Pith. sign in

REVIEW 4 major objections 6 minor 36 references

Take a Peek: Efficient Encoder Adaptation for Few-Shot Semantic Segmentation via LoRA

T0 review · 4 major / 6 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 Take a Peek fine-tunes the encoder with LoRA on the support set at test time, improving few-shot segmentation across diverse models and benchmarks.

desk verdict Solid plug-and-play encoder adaptation recipe for FSS; evidence needs error bars and an SVD comparison before the 'consistent gains' claim holds up. read the letter →

arxiv 2512.10521 v3 pith:BBRDQWQB submitted 2025-12-11 cs.CV

Pasquale De Marinis , Gennaro Vessio , Giovanna Castellano This is my paper
classification cs.CV
keywords few-shotsemanticsegmentationLow-RankAdaptationtest-timeencodercross-domainmulti-classparameter-efficientfine-tuningfeature-spaceshift
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 argues that the main bottleneck in few-shot semantic segmentation is not the decoder but the frozen encoder's poor features for novel classes. It proposes Take a Peek (TaP): during inference, briefly fine-tune the encoder on the support set via Low-Rank Adaptation, treating support images as pseudo-queries, then segment the query. TaP is model-agnostic and adds only a small fraction of trainable parameters. Across COCO-20i, Pascal-5i, and three cross-domain datasets (DeepGlobe, ISIC, Chest X-ray), it consistently lifts mIoU over vanilla baselines, with the largest gains in multi-class and strong domain-shift settings. The method works by inducing a support-conditioned feature-space shift in the encoder, and it degrades only when ranks become large enough to trigger catastrophic forgetting.

What carries the argument

Low-Rank Adaptation (LoRA): the encoder's weight matrices are updated by adding a low-rank product W' = W + α·A·B, with A ∈ R^{m×r}, B ∈ R^{r×n} and rank r much smaller than the matrix dimensions. TaP applies this only to attention layers in transformer backbones and to 1×1 convolutions in CNN backbones. The support set is re-organized episodically: each support image is treated as a pseudo-query and predicted from the remaining supports (substitution), providing gradient supervision with a class-balanced focal loss. This keeps adaptation in the same episodic distribution the decoder expects, and the low-rank constraint prevents overfitting and catastrophic forgetting.

What would settle it

Run TaP on an FSS model whose decoder contains normalization layers (e.g., batch or instance norm) whose statistics were fit to the frozen encoder's features; if mean IoU on COCO-20i drops below the vanilla baseline when the encoder is LoRA-adapted, the model-agnostic compatibility claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a frozen pre-trained encoder limits FSS performance more than decoder sophistication, and that a lightweight, test-time LoRA fine-tune of the encoder on the support set can recover much of this gap. Concretely, TaP updates only low-rank matrices in the encoder's attention or pointwise-convolution layers for a few iterations, using each support image as a pseudo-query supervised by its mask. This shifts the encoder's feature distribution toward the novel classes while leaving the decoder untouched. The authors report consistent mIoU improvements across five FSS models on standard benchmarks and a cross-domain model, including up to +20.65 on Chest X-ray, and show tha

Load-bearing premise

The decoder was trained together with a frozen version of the encoder, so TaP assumes that after the encoder's features are shifted by LoRA, the decoder's original mapping still applies; the paper offers no analysis of this compatibility, only empirical results.

Editorial extensions

If this is right

  • If correct, TaP offers a plug-in improvement for any encoder-decoder FSS model, raising mIoU without retraining or decoder changes.
  • Gains scale with the number of shots, meaning richer support sets can be converted into better features, especially in cross-domain scenarios.
  • Because only a few million parameters are trained, the method is feasible in constrained settings and can trade accuracy for latency via rank and iteration count.
  • The observed degradation at high ranks (r ≥ 2^9) identifies an overparameterization boundary, guiding rank selection for stable adaptation.
  • The explicit exclusion of 1-shot settings indicates a practical limitation: with a single support pair, test-time adaptation provides marginal and unstable gains.

