Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation

Guangyao Chen; Jintao Tong; Ran Ma; Ruixuan Li; Yixiong Zou; Yuhua Li

REVIEW 4 major objections 7 minor 23 references

Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A residual adapter inserted deep in a frozen pretrained backbone naturally decouples domain-specific from domain-agnostic features, and the navigator built on this principle reaches state-of-the-art cross-domain few-shot segmentation.

desk verdict The DFN/SAM-SVN method delivers real CD-FSS gains, but the 'adapter naturally decouples domains' claim rests on CKA changes without a random-adapter control and a theory that assumes its conclusion. read the letter →

arxiv 2506.07376 v1 pith:OPVZDWIN submitted 2025-06-09 cs.CV cs.AIcs.LG

Jintao Tong , Ran Ma , Yixiong Zou , Guangyao Chen , Yuhua Li , Ruixuan Li This is my paper

classification cs.CVcs.AIcs.LG

keywords cross-domainfew-shotsegmentationsemanticdomaindecouplingadaptermodulesparameter-efficientfine-tuningsharpness-awareminimizationsingularvaluedecompositiongeneralization

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 claims that an adapter module, placed with a residual connection deep inside a frozen pretrained backbone, does more than enable efficient fine-tuning: it naturally absorbs domain-specific information, leaving the downstream encoder-decoder to learn domain-agnostic knowledge. The authors show this through centered kernel alignment similarity measurements and then exploit it in a proposed Domain Feature Navigator (DFN), a single $1\times1$ convolution with a residual skip. To keep the DFN from overfitting source-domain samples, they apply sharpness-aware minimization only to the singular value matrix of the adapter. On the standard cross-domain few-shot segmentation benchmark, this structure-based decoupler outperforms the previous state of the art by 2.69% and 4.68% mean intersection-over-union in 1-shot and 5-shot settings respectively.

What carries the argument

The central object is the Domain Feature Navigator (DFN), a $1\times1$ convolutional adapter with matched input and output channels, attached by residual connection to deep layers of a frozen pretrained backbone and connected across low, middle, and high-level features so that its output is added to the backbone features before the hypercorrelation encoder-decoder. Its companion mechanism is SAM-SVN, which performs singular value decomposition on the adapter weight matrix and applies sharpness-aware minimization only to the diagonal singular value matrix, preserving the adapter's ability to absorb domain information while preventing excessive overfitting to source-domain training samples.

What would settle it

Train the identical model with the DFN left at random initialization, or replaced by a fixed random residual 1-by-1 convolution, while freezing the backbone, and measure the same CKA and MMD shifts and the final MIoU: if a random adapter produces the same drop in backbone domain similarity and the same rise in encoder domain similarity, the natural-decoupler interpretation collapses even if the performance gains remain.

Watch

Extended reading notes

Core claim

The central discovery is that the decoupling of domain information is a structural property of a specific adapter placement, not a learned loss-based effect. When a small adapter is trained from scratch at the output of a fixed, pretrained deep backbone, its output carries the domain-specific components of the representation, while the residual path carries the general features; the subsequent learnable encoder and decoder therefore receive a cleaner domain-agnostic signal. This is evidenced by centered kernel alignment (CKA): the backbone output becomes less similar across domains when the adapter is attached, while the encoder output becomes more similar. The adapter's much smaller capacity, together with the explicit residual identity $\mathbf{I}$ in the gradient path, separates the two learning objectives; the paper formalizes this with an information-bottleneck argument and a gradient algebra in which the adapter approximates the negative projected domain-specific feature.

Load-bearing premise

The load-bearing premise is that the observed similarity changes are caused by the adapter learning domain-specific information; the paper does not rule out that any randomly initialized residual convolution of the same shape produces the same effect, and its appendix derivation assumes this conclusion.

Editorial extensions

If this is right

Adding the DFN alone raises average 1-shot MIoU from 47.57 to 59.89 and 5-shot from 51.38 to 66.59 across the four target datasets.
Adding SAM-SVN further raises the averages to 61.98 (1-shot) and 68.81 (5-shot), exceeding the previous state of the art by 2.69 and 4.68 points respectively.
Domain-similarity measurements indicate the effect is structural: backbone output becomes less domain-similar and encoder output becomes more domain-similar after the adapter is attached.
During target adaptation the backbone is frozen and only the DFN is fine-tuned, so domain adaptation is achieved with a small number of learnable parameters.
Feature-map visualizations show the DFN redirects the model's attention to the target object and its distinctive parts, consistent with a domain-agnostic learning signal.

