Pith. sign in

REVIEW 4 major objections 5 minor 94 references

Orthogonal Projection Subspace to Aggregate Online Prior-knowledge for Continual Test-time Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that freezing the source model, tuning only low-rank adapters under a soft orthogonality penalty, and training an aggressively masked student on teacher pseudo labels solves continual test-time adaptation for semantic…

desk verdict Strong empirical recipe for continual TTA; the 'orthogonal projection subspace' mechanism is not supported by Eq. (4), but the benchmark work and ablations justify sending to referees. read the letter →

arxiv 2506.19022 v1 pith:JSNY3WJF submitted 2025-06-23 cs.CV

Jinlong Li , Dong Zhao , Qi Zang , Zequn Jie , Lin Ma , Nicu Sebe This is my paper · ORCID
classification cs.CV
keywords continualtest-timeadaptationsemanticsegmentationorthogonalregularizationlow-rankLoRAimagemaskingteacher-studentframeworkdomainshift
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 is trying to establish that continual test-time adaptation—a model adapting online to a never-ending stream of unlabeled images from changing domains, seeing each sample once—can be made practical for semantic segmentation by freezing the entire source model and tuning only low-rank adapters. Its central claim is that softly pushing those adapter updates toward an orthogonal subspace preserves the semantic knowledge already encoded in the frozen weights, so the model can take in foggy, rainy, night, and snowy domains without forgetting earlier ones. A second mechanism, an image-masking strategy, feeds the student aggressively masked images while an EMA teacher supplies pseudo labels from the original image, which the paper argues reduces error accumulation by forcing the student to exploit context. If the paper is right, state-of-the-art continual adaptation follows from roughly one million trainable parameters (1.04M for the rank-4 variant), with the learned updates merged back into the original weights so deployment costs nothing extra.

What carries the argument

The central object is the orthogonal projection subspace: the set of low-rank update matrices $\Delta W = BA$ whose columns are softly constrained by $L_{\text{orth}} = \|(BA)^T(BA) - I\|^2$ to act like an orthogonal transform of the frozen source weights, so that adaptation proceeds by adjusting coordinate angles rather than magnitudes. The second mechanism is the Image Masking Strategy (IMS), which builds a uniform random grid mask, thresholds it at ratio $\alpha$ with grid size $s$, and fills masked student-input pixels with 0 or 255 proxy values to simulate night or snow-like conditions while the teacher sees the original image. The two mechanisms work through a teacher–student (EMA) loop: the teacher's pseudo labels supervise the student's segmentation loss on masked input, and the student's updated low-rank adapters are then averaged back into the teacher. OPS is the component aimed at catastrophic forgetting; IMS is the component aimed at error accumulation.

What would settle it

Measure the actual orthogonality of the learned updates during an ACDC run, for example the singular values of $BA$ and the value of $\|(BA)^T(BA)-I\|$: if the updates remain far from orthogonal while forgetting stays low, the orthogonality term is not the operative mechanism. A second check is to re-run the angular-reconstruction toy on SegFormer-B5 features instead of the Stanford Dogs encoder-decoder; if angles alone do not reconstruct those features, the stated motivation for OPS is undercut.

Watch

Extended reading notes

Core claim

The paper's central claim, stated as a sympathetic reader would state it, is that the two failure modes of continual test-time adaptation—catastrophic forgetting and error accumulation—can be attacked simultaneously with a simple recipe. The proposed pipeline, OoPk, freezes the pre-trained source weights $W_0$ and learns only the low-rank update $\Delta W = BA$, adding the soft orthogonality loss $L_{\text{orth}} = \|(BA)^T(BA) - I\|^2$ so that the update behaves like a rotation of the source weight directions; the teacher model is an exponential moving average of the student and provides pseudo labels from unmasked inputs, while the student sees inputs with random grid patches replaced by 0 or 255 proxy pixels under the Image Masking Strategy. On the Cityscapes-to-ACDC continual benchmark the rank-4 variant reaches 61.3 mean intersection-over-union (mIoU) without warmup using 1.04M trainable parameters (63.8 at rank 32 with warmup), and on SHIFT it reaches 69.32 mean mIoU; the paper also reports a drop in CIFAR100-to-CIFAR100C average error to 34.9% at rank 32, in each case matching or beating prior CTTA methods with fewer trainable parameters and no inference overhead after merging $A$ and $B$ into $W_0$. The motivation is supported by a toy experiment on Stanford Dogs in which images can be reconstructed from the angular information of the weights alone, suggesting that rotating weight directions is a knowledge-preserving way to adapt.

Load-bearing premise

The load-bearing premise is that keeping the directions of the added low-rank weights close to perpendicular to each other preserves what the frozen source model already knows, so the model can learn new domains without forgetting old ones; nothing in the paper proves that this holds for low-rank updates where the perpendicularity constraint can only be approximate.

