3] [6]

Simmim: a simple framework for masked image modeling.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: a simple framework for masked image modeling.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

2022

[4] [7]

Robust test-time adaptation in dynamic scenarios.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

Longhui Yuan, Binhui Xie, and Shuangliang Li. Robust test-time adaptation in dynamic scenarios.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

2023

[5] [8]

Ot-vp: Optimal transport-guided visual prompting for test-time adaptation

Yunbei Zhang, Akshay Mehra, and Jihun Hamm. Ot-vp: Optimal transport-guided visual prompting for test-time adaptation. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1122–1132. IEEE, 2025a. Yunbei Zhang, Akshay Mehra, Shuaicheng Niu, and Jihun Hamm. DPCore: Dynamic prompt coreset for continual test-time adaptation. InFort...

2025

[6] [10]

consists of top-down video frames ofPunctatus Oplegnathus cultured in a recirculating tank (diameter 3m, depth 0.75m) in Yantai City, Shandong Province, China. The task—feeding-intensity recognition—targets a core aquaculture monitoring problem: estimating feeding activity from overhead cameras to support operational decisions (e.g., feeding control) with...

2018

[7] [2022]

Fremim: Fourier transform meets masked image modeling for medical image segmentation.2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV),

Wenxuan Wang, Jing Wang, Chen Chen, Jianbo Jiao, Lichao Sun, Yuanxiu Cai, Shanshan Song, and Jiangyun Li. Fremim: Fourier transform meets masked image modeling for medical image segmentation.2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV),

2024

[8] [2023]

Is less more? exploring token condensation as training-free test-time adaptation

Zixin Wang, Dong Gong, Sen Wang, Zi Huang, and Yadan Luo. Is less more? exploring token condensation as training-free test-time adaptation. InInternational Conference on Computer Vision (ICCV), 3 2025b. URLhttp://arxiv.org/abs/2410.14729. Chen Wei, Haoqi Fan, Saining Xie, Chaoxia Wu, Alan Loddon Yuille, and Christoph Feichtenhofer. Masked feature predicti...

arXiv 2022