Reading between the lines

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

A testable consequence the paper does not state: the decoupling should grow with insertion depth, so the same adapter placed at shallow backbone layers should produce little or no CKA separation; the paper's position analysis only samples a few depths.
Because DFN is a 1-by-1 convolution, one pretrained model could serve multiple target domains by storing one fine-tuned DFN per domain and swapping it at inference time; the paper fine-tunes per target but does not discuss this deployment pattern.
The SAM-SVN recipe (singular value decomposition of a small weight matrix plus sharpness minimization on singular values) could be applied to other parameter-efficient modules such as low-rank adapters, where the same overfitting-versus-adaptation trade-off appears.
The paper's own impact statement flags that many-shot scenarios are not covered; extending DFN to many-shot target support sets is an immediate next direction.

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

Two things to know about this paper. First, the method works: DFN plus SAM-SVN gives real, consistent MIoU gains on the CD-FSS benchmark, on both ResNet-50 and ViT, and the ablations are clean. Second, the paper's headline interpretation—that adapters 'naturally' decouple domain information—is not actually established. The evidence is a few CKA numbers and a derivation that assumes the conclusion.

Credit where it's due. The DFN is a simple residual 1x1 conv placed at specific depths, and SAM-SVN restricts sharpness-aware minimization to the singular value matrix, which is a clever way to regularize without killing the adapter's flexibility. The empirical story is convincing as engineering: the gains over prior SOTA are meaningful in this subfield, and the ablation study decomposes the contributions clearly. They also go out of their way to compare fairly with IFA under a batch-size-96 setting, which is a nice touch.

The soft spot is the central claim. The paper infers that adapters 'naturally' store domain-specific information from Table 1: CKA drops at stage-4 and rises at the encoder after adding the adapter. But a randomly initialized 1x1 residual adapter would also change CKA, because CKA is not invariant under arbitrary affine transforms. Without an untrained-adapter baseline, we cannot tell whether the adapter is learning domain structure or just adding a generic perturbation. The theory in Section 6 / Appendix A does not fix this: Step 4 postulates g(f(x)) ≈ -P_spec f(x), and the subsequent steps use that assumption to derive that gradients on the encoder-decoder pass through P_inv. That is an assumption of the conclusion, not a derivation. The paper also gives no error bars on the main tables and no code, which is a real limitation for a paper that is partly making a structural claim about what networks 'naturally' do.

Who is this for? Anyone working on CD-FSS or parameter-efficient transfer will get something from the method and the empirical results. The conceptual framing needs much more work before it should be taken as a general principle. If I were editing, I would send this to peer review because the method is useful and the question is worth asking, but I would expect a major revision: add the random-adapter baseline, fix or drop the theory, and report variance across seeds.

Referee Report

4 major / 7 minor

Summary. This paper studies cross-domain few-shot semantic segmentation (CD-FSS). The authors observe that inserting a 1x1 convolutional adapter with a residual connection into deep layers of a frozen backbone reduces CKA similarity between source and target features at the backbone output while increasing it at the encoder output. They interpret this as the adapter naturally decoupling domain-specific information from domain-invariant information, without any domain-decoupling loss. Based on this interpretation, they propose the Domain Feature Navigator (DFN), a set of residual 1x1 adapters attached at multiple backbone levels, and SAM-SVN, a sharpness-aware minimization variant that perturbs only the singular values of the adapter. On the PATNet benchmark, the method reports state-of-the-art MIoU on FSS-1000, Deepglobe, ISIC, and Chest X-ray under both ResNet-50 and ViT backbones. The paper also includes an information-bottleneck-based derivation intended to explain the decoupling phenomenon, analyses of loss landscape sharpness, MMD, CKA, and qualitative feature visualizations.