Editorial extensions

If this is right

  • On the Cityscapes-to-ACDC benchmark, the rank-4 variant reaches 61.3 mean mIoU without any warmup using only 1.04M trainable parameters, surpassing BECoTTA and matching heavier methods such as SVDP and C-MAE.
  • On the SHIFT continuous validation set, OoPk reaches 69.32 mean mIoU, above the source model and all prior CTTA methods listed in the paper.
  • Because $A$ and $B$ can be merged into $W_0$ after adaptation, the deployed model has exactly the inference cost of the frozen source model, unlike prompt- or mixture-of-experts methods that must store and compute extra routing structures.
  • The pipeline is architecture-agnostic: the paper shows it works for SegFormer semantic segmentation and for a WideResNet classifier, suggesting the recipe transfers across tasks.
  • Each component contributes independently: in the paper's ablation, LoRA alone gives 58.63 mIoU, adding the orthogonality loss gives 59.82, and adding image masking gives 61.30, with performance insensitive to the order in which weather domains arrive.

Reading between the lines

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

  • Editorial inference: the same freeze-and-rotate recipe could be applied to other dense prediction tasks, such as monocular depth estimation or panoptic segmentation under continuous domain shift, since nothing in the pipeline depends on semantic classes.
  • Editorial inference: the paper's angular-information story predicts that a genuinely orthogonal parameterization of the updates, for example via Cayley or QR reparameterization instead of a soft penalty, would hold the forgetting curve flat for even longer; the paper does not test this.
  • Editorial inference: because the orthogonality penalty on $BA$ cannot drive $(BA)^T(BA)$ toward $I$ when rank $r$ is much smaller than the layer width, the method's orthogonal subspace is necessarily an approximation, and one measurable consequence is that results should depend nontrivially on rank, which the paper's own ablation confirms.
  • Editorial inference: making the masking ratio or grid size adaptive to per-batch uncertainty could push the IMS gain further, but that is an extension the paper leaves open, not a claim it makes.
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, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes OoPk, a continual test-time adaptation (CTTA) pipeline for semantic segmentation that combines low-rank LoRA-style tuning with a soft orthogonality regularizer (OPS) and an image masking strategy (IMS) inside an EMA teacher-student framework. The stated mechanism is that rotating pre-trained weights inside an orthogonal subspace preserves source knowledge while absorbing new domain knowledge, thereby reducing catastrophic forgetting and error accumulation. The method is evaluated on Cityscapes-to-ACDC, SHIFT, and CIFAR-100-to-CIFAR-100C, reporting competitive mIoU/accuracy with 1.04M trainable parameters and no inference overhead via reparameterization.

Significance. If the empirical results hold, OoPk is practically valuable: it achieves strong CTTA performance with very few trainable parameters, works with a frozen source model, and adds no inference latency after merging the low-rank factors. The ablations clearly show that both the orthogonality loss and the masking strategy contribute positively, and the experiments cover a dense-prediction setting (ACDC/SHIFT) that is less common than classification benchmarks. However, the paper does not ship code or data, reports single runs without variance, and, most importantly, the theoretical explanation in §3.2 is not supported by the actual regularizer in Eq. (4), which cannot make a low-rank update orthogonal. The contribution should therefore be viewed as an empirical recipe whose mechanism is not yet established.

major comments (4)
  1. [§3.2, Eq. (4)] The regularization objective cannot be satisfied for the low-rank update actually used. With B ∈ R^{d×r} and A ∈ R^{r×k}, the matrix (BA)^T(BA) has rank at most r, so it cannot equal the k×k identity when r ≪ min(d,k); the loss has a positive lower bound and the optimization can only push the r nonzero singular values of BA toward 1. Consequently, the adapted weight W0+BA is not an orthogonal transformation of W0, and the claimed "orthogonal projection subspace" mechanism for preserving source knowledge is not implemented as stated. The authors should either reformulate the regularizer so that the orthogonality claim is mathematically accurate (e.g., regularizing singular values or projecting onto the Stiefel manifold) or provide direct evidence that this specific soft constraint, rather than a generic low-rank regularizer, preserves source semantics.
  2. [Tables 2 and 8] The abstract and introduction state that OoPk 'surpasses' previous CTTA methods, but in the no-warmup setting OoPk r=4 obtains 61.3 mIoU on Cityscapes-to-ACDC, which equals SVDP and is below C-MAE's 61.8; OoPk r=32 ties C-MAE at 61.8. The state-of-the-art claim is therefore only valid under the additional criterion of parameter efficiency, and the paper should either state this explicitly or provide a Pareto-style comparison of performance against trainable parameters.
  3. [§9.1 and Table 7] The warmup variant 'OoPk*' is trained on style-transferred Cityscapes images with full label accessibility before adaptation (Table 7: 'Label accessibility Yes'). This is not consistent with the source-free CTTA protocol described in §3.1, and comparing warmup variants against methods that do not use source labels is not apples-to-apples. The paper should clearly separate the warmup and no-warmup protocols in the abstract and contribution claims, and should not present warmup results as evidence for the core CTTA contribution.
  4. [Supplementary §6, Fig. 7] The toy experiment demonstrates that angular information alone can reconstruct images for a small encoder-decoder on Stanford Dogs, but it does not test whether the proposed LoRA-plus-soft-orthogonality update preserves source semantics on the actual SegFormer-B5 model or on the ACDC/SHIFT distributions. Since the orthogonality mechanism is the stated explanation for reduced forgetting, the paper needs either a direct test of whether W0+BA with the Eq. (4) regularizer behaves as a direction-preserving update, or a more cautious claim that L_orth is an empirically useful regularizer rather than a mechanism that provably preserves knowledge.
minor comments (5)
  1. [Tables 2 and 8] The parameter count for OoPk* r=4 is reported as 1.04M in Table 2 but 8.32M in Table 8; please clarify whether warmup changes the number of trainable parameters or whether one of the tables contains a typo.
  2. [§4 experiments] All results are reported as single runs without standard deviations or multiple seeds. For a CTTA method that involves random masking and online adaptation, reporting variance over at least a few seeds would materially improve reproducibility and support the claimed improvements over baselines.
  3. [§3.2, Eq. (2)] Eq. (2) applies W_i^T W_i − I without specifying whether W_i is square; for non-square layer weights the identity has ambiguous dimension. Please state the intended identity dimension and how the regularizer is applied to non-square matrices.
  4. [Supplementary, Sec. 9] The text says 'we will release our synthetic data' but does not mention code release; please state a clear availability plan for both code and data, and use stable repository URLs instead of anonymous links.
  5. [Figures 6 and 8] The axis labels and legends in Figs. 6 and 8 are difficult to read in the provided resolution; please ensure the final figures are legible or enlarge the relevant panels.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: OoPk's claims are supported by external benchmark comparisons and independent ablations.

full rationale

The paper's central claims—that LoRA with a soft orthogonality regularizer (Eq. 4) plus an image-masking student-teacher framework improves continual test-time adaptation—are validated against external benchmarks (Cityscapes-to-ACDC, SHIFT, CIFAR100-to-CIFAR100C) and through component-level ablations (Table 6). No fitted parameter is renamed as a prediction: hyperparameters such as rank r, masking grid size s, masking ratio α, and loss weight λ are ablated and reported as design choices, not derived from the target metrics. The orthogonality regularizer is justified by a toy experiment (Fig. 4/7) that is independent of the final benchmarks and does not encode the reported mIoU values. Self-citations appear only in a future-work enumeration and are not load-bearing for any derivation. The concern that Eq. (4) cannot drive (BA)^T(BA) to I when rank r << k is a correctness/implementability issue about the mechanism, not a circularity: the claim does not reduce to its inputs by construction, because the objective is not defined in terms of the reported outcomes.

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

The method does not introduce new physical or architectural entities; its components are LoRA adapters, an orthogonality regularization loss, and a masking augmentation. The main untested assumptions are the semantic-preservation property of orthogonal updates and the usefulness of aggressive 0/255 masking. The warmup variant adds a labeled synthetic data assumption that is not part of the strict CTTA protocol. Several hyperparameters are tuned per benchmark, most notably rank, grid size, masking ratio, and lambda.

free parameters (5)
  • LoRA rank r = 32
    Selected by ablation on ACDC; lower ranks underfit and higher ranks overfit (Fig. 6, Section 4.2).
  • IMS masking ratio alpha = 0.75
    Chosen as best in the masking ratio ablation (Fig. 6 and Fig. 9).
  • IMS grid size s = 32
    Selected as best grid size; larger grid sizes degrade performance (Section IMS analysis).
  • Orthogonality loss weight lambda = 1.0
    Set by default after sensitivity analysis; performance is reported as insensitive to lambda.
  • EMA smoothing factor beta = not reported
    Defined in Eq. (9) as the teacher update weight but no numeric value is given in implementation details.
assumptions (5)
  • ad hoc to paper Angles and directions of network weights encode the semantically important information, so rotating weight directions should preserve source knowledge.
    Used to justify the orthogonality penalty in Section 3.1; supported only by a toy experiment on Stanford Dogs, not by evidence on the actual segmentation model.
  • domain assumption Randomly masking 75% of the student input with 0 or 255 proxy pixels mimics real target dynamism such as fog, night, rain, and snow.
    This is the core IMS assumption and is validated only empirically through ablations (Table 6).
  • domain assumption Teacher-student EMA with teacher-generated pseudo labels provides reliable supervision in an online one-pass setting.
    Inherited from CoTTA and standard in the CTTA literature; not independently validated here.
  • ad hoc to paper Warmup on style-transferred Cityscapes images with labels is compatible with the source-free CTTA protocol.
    Table 7 shows label accessibility 'Yes' during warmup, so the OoPk* results are not strictly source-free.
  • ad hoc to paper The low-rank matrix BA can be regularized toward orthogonality via the objective ||(BA)^T(BA) - I||^2.
    With rank r=32 and layer width k much larger than 32, (BA)^T(BA) cannot equal the identity matrix, so the stated objective is unattainable and can at best be a soft partial-isometry proxy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Orthogonal Projection Subspace to Aggregate Online Prior-knowledge for Continual Test-time Adaptation." pith.science (2026). https://pith.science/paper/JSNY3WJF