Significance. If the central claim holds, the paper offers a structural alternative to loss-based domain decoupling, which could be useful beyond CD-FSS because adapters are widely used. The proposed method is simple, parameter-efficient, and shows large empirical gains over strong baselines on multiple target datasets and two backbone architectures. The paper is also careful in several respects: it reports computational overhead, compares against IFA under that method's specific batch-size setting, and includes ablations over adapter position and structure. However, the main conceptual novelty — that adapters 'naturally' serve as decouplers — is not yet convincingly established, because the empirical evidence lacks a random-adapter control and the theoretical derivation assumes the conclusion. The empirical performance may still be of practical value, but the paper's narrative currently overstates the interpretational support.

major comments (4)

[Section 2.1, Table 1] The central observation — that adding a residual 1x1 adapter lowers stage-4 CKA and raises encoder CKA — is not sufficient to show that the adapter has learned domain-specific information. Because the adapter changes the representation by an affine map A = I + W, and CKA is invariant only under orthogonal transformations (not general linear maps), the same CKA movements could occur for an untrained or randomly initialized adapter that has learned no domain structure at all. The paper does not report a random-adapter baseline, a frozen-adapter control, a null distribution, or seed-level repetitions. Without such a control, the causal attribution to 'the adapter captures domain-specific information' is unsupported. This is load-bearing because the title and the method design both rest on the decoupler interpretation.
[Appendix A, Step 4 / Eq. (21)] The theoretical derivation assumes the conclusion. Eq. (21) postulates g(f(x)) ≈ -P_spec f(x), and Eqs. (22)-(27) then derive that gradients on the encoder-decoder parameters are filtered by P_inv. Since the claim that the adapter decouples domain information is essentially the claim that the adapter learns to apply a negative domain-specific projection, the derivation does not provide independent support; the paper itself labels this a hypothesis. The preceding assertion in Eq. (12), that low adapter capacity implies I(X_spec; Z_adapter) >> I(X_inv; Z_adapter), is also asserted rather than derived. I recommend replacing this with either a derivation from the training dynamics or a falsifiable empirical test, such as probing the trained adapter's transformation to measure how much domain-specific variance it actually absorbs compared with a random adapter.
[Section 4, Tables 6, 7, 14, 15] All main results and ablations are reported as single numbers without variance or statistical significance. CD-FSS evaluation is sensitive to the sampled support-query episodes and training seeds, so the headline gains (2.69% and 4.68% MIoU over APSeg in the 1-shot and 5-shot settings) may be within run-to-run noise. The paper should report mean ± std over at least several seeds or episode resamplings, and ideally paired tests for the ablations in Tables 7 and 14. Table 8, which reports only the fluctuation between five checkpoints, does not substitute for variance in the main evaluation.
[Section 4, Implementation Details] The target-domain fine-tuning protocol uses per-dataset learning rates that differ by orders of magnitude (5e-1 for Deepglobe versus 1e-3 for FSS-1000 and 5e-3 for ISIC/Chest X-ray) with a fixed 50 iterations. It is not stated whether these values were selected using target-domain validation labels, nor whether comparable parameter-efficient fine-tuning was applied to adapter-based baselines such as ABCDFSS or APM. To support a fair comparison, the paper should describe the selection rule for these hyperparameters and the fine-tuning protocol used for baselines.

minor comments (7)

[Abstract and Section 3] In the Abstract, the sentence 'we freeze the model and fine-tune the DFN to learn target-specific knowledge specific' contains a duplicated word; it should read 'target-specific knowledge'.
[Table 2] The columns of Table 2 are garbled in the submitted text: the header labels 'BKB shallower', 'BKB deeper', and 'between enc-dec' are merged, and the CKA values for the encoder are difficult to associate with the correct position. The table should be redrawn so that each insertion position has clearly separated stage-4 and encoder columns.
[Section 3.2, Eqs. (2)-(3)] The notation N_alpha({F_l}) suggests that the same 1x1 convolution is applied to the entire set of feature maps, but the text says DFN is connected to low, middle, and high-level features separately. Please clarify whether there are multiple adapters with shared or separate parameters across the L feature levels.
[Section 3.3, Eqs. (5)-(7)] After the SVD alpha' = U S V^T, the perturbation epsilon in Eq. (7) is added to the singular value matrix S. If epsilon is not constrained to be diagonal, the expression U(S+epsilon)V^T is no longer an SVD and the resulting weight matrix may not be reconstructible as a valid convolution with the stated rank-one factorization. Please clarify how epsilon is parameterized and how Eq. (7) is computed in practice.
[Figure 6] Figure 6 is very dense and the internal structures of DFN and SAM-SVN are hard to read at the printed size. The green-box details should be enlarged or split into a separate diagram.
[Appendix F, Table 12] The right-hand side of Table 12 reports only parameter and FLOPs counts; the 'complexity analysis' would be more informative if it also reported wall-clock training time per iteration for baseline and ours. Also, the row 'increase ratio 21% 0.02%' should be split into two clearly labeled rows for parameters and FLOPs.
[Table 6] In Table 6, the method name 'RestNet' (Huang et al., 2023) is likely a typo for 'ResTNet' or 'Residual Transformation Network'; please verify the spelling against the cited paper.