@misc{pith2026250619022,
  author       = {Pith},
  title        = {Pith review of: Orthogonal Projection Subspace to Aggregate Online Prior-knowledge for Continual Test-time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSNY3WJF}},
  note         = {Machine review of arXiv:2506.19022}
}
read the original abstract

Continual Test Time Adaptation (CTTA) is a task that requires a source pre-trained model to continually adapt to new scenarios with changing target distributions. Existing CTTA methods primarily focus on mitigating the challenges of catastrophic forgetting and error accumulation. Though there have been emerging methods based on forgetting adaptation with parameter-efficient fine-tuning, they still struggle to balance competitive performance and efficient model adaptation, particularly in complex tasks like semantic segmentation. In this paper, to tackle the above issues, we propose a novel pipeline, Orthogonal Projection Subspace to aggregate online Prior-knowledge, dubbed OoPk. Specifically, we first project a tuning subspace orthogonally which allows the model to adapt to new domains while preserving the knowledge integrity of the pre-trained source model to alleviate catastrophic forgetting. Then, we elaborate an online prior-knowledge aggregation strategy that employs an aggressive yet efficient image masking strategy to mimic potential target dynamism, enhancing the student model's domain adaptability. This further gradually ameliorates the teacher model's knowledge, ensuring high-quality pseudo labels and reducing error accumulation. We demonstrate our method with extensive experiments that surpass previous CTTA methods and achieve competitive performances across various continual TTA benchmarks in semantic segmentation tasks.

Figures

Figures reproduced from arXiv: 2506.19022 by the authors.

Figure 1
Figure 1. CoTTA scenario is illustrated above. Bubble chart be [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Principle comparison with CoTTA[68], EcoTTA[64] and BECoTTA[34]. Our method has fewer trainable parameters and does not rely on adding new structures via orthogonal projection. like fog, rain, night, or snow, we empirically observe that continuously changing target distribution often comes with serious occlusion, blur, or darkness, etc. We further elabo￾rate a novel prior-knowledge aggregation strategy to mimic the … view at source ↗
Figure 3
Figure 3. The overall paradigm of our method, OoPk. Given unlabeled data from a new domain, its original and simulated version by our IMS for unseen domains are fed into the teacher and student models, respectively. The teacher model then supervises the student model’s adaptation to the new domain using our proposed orthogonal projection subspace tuning. This method fixes the original model’s weights Ws, introduces orthogonal… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: A toy experiment to demonstrate the motivation of Or [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of our OoPkr=32 (no warmup) with previous SOTA methods on the ACDC dataset. Our method could better segment different pixel-wise classes such as shown in the white box. Best view zoom in and out. Scenarios Daytime − N ight Clear − F oggy Clear − …
Figure 6
Figure 6. Figure 6: Ablations about each component. We validate by [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: A toy experiment to demonstrate the motivation of Or [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Comparisons on ablation for various OPS positions equipped into source model and corresponding tunable parameters counting. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Image Masking Strategy Variants visualizations. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison of our method in foggy weather on ACDC val. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison of our method in night weather on ACDC val. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of our method in rainy weather on ACDC val. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Qualitative comparison of our method in snowy weather on ACDC val. [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Visualization of our synthetic data generation. [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: Visualization of our synthetic data generation. [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of our method on SHIFT continuous val. [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

94 extracted references · 67 canonical work pages

  1. [1]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. InICLR, 2022. 3, 4

    2022

  2. [2]

    In search for a general- izable method for source free domain adaptation

    Malik Boudiaf, Tom Denton, Bart Van Merri¨enboer, Vincent Dumoulin, and Eleni Triantafillou. In search for a general- izable method for source free domain adaptation. InICML, pages 2914–2931. PMLR, 2023. 3

    2023

  3. [3]

    Large scale GAN training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. InICLR, 2019. 3

    2019

  4. [4]

    Angular visual hardness

    Beidi Chen, Weiyang Liu, Zhiding Yu, Jan Kautz, Anshu- mali Shrivastava, Animesh Garg, and Animashree Anandku- mar. Angular visual hardness. InICML, pages 1637–1648. PMLR, 2020. 12

    2020

  5. [5]

    Contrastive test-time adaptation

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. InCVPR, pages 295–305, 2022. 3

    2022

  6. [6]

    Robustnet: Improving domain generalization in urban-scene segmentation via in- stance selective whitening

    Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via in- stance selective whitening. InCVPR, pages 11580–11590,

  7. [7]

    Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes

    Sungha Choi, Seunghan Yang, Seokeon Choi, and Sun- grack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In ECCV, pages 440–458. Springer, 2022. 8

    2022

  8. [8]

    MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark.https : / / github

    MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark.https : / / github . com / open - mmlab/mmsegmentation, 2020. 6

    2020

Show all 94 references
  1. [9]

    Qlora: Efficient finetuning of quantized llms

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. InNeurIPS, 2024. 3

    2024

  2. [10]

    Bert: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. InACL, 2019. 3, 4

    2019

  3. [11]

    Improved regular- ization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017

    Terrance DeVries and Graham W Taylor. Improved regular- ization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017. 3

    2017 arXiv

  4. [12]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. InICLR, 2020. 1

    2020

  5. [13]

    Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity.JMLR, 23(120):1–39, 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity.JMLR, 23(120):1–39, 2022. 2, 3

    2022

  6. [14]

    Decorate the newcomers: Visual domain prompt for continual test time adaptation

    Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595–7603, 2023. 6, 7

    2023

  7. [15]

    Test-time training with masked autoencoders

    Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. InNeurIPS, pages 29374–29385, 2022. 1

    2022

  8. [16]

    Clip-adapter: Better vision-language models with feature adapters.IJCV, 132(2):581–595, 2024

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters.IJCV, 132(2):581–595, 2024. 3

    2024

  9. [17]

    Mimic before reconstruct: Enhancing masked autoencoders with feature mimicking.IJCV, 132(5):1546–1556, 2024

    Peng Gao, Ziyi Lin, Renrui Zhang, Rongyao Fang, Hongyang Li, Hongsheng Li, and Yu Qiao. Mimic before reconstruct: Enhancing masked autoencoders with feature mimicking.IJCV, 132(5):1546–1556, 2024. 4

    2024

  10. [18]

    Vi- sual prompt tuning for test-time domain adaptation.arXiv preprint arXiv:2210.04831, 2022

    Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Vi- sual prompt tuning for test-time domain adaptation.arXiv preprint arXiv:2210.04831, 2022. 3, 6, 7

    2022 arXiv

  11. [19]

    Towards a unified view of parameter-efficient transfer learning

    Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg- Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. InICLR, 2022. 3

    2022

  12. [20]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR, pages 770–778, 2016. 1

    2016

  13. [21]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. InCVPR, pages 9729–9738, 2020

    2020

  14. [22]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. InCVPR, pages 16000–16009, 2022. 1, 3, 4

    2022

  15. [23]

    Milan: Masked image pretraining on language assisted representation.arXiv preprint arXiv:2208.06049,

    Zejiang Hou, Fei Sun, Yen-Kuang Chen, Yuan Xie, and Sun- Yuan Kung. Milan: Masked image pretraining on language assisted representation.arXiv preprint arXiv:2208.06049,

    arXiv

  16. [24]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. InICML, pages 2790–2799. PMLR, 2019. 2, 3

    2019

  17. [25]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3

    2022

  18. [26]

    Densely connected convolutional net- works

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kil- ian Q Weinberger. Densely connected convolutional net- works. InCVPR, pages 4700–4708, 2017. 1

    2017

  19. [27]

    Test-time classifier ad- justment module for model-agnostic domain generalization

    Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier ad- justment module for model-agnostic domain generalization. InNeurIPS, pages 2427–2440, 2021. 1

    2021

  20. [28]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. InECCV, pages 709–727. Springer,

  21. [29]

    TSIT: A simple and versatile framework for image-to-image translation

    Liming Jiang, Changxu Zhang, Mingyang Huang, Chunxiao Liu, Jianping Shi, and Chen Change Loy. TSIT: A simple and versatile framework for image-to-image translation. In ECCV, 2020. 13

    2020

  22. [30]

    Novel dataset for fine-grained image categorization: Stanford dogs

    Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. InCVPR workshop on fine- grained visual categorization (FGVC), 2011. 12

    2011

  23. [31]

    Kingma and J

    D. Kingma and J. Ba. Adam: A method for stochastic opti- mization. InICLR, 2015. 6, 12

    2015

  24. [32]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 14 9

    2009

  25. [33]

    Universal source-free domain adaptation

    Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. InCVPR, pages 4544–4553, 2020. 3

    2020

  26. [34]

    Becotta: Input-dependent online blending of experts for continual test-time adaptation

    Daeun Lee, Jaehong Yoon, and Sung Ju Hwang. Becotta: Input-dependent online blending of experts for continual test-time adaptation. InICML, 2024. 2, 3, 4, 5, 6, 7, 8, 13, 15

    2024

  27. [35]

    The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021. 3

    2021 arXiv

  28. [36]

    Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization

    Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. InCVPR, pages 8300–8311, 2021. 1

    2021

  29. [37]

    Expansion and shrinkage of localization for weakly- supervised semantic segmentation

    Jinlong Li, Zequn Jie, Xu Wang, Xiaolin Wei, and Lin Ma. Expansion and shrinkage of localization for weakly- supervised semantic segmentation. InAdvances in neural information processing systems, pages 16037–16051, 2022. 14

    2022

  30. [38]

    Weakly supervised semantic segmentation via pro- gressive patch learning.IEEE Transactions on multimedia, 25:1686–1699, 2022

    Jinlong Li, Zequn Jie, Xu Wang, Yu Zhou, Xiaolin Wei, and Lin Ma. Weakly supervised semantic segmentation via pro- gressive patch learning.IEEE Transactions on multimedia, 25:1686–1699, 2022

    2022

  31. [39]

    Weakly supervised semantic segmentation via self-supervised destruction learning.Neurocomputing, 561: 126821, 2023