Circularity Check

1 steps flagged · score 6.0 of 10

Appendix A's 'derivation' postulates the decoupler behavior it then concludes; the theoretical support for the central 'adapter naturally decouples' claim is circular, though the benchmark gains are independent.

self definitional [Appendix A, Step 4 (Eq. 21) and closing paragraph; relied on by Section 6, Eqs. (10)-(16)]
"We hypothesize g(f(x))≈−fspec(x). Using the projection, g(f(x))≈−Pspecf(x). Thus, the Jacobian of the adapter function is ... ∂g(f(x))/∂f(x) ≈ −Pspec ... This derivation shows that, under these assumptions, the gradient updates to the encoder-decoder parameters θf are predominantly guided by the domain-invariant components of the features."

The load-bearing Step 4 asserts exactly the phenomenon to be derived: the adapter represents the negative domain-specific projection. Equations (22)-(27) substitute this assumption into the chain rule and conclude that the encoder-decoder gradient is multiplied by P_inv, i.e., that the adapter filters out domain-specific gradients. That conclusion is the Step 4 postulate restated; no independent mechanism is derived. The paper's Section 6 text states the same conclusion ('the adapter... filters or projects out the domain-specific gradient components that would otherwise update θf') immediately after relying on this Appendix.

full rationale

The paper's headline benchmark results (Tables 6, 7, 9, 14-16) are self-contained: DFN is a residual 1x1 adapter with SAM-SVN, and the mIoU gains are measured against external methods, so the empirical performance claims are not circular. The conceptual claim that the adapter 'naturally serves as a domain information decoupler' is supported by two kinds of evidence: CKA similarity shifts (Tables 1, 2, 4) and the Appendix A theoretical analysis. The Appendix A analysis is circular: Step 4 hypothesizes g(f(x))≈−P_spec f(x), which is the decoupling property itself, and the subsequent equations only propagate that postulate to the encoder-decoder gradient. Thus the formal derivation reduces to its own input. The CKA tables show a real correlational pattern, but the paper reports no control with an untrained/random residual 1x1 convolution, so the tables do not independently establish that the adapter stores domain-specific information rather than generically perturbing features. Self-citations (Zou et al. 2022; Tong et al. 2024b) are used only to motivate the CKA measure and are not load-bearing. Overall, this is partial circularity: the central interpretive derivation assumes its conclusion, while the empirical method and benchmark gains retain independent content.

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

The theoretical section rests on a decomposition of features into invariant and specific subspaces and on the assumption that the adapter learns the negative of the specific projection; these assumptions are not independent of the result being derived. The empirical insight relies on a CKA interpretation that is not controlled. There are three tuned hyperparameters.

free parameters (3)

SAM-SVN perturbation radius rho = 0.5
Set in Implementation Details; controls the strength of the sharpness perturbation and is chosen by hand without sensitivity analysis.
Target-domain fine-tuning learning rates = 1e-3 (FSS-1000), 5e-1 (Deepglobe), 5e-3 (ISIC, Chest X-ray)
Per-dataset tuned (Appendix E, Fig.14); final MIoU values depend on these choices.
Target-domain fine-tuning iterations = 50
Fixed for all target tasks in Section 4; no principled stopping criterion; affects adaptation quality.

assumptions (5)

ad hoc to paper Input features decompose into orthogonal domain-invariant and domain-specific subspaces: f(x) = f_inv(x) + f_spec(x), with projection operators P_inv and P_spec satisfying P_inv + P_spec approximately equals I and P_inv P_spec approximately equals 0.
Appendix A, Eq. 17 and surrounding text. This decomposition is assumed, not derived, and is used to prove the decoupling claim.
ad hoc to paper The adapter with limited capacity tends to absorb primarily the domain-specific signal (Eq.12: I(X_spec;Z_adapter) much greater than I(X_inv;Z_adapter)).
Section 6. This statement restates what the paper is trying to prove rather than deriving it from independent evidence.
ad hoc to paper The adapter approximates the negative of the projected domain-specific representation: g(f(x)) approximately equals -P_spec f(x).
Appendix A, Step 4. This assumption is equivalent to the conclusion that the adapter decouples domain information, making the derivation circular.
domain assumption Deeper layers of the backbone encode more domain-specific semantic information.
Section 2.2.1 and Table 3. Plausible but not proven; used to justify inserting the adapter only into deeper layers.
domain assumption CKA similarity between source and target features is a valid indicator of domain-specific versus domain-agnostic content (lower CKA means more domain-specific).
Section 2.1. No control for random projections is reported, so the interpretation is not uniquely determined.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation." pith.science (2026). https://pith.science/paper/OPVZDWIN