    Jinlong Li, Zequn Jie, Xu Wang, Yu Zhou, Lin Ma, and Jian- min Jiang. Weakly supervised semantic segmentation via self-supervised destruction learning.Neurocomputing, 561: 126821, 2023

    2023

  32. [40]

    Cross-modal and uncertainty-aware agglomeration for open- vocabulary 3d scene understanding

    Jinlong Li, Cristiano Saltori, Fabio Poiesi, and Nicu Sebe. Cross-modal and uncertainty-aware agglomeration for open- vocabulary 3d scene understanding. InCVPR, pages 19390– 19400, 2025. 14

    2025

  33. [41]

    Prefix-tuning: Optimiz- ing continuous prompts for generation.arXiv preprint arXiv:2101.00190, 2021

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimiz- ing continuous prompts for generation.arXiv preprint arXiv:2101.00190, 2021. 3

    2021 arXiv

  34. [42]

    Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation

    Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation. InICML, pages 6028–6039. PMLR, 2020. 3

    2020

  35. [43]

    Vida: Homeostatic visual domain adapter for continual test time adaptation

    Jiaming Liu, Senqiao Yang, Peidong Jia, Renrui Zhang, Ming Lu, Yandong Guo, Wei Xue, and Shanghang Zhang. Vida: Homeostatic visual domain adapter for continual test time adaptation. InICLR, 2023. 2, 3, 5, 13, 15

    2023

  36. [44]

    Continual-mae: Adaptive distribution masked autoencoders for continual test-time adaptation

    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 test-time adaptation. InCVPR, pages 28653– 28663, 2024. 4, 5, 7, 13

    2024

  37. [45]

    Deep hyperspherical learning

    Weiyang Liu, Yan-Ming Zhang, Xingguo Li, Zhiding Yu, Bo Dai, Tuo Zhao, and Le Song. Deep hyperspherical learning. InNeurIPS, 2017. 12

    2017

  38. [46]

    Learning towards minimum hy- perspherical energy

    Weiyang Liu, Rongmei Lin, Zhen Liu, Lixin Liu, Zhiding Yu, Bo Dai, and Le Song. Learning towards minimum hy- perspherical energy. InNeurIPS, 2018. 2

    2018

  39. [47]

    Decoupled net- works

    Weiyang Liu, Zhen Liu, Zhiding Yu, Bo Dai, Rongmei Lin, Yisen Wang, James M Rehg, and Le Song. Decoupled net- works. InCVPR, pages 2771–2779, 2018. 12

    2018

  40. [48]

    Rehg, Liam Paull, Li Xiong, Le Song, and Adrian Weller

    Weiyang Liu, Rongmei Lin, Zhen Liu, James M. Rehg, Liam Paull, Li Xiong, Le Song, and Adrian Weller. Orthogonal over-parameterized training. InCVPR, 2021. 2

    2021

  41. [49]

    Less: Label-efficient and single-stage referring 3d instance segmentation

    Xuexun Liu, Xu Xiaoxu, Jinlong Li, Qiudan Zhang, Xu Wang, Nicu Sebe, Ma Lin, et al. Less: Label-efficient and single-stage referring 3d instance segmentation. InNeurIPS. NeurIPS, 2024. 14

    2024

  42. [50]

    Ttt++: When does self-supervised test-time training fail or thrive? InNeurIPS, pages 21808–21820, 2021

    Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? InNeurIPS, pages 21808–21820, 2021. 1

    2021

  43. [51]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012–10022, 2021. 1

    2021

  44. [52]

    Unsupervised domain adaptation with residual trans- fer networks.NeurIPS, 29, 2016

    Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual trans- fer networks.NeurIPS, 29, 2016. 1

    2016

  45. [53]

    Mmevol: Empowering multimodal large language models with evol-instruct.arXiv preprint arXiv:2409.05840, 2024

    Run Luo, Haonan Zhang, Longze Chen, Ting-En Lin, Xiong Liu, Yuchuan Wu, Min Yang, Minzheng Wang, Pengpeng Zeng, Lianli Gao, et al. Mmevol: Empowering multimodal large language models with evol-instruct.arXiv preprint arXiv:2409.05840, 2024. 14

    2024 arXiv

  46. [54]

    Evaluating prediction-time batch normalization for robust- ness under covariate shift.arXiv preprint arXiv:2006.10963,

    Zachary Nado, Shreyas Padhy, D Sculley, Alexander D’Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robust- ness under covariate shift.arXiv preprint arXiv:2006.10963,

    2006 arXiv

  47. [55]

    Efficient test- time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test- time model adaptation without forgetting. InICML, pages 16888–16905. PMLR, 2022. 8

    2022

  48. [56]

    Towards stable test-time adaptation in dynamic wild world

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. InICLR, 2023. 3, 15

    2023

  49. [57]

    One-step image translation with text-to-image models.arXiv preprint arXiv:2403.12036, 2024

    Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models.arXiv preprint arXiv:2403.12036, 2024. 13

    2024 arXiv

  50. [58]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6

    2019

  51. [59]

    Adapters: A unified library for parameter-efficient and modular trans- fer learning

    Clifton Poth, Hannah Sterz, Indraneil Paul, Sukannya Purkayastha, Leon Engl ¨ander, Timo Imhof, Ivan Vuli ´c, Se- bastian Ruder, Iryna Gurevych, and Jonas Pfeiffer. Adapters: A unified library for parameter-efficient and modular trans- fer learning. InEMNLP, pages 149–160, Sin...