@misc{pith2026250607376,
  author       = {Pith},
  title        = {Pith review of: Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OPVZDWIN}},
  note         = {Machine review of arXiv:2506.07376}
}

read the original abstract

Cross-domain few-shot segmentation (CD-FSS) is proposed to pre-train the model on a source-domain dataset with sufficient samples, and then transfer the model to target-domain datasets where only a few samples are available for efficient fine-tuning. There are majorly two challenges in this task: (1) the domain gap and (2) fine-tuning with scarce data. To solve these challenges, we revisit the adapter-based methods, and discover an intriguing insight not explored in previous works: the adapter not only helps the fine-tuning of downstream tasks but also naturally serves as a domain information decoupler. Then, we delve into this finding for an interpretation, and find the model's inherent structure could lead to a natural decoupling of domain information. Building upon this insight, we propose the Domain Feature Navigator (DFN), which is a structure-based decoupler instead of loss-based ones like current works, to capture domain-specific information, thereby directing the model's attention towards domain-agnostic knowledge. Moreover, to prevent the potential excessive overfitting of DFN during the source-domain training, we further design the SAM-SVN method to constrain DFN from learning sample-specific knowledge. On target domains, we freeze the model and fine-tune the DFN to learn target-specific knowledge specific. Extensive experiments demonstrate that our method surpasses the state-of-the-art method in CD-FSS significantly by 2.69% and 4.68% MIoU in 1-shot and 5-shot scenarios, respectively.

Figures

Figures reproduced from arXiv: 2506.07376 by the authors.

**Figure 3.** Three different positions for adapters. Position baseline(w/o adapter) BKB shallower BKB deeper between enc-dec BKB encoder BKB encoder BKB encoder BKB encoder FSS-1000 0.5371 0.0709 0.4679↓ 0.0737↑ 0.4788↓ 0.0733↑ 0.5371- 0.0695↓ Deepglobe 0.4147 0.0498 0.3736↓ 0.0452↓ 0.3687↓ 0.0679↑ 0.4147- 0.0469↓ ISIC 0.5266 0.0678 0.5028↓ 0.0633↓ 0.4653↓ 0.0724↑ 0.5266- 0.0658↓ Chest X-ray 0.4865 0.0554 0.4639↓ 0.0532↓ 0.4390↓… view at source ↗

**Figure 4.** The visualization results of the feature maps with [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (9 more)

**Figure 5.** Figure 5: Two different design and connection structures. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: Overview of our method in a 1-shot example. After obtaining the pyramid features of support and query images, [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Qualitative results of our model for 1-shot setting. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 9.** Figure 9: Feature map visualization: (row 1) DFN guides [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

**Figure 11.** Figure 11: The deeper the neural network is, the more [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

**Figure 12.** Figure 12: The adapter serves as a natural boundary line for [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]

**Figure 14.** Figure 14: The Learning rate for the fine-tuning of four [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]

**Figure 13.** Figure 13: Sharpness-Aware Minimization (SAM): flattens [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

**Figure 16.** Figure 16: Example of segmentation for four datasets. [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 10 canonical work pages

[1]

Multimodality helps few-shot 3d point cloud semantic segmentation.arXiv preprint arXiv:2410.22489, 2024a

An, Z., Sun, G., Liu, Y ., Li, R., Wu, M., Cheng, M.-M., Konukoglu, E., and Belongie, S. Multimodality helps few-shot 3d point cloud semantic segmentation.arXiv preprint arXiv:2410.22489, 2024a. An, Z., Sun, G., Liu, Y ., Liu, F., Wu, Z., Wang, D., Van Gool, L., and Belongie, S. Rethinking few-shot 3d point cloud semantic segmentation. InProceedings of th...

arXiv
[4]