    2023

  52. [60]

    Controlling text-to-image diffusion by orthogo- nal finetuning

    Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Sch¨olkopf. Controlling text-to-image diffusion by orthogo- nal finetuning. InNeurIPS, pages 79320–79362, 2023. 2

    2023

  53. [61]

    Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer. InICLR, 2017. 2, 3 10

    2017

  54. [62]

    Mm-tta: multi-modal test-time adaptation for 3d se- mantic segmentation

    Inkyu Shin, Yi-Hsuan Tsai, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Sparsh Garg, In So Kweon, and Kuk-Jin Yoon. Mm-tta: multi-modal test-time adaptation for 3d se- mantic segmentation. InCVPR, pages 16928–16937, 2022. 3

    2022

  55. [63]

    Cd-tta: Compound domain test-time adaptation for semantic segmentation.arXiv preprint arXiv:2212.08356,

    Junha Song, K Park, Inkyu Shin, Sanghyun Woo, and In So Kweon. Cd-tta: Compound domain test-time adaptation for semantic segmentation.arXiv preprint arXiv:2212.08356,

    arXiv

  56. [64]

    Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization

    Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. InCVPR, pages 11920–11929,

  57. [65]

    Shift: a synthetic driving dataset for continuous multi-task domain adaptation

    Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. InCVPR, pages 21371–21382, 2022. 5, 6

    2022

  58. [66]

    On orthogonality and learning recurrent networks with long term dependencies

    Eugene V orontsov, Chiheb Trabelsi, Samuel Kadoury, and Chris Pal. On orthogonality and learning recurrent networks with long term dependencies. InICML, pages 3570–3578. PMLR, 2017. 3

    2017

  59. [67]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. InICLR, 2021. 1, 3, 6, 7, 8, 15

    2021

  60. [68]

    Con- tinual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Con- tinual test-time domain adaptation. InCVPR, pages 7201– 7211, 2022. 1, 2, 3, 5, 6, 7, 8, 15

    2022

  61. [69]

    Segformer: Simple and ef- ficient design for semantic segmentation with transformers

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and ef- ficient design for semantic segmentation with transformers. InNeurIPS, pages 12077–12090, 2021. 6

    2021

  62. [70]

    Simmim: A simple framework for masked image modeling

    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. InCVPR, pages 9653–9663, 2022. 3, 4

    2022

  63. [71]

    3d weakly supervised semantic segmentation with 2d vision-language guidance

    Xiaoxu Xu, Yitian Yuan, Jinlong Li, Qiudan Zhang, Zequn Jie, Lin Ma, Hao Tang, Nicu Sebe, and Xu Wang. 3d weakly supervised semantic segmentation with 2d vision-language guidance. InECCV, pages 87–104. Springer, 2024. 14

    2024

  64. [72]

    Generalized source-free domain adaptation

    Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Her- ranz, and Shangling Jui. Generalized source-free domain adaptation. InICCV, pages 8978–8987, 2021. 3

    2021

  65. [73]

    Exploring sparse visual prompt for domain adaptive dense prediction

    Senqiao Yang, Jiarui Wu, Jiaming Liu, Xiaoqi Li, Qizhe Zhang, Mingjie Pan, Yulu Gan, Zehui Chen, and Shanghang Zhang. Exploring sparse visual prompt for domain adaptive dense prediction. InAAAI, pages 16334–16342, 2024. 2, 3, 4, 5, 6, 7, 13

    2024

  66. [74]

    Robust test-time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. InCVPR, pages 15922– 15932, 2023. 3

    2023

  67. [75]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. InICCV, pages 6023–6032, 2019. 3

    2019

  68. [76]

    Generalized source- free domain-adaptive segmentation via reliable knowledge propagation

    Qi Zang, Shuang Wang, Dong Zhao, Yang Hu, Dou Quan, Jinlong Li, Nicu Sebe, and Zhun Zhong. Generalized source- free domain-adaptive segmentation via reliable knowledge propagation. InACM MM, pages 5967–5976, 2024. 14

    2024

  69. [77]

    Boosting novel category dis- covery over domains with soft contrastive learning and all in one classifier

    Zelin Zang, Lei Shang, Senqiao Yang, Fei Wang, Baigui Sun, Xuansong Xie, and Stan Z Li. Boosting novel category dis- covery over domains with soft contrastive learning and all in one classifier. InICCV, pages 11858–11867, 2023. 3

    2023

  70. [78]

    S2 transformer for image captioning

    Pengpeng Zeng, Haonan Zhang, Jingkuan Song, and Lianli Gao. S2 transformer for image captioning. InIJCAI, pages 1608–1614, 2022. 14

    2022

  71. [79]

    mixup: Beyond empirical risk minimiza- tion

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion. InICLR, 2018. 3

    2018

  72. [80]