Deep- globe 2018: A challenge to parse the earth through satel- lite images

Demir, I., Koperski, K., Lindenbaum, D., Pang, G., Huang, J., Basu, S., Hughes, F., Tuia, D., and Raskar, R. Deep- globe 2018: A challenge to parse the earth through satel- lite images. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 172–181,

work page 2018
[9]

J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W

Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

arXiv
[13]

K., Ruder, S., Dehghani, M., and Hender- son, J

Mahabadi, R. K., Ruder, S., Dehghani, M., and Hender- son, J. Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks.arXiv preprint arXiv:2106.04489,

arXiv
[14]

Normalization Layers Are All That Sharpness-Aware Minimization Needs

Mueller, M., Vlaar, T., Rolnick, D., and Hein, M. Normal- ization layers are all that sharpness-aware minimization needs.arXiv preprint arXiv:2306.04226,

work page Pith review arXiv
[15]

One- shot learning for semantic segmentation.arXiv preprint arXiv:1709.03410,

Shaban, A., Bansal, S., Liu, Z., Essa, I., and Boots, B. One- shot learning for semantic segmentation.arXiv preprint arXiv:1709.03410,

arXiv
[16]

Prototype mixture models for few-shot semantic segmentation

Yang, B., Liu, C., Li, B., Jiao, J., and Ye, Q. Prototype mixture models for few-shot semantic segmentation. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII 16, pp. 763–778. Springer,

work page 2020
[17]

Object-contextual rep- resentations for semantic segmentation

Yuan, Y ., Chen, X., and Wang, J. Object-contextual rep- resentations for semantic segmentation. InComputer Vision–ECCV 2020: 16th European Conference, Glas- gow, UK, August 23–28, 2020, Proceedings, Part VI 16, pp. 173–190. Springer,

work page 2020

Show all 23 references

[18]

Flatten long- range loss landscapes for cross-domain few-shot learning, 2024a

Zou, Y ., Liu, Y ., Hu, Y ., Li, Y ., and Li, R. Flatten long- range loss landscapes for cross-domain few-shot learning, 2024a. Zou, Y ., Ma, R., Li, Y ., and Li, R. Attention temperature matters in vit-based cross-domain few-shot learning.Ad- vances in Neural Information Proc...

arXiv
[19]

When domain shift occurs, the weights learned by the DFN become misaligned on the target domain

During source domain training, the DFN absorbs domain-specific informa- tion, leading to loss minima that are specific to the source domain, i.e., source minima. When domain shift occurs, the weights learned by the DFN become misaligned on the target domain. If the loss landsc...

2022
[21]

We employ PASCAL- 5i as our source domain for training

is an extended version of PASCAL VOC 2012 (Everingham et al., 2010), incorporat- ing supplementary annotation enhancement details from the SDS dataset (Hariharan et al., 2011). We employ PASCAL- 5i as our source domain for training. Subsequently, we assess the performance of t...

2012
[22]

As ground-truth labels are only provided in the training set, we rely on the official training dataset, consist- ing of 803 images, to present our results

comprises satellite images with dense pixel-level annotations across 7 categories: ur- ban, agriculture, rangeland, forest, water, barren, and un- known. As ground-truth labels are only provided in the training set, we rely on the official training dataset, consist- ing of 803...

2019
[23]

The dataset is processed and utilized in accordance with the standards set by PATNet

is designed for skin cancer screening, and comprises lesion images, with each image containing precisely one primary lesion. The dataset is processed and utilized in accordance with the standards set by PATNet. And we regard ISIC2018 as a target domain for testing. Chest X-ray...

2013
[96]

Related Work Few-shot learningFew-shot learning focuses on developing robust representations for novel concepts with limited anno- tated samples (An et al., 2024a;b)

I. Related Work Few-shot learningFew-shot learning focuses on developing robust representations for novel concepts with limited anno- tated samples (An et al., 2024a;b). Existing approaches can be broadly categorized into three primary frameworks: met- ric learning methods (Sn...

2017
[2012]

C., Karlinsky, L., Codella, J

Guo, Y ., Codella, N. C., Karlinsky, L., Codella, J. V ., Smith, J. R., Saenko, K., Rosing, T., and Feris, R. A broader study of cross-domain few-shot learning. InComputer vision–ECCV 2020: 16th European conference, glasgow, UK, August 23–28, 2020, proceedings, part XXVII 16, ...