    Mpt: Multi-grained prompt tuning for text-video retrieval

    Haonan Zhang, Pengpeng Zeng, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Mpt: Multi-grained prompt tuning for text-video retrieval. InACM MM, pages 1206–1214, 2024. 14

    2024

  73. [81]

    Tip-adapter: Training-free clip-adapter for better vision- language modeling

    Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision- language modeling. InECCV, 2022. 3

    2022

  74. [82]

    Auxadapt: Stable and efficient test-time adaptation for temporally consistent video semantic segmentation

    Yizhe Zhang, Shubhankar Borse, Hong Cai, and Fatih Porikli. Auxadapt: Stable and efficient test-time adaptation for temporally consistent video semantic segmentation. In WACV, pages 2339–2348, 2022. 3

    2022

  75. [83]

    Fishertune: Fisher- guided robust tuning of vision foundation models for domain generalized segmentation

    Dong Zhao, Jinlong Li, Shuang Wang, Mengyao Wu, Qi Zang, Nicu Sebe, and Zhun Zhong. Fishertune: Fisher- guided robust tuning of vision foundation models for domain generalized segmentation. InCVPR, pages 15043–15054,

  76. [84]

    Random erasing data augmentation

    Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. InAAAI, pages 13001–13008, 2020. 3

    2020

  77. [85]

    Taming sparsely activated transformer with stochastic experts.arXiv preprint arXiv:2110.04260, 2021

    Simiao Zuo, Xiaodong Liu, Jian Jiao, Young Jin Kim, Hany Hassan, Ruofei Zhang, Tuo Zhao, and Jianfeng Gao. Taming sparsely activated transformer with stochastic experts.arXiv preprint arXiv:2110.04260, 2021. 3 11 Orthogonal Projection Subspace to Aggregate Online Prior-knowled...

    2021 arXiv

  78. [86]

    First, based on previous studies [4, 45, 47], the angles of weights in neural networks capture most informative char- acteristics

    Toy Experiment to Clarify Motivation of Or- thogonal Projection Subspace tuning In this section, we make more explanations for the moti- vation for the proposed Orthogonal Projection Subspace. First, based on previous studies [4, 45, 47], the angles of weights in neural networ...

  79. [87]

    Additional Ablations of Orthogonal Projec- tion Subspace tuning Position In this section, we ablate the proposed Orthogonal Projec- tion Subspace tuning (OPS) in various positions given the source pre-trained model, SegFormer B5, including both Attention&FFN, Attention and FFN...

  80. [88]

    As shown in Tab

    Additional Results of 10 Rounds In this section, we provide more comparative results in terms of 10 rounds on the ACDC dataset. As shown in Tab. 8, ourOoPk r=4 can obtain 61.2% and 62.9% mIoU fornowarmup and warmup settings, which surpasses BECoTTA M andBECoTTA ∗ M and present...

  81. [89]

    Warmup Implementations 9.1. Warmup data synthesis Following previous works, like BECoTTA [34], we utilize pre-trained style transformer [29, 57] to first generate po- tential domains using Cityscapes RGB images only. To achieve this objective, we set the candidate domains with...

  82. [90]

    In our main paper, we ablate the experimental results with differ- entgrid sizesand masking ratioα

    Additional Image Masking Strategy Anal- ysis In this section, we visualize our Image Masking Strategy (IMS) to compare various masking implementations. In our main paper, we ablate the experimental results with differ- entgrid sizesand masking ratioα. From Fig. 9, we can see t...

  83. [91]

    We obtain 59.6% mIoU and 59.7% mIoU for reimple-BECoTTAM and BECoTTAM w masking, respectively

    Image Masking Strategy to BECoTTA To validate that our OPS and IMS work cooperatively, we then conduct comparative experiments that employ our proposed IMS to BECoTTA, reimple-BECoTTAM and BECoTTAM wmasking. We obtain 59.6% mIoU and 59.7% mIoU for reimple-BECoTTAM and BECoTTAM...

  84. [92]

    We adopt the same comparative environment with one single Nvidia A6000 48GB

    Training/Inference time compared with BECoTTA While our method requires much fewer tunable parame- ters than BECoTTA, BECoTTA 2.15Mv.s.ours 1.04M, We further conduct training time comparisons, obtain- ing 1.38 hours for BECoTTA M , 1.01 hours for our OoPkr=4 and 1.02 hours for...

  85. [93]

    Our implemen- tal hyper-parameters are shown in Tab: 9

    SHIFT implemental details To conduct experiments on the SHIFT dataset, we first pre- train the source model on discrete datasets. Our implemen- tal hyper-parameters are shown in Tab: 9

  86. [94]

    However, it requires detailed hyperparameters choices, in- cluding the rankrin OPS, the maskinggrid size sand masking ratioαin IMS, and loss weight tuningλin Lorth

    Limitation and Future Works We verify our OoPk demonstrates superior performance with fewer parameters compared to other CTTA baselines. However, it requires detailed hyperparameters choices, in- cluding the rankrin OPS, the maskinggrid size sand masking ratioαin IMS, and loss...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.