2020
[2013]

Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and Yuille, A. L. Semantic image segmentation with deep convolutional nets and fully connected crfs.arXiv preprint arXiv:1412.7062,

arXiv
[2015]

Domain- invariant feature exploration for domain generalization

Lu, W., Wang, J., Li, H., Chen, Y ., and Xie, X. Domain- invariant feature exploration for domain generalization. arXiv preprint arXiv:2207.12020,

arXiv
[2017]

Sharpness-aware minimization for efficiently improving generalization.arXiv preprint arXiv:2010.01412,

Foret, P., Kleiner, A., Mobahi, H., and Neyshabur, B. Sharpness-aware minimization for efficiently improving generalization.arXiv preprint arXiv:2010.01412,

2010 arXiv
[2018]

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

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

2010 arXiv
[2019]

E., Dusza, S., Gutman, D., Helba, B., Kalloo, A., Liopyris, K., Marchetti, M., et al

Codella, N., Rotemberg, V ., Tschandl, P., Celebi, M. E., Dusza, S., Gutman, D., Helba, B., Kalloo, A., Liopyris, K., Marchetti, M., et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic).arXiv prep...

2018 arXiv
[2020]

and Bruna, J

Garcia, V . and Bruna, J. Few-shot learning with graph neural networks.arXiv preprint arXiv:1711.04043,

arXiv
[2021]

Restnet: Boosting cross- domain few-shot segmentation with residual transforma- tion network.arXiv preprint arXiv:2308.13469,

Huang, X., Zhu, C., and Chen, W. Restnet: Boosting cross- domain few-shot segmentation with residual transforma- tion network.arXiv preprint arXiv:2308.13469,

arXiv
[2022]

Decoupling representation and classifier for long-tailed recognition.arXiv preprint arXiv:1910.09217,

Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., and Kalantidis, Y . Decoupling representation and classifier for long-tailed recognition.arXiv preprint arXiv:1910.09217,

1910 arXiv

Pith tools

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

[1] [1]

Multimodality helps few-shot 3d point cloud semantic segmentation.arXiv preprint arXiv:2410.22489, 2024a

An, Z., Sun, G., Liu, Y ., Li, R., Wu, M., Cheng, M.-M., Konukoglu, E., and Belongie, S. Multimodality helps few-shot 3d point cloud semantic segmentation.arXiv preprint arXiv:2410.22489, 2024a. An, Z., Sun, G., Liu, Y ., Liu, F., Wu, Z., Wang, D., Van Gool, L., and Belongie, S. Rethinking few-shot 3d point cloud semantic segmentation. InProceedings of th...

arXiv

[2] [4]

Deep- globe 2018: A challenge to parse the earth through satel- lite images

Demir, I., Koperski, K., Lindenbaum, D., Pang, G., Huang, J., Basu, S., Hughes, F., Tuia, D., and Raskar, R. Deep- globe 2018: A challenge to parse the earth through satel- lite images. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 172–181,

work page 2018

[3] [9]

J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W

Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

arXiv

[4] [13]

K., Ruder, S., Dehghani, M., and Hender- son, J

Mahabadi, R. K., Ruder, S., Dehghani, M., and Hender- son, J. Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks.arXiv preprint arXiv:2106.04489,

arXiv

[5] [14]

Normalization Layers Are All That Sharpness-Aware Minimization Needs

Mueller, M., Vlaar, T., Rolnick, D., and Hein, M. Normal- ization layers are all that sharpness-aware minimization needs.arXiv preprint arXiv:2306.04226,

work page Pith review arXiv

[6] [15]

One- shot learning for semantic segmentation.arXiv preprint arXiv:1709.03410,

Shaban, A., Bansal, S., Liu, Z., Essa, I., and Boots, B. One- shot learning for semantic segmentation.arXiv preprint arXiv:1709.03410,

arXiv

[7] [16]

Prototype mixture models for few-shot semantic segmentation

Yang, B., Liu, C., Li, B., Jiao, J., and Ye, Q. Prototype mixture models for few-shot semantic segmentation. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII 16, pp. 763–778. Springer,

work page 2020

[8] [17]

Object-contextual rep- resentations for semantic segmentation

Yuan, Y ., Chen, X., and Wang, J. Object-contextual rep- resentations for semantic segmentation. InComputer Vision–ECCV 2020: 16th European Conference, Glas- gow, UK, August 23–28, 2020, Proceedings, Part VI 16, pp. 173–190. Springer,

work page 2020