Reading between the lines

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

  • If the feature-space shift is the true mechanism, TaP should transfer to other dense prediction tasks that use the same encoder-decoder structure, such as few-shot object detection or instance segmentation, where support-set adaptation of the backbone could similarly help.
  • The method's dependence on decoder compatibility suggests a testable extension: measuring how the LoRA-induced feature distribution shift interacts with decoder normalization layers; a decoder that is sensitive to feature statistics might require adapting those statistics too.
  • Because gains increase with shot count, an obvious extension is to use only a subset of support images during adaptation (as done for DMTNet), potentially enabling faster adaptation in high-shot settings with minimal loss.
  • Since the paper finds 1-shot unstable, a data-augmentation strategy on the single support pair (a direction the authors mention) could be a direct follow-up to make TaP work in the hardest setting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Take a Peek (TaP), a test-time adaptation method for few-shot semantic segmentation (FSS). TaP uses LoRA to fine-tune the frozen encoder on the support set for a few iterations, treating each support image as a pseudo-query and using the corresponding mask as supervision. The decoder is left unchanged, and after adaptation the query is processed with the adapted encoder. The method is evaluated on COCO-20i and Pascal-5i using five FSS models (BAM, DCAMA, FPTrans, HDMNet, Label Anything) and on three cross-domain datasets using DMTNet, with comparisons to decoder-only fine-tuning and a modified AdaptiveFSS baseline. The paper reports consistent mIoU improvements, especially in 2-way and cross-domain settings, and provides rank/iteration ablations and computational cost profiling.

Significance. If the empirical claims hold, TaP is a simple, model-agnostic plug-in that addresses a recognized bottleneck in FSS — the encoder's poor feature extraction for novel classes — without modifying the decoder or requiring large amounts of trainable parameters. The paper includes several strengths: it evaluates across diverse architectures and encoder types (ResNet, Swin, ViT), releases code, compares multiple adapter variants, and reports parameter counts and memory/time. However, the central claim of "consistent" improvement is currently supported only by mean mIoU differences over five runs with no error bars or significance tests, and some results, including the excluded 1-shot setting, show marginal or negative gains. The efficiency claims also need qualification in light of Table 4. The idea is plausible and useful, but the evidence as presented is not yet at the level required for the strength of the claims.

major comments (4)
  1. [Sec. 4, Tables 1–2, App. C] The abstract and Sec. 4 claim that TaP "consistently improves segmentation performance across diverse models and shot settings." The evidence does not currently support the word "consistent": results are reported as means over five runs of 1000 episodes without standard deviations, confidence intervals, or paired significance tests. Table 1 shows negative per-fold deltas (e.g., HDMNet Fold 1 on COCO 1-way 5-shot: -0.84; BAM Fold 1 on Pascal 1-way 5-shot: -0.86), and App. C reports a mean 1-shot improvement of only +0.13 with negative folds and an instability that the authors themselves call "marginal" and "unstable." Given the small absolute gains for some models (FPTrans +0.66 on COCO 1-way 5-shot), sampling noise is a plausible alternative explanation. The authors should report variance/confidence intervals, run paired tests across episodes or folds, or substantially soften the consist
  2. [Sec. 3.2, compatibility assumption] The claim that TaP is compatible with "any FSS model that follows the encoder-decoder paradigm" rests on the unstated assumption that the decoder's learned mapping remains valid when the LoRA-shifted encoder produces features outside the distribution seen during decoder training. This is the central assumption behind the model-agnostic claim, but the paper provides no analysis or control experiment to test it. I would expect at least one of the following: (a) measuring support-set reconstruction (pseudo-query mIoU) before and after adaptation to show the decoder remains aligned; (b) quantifying the feature shift induced by LoRA; or (c) explicitly reporting cases where adaptation hurts and discussing why. Without this, the generality claim is not established.
  3. [Sec. 4.4, Table 4] The text says TaP is "substantially lighter than full fine-tuning," but Table 4 does not consistently support this for memory or time. For HDMNet, TaP uses 29,568 MB versus 27,251 MB for full fine-tuning; for DMTNet, 6,383 MB versus 6,460 MB; and TaP inference times are equal to or longer than full fine-tuning for BAM and DMTNet. TaP is clearly lighter in terms of trainable parameters (e.g., 2.29M vs 51.65M for BAM), but memory and runtime are not universally lower. The efficiency claims should be stated in terms of trainable parameters and the specific architectures where memory/time savings actually occur.
  4. [Sec. 4 and Fig. 2, hyperparameter choice] The method sets T=8 and r=2^6 for all experiments, but Fig. 2 shows that performance peaks at different iterations depending on rank and then declines in many cases. The choice of T=8 is not justified across the different models, datasets, and shot settings used in the paper. Since the adaptation schedule is a key component of the method, the authors should either provide a sensitivity analysis across settings or explain how T can be selected in practice. The current ablation covers only DCAMA on COCO 1-way 5-shot, which is too narrow to support the global choice.
minor comments (6)
  1. [Abstract / Sec. 4.1] The abstract says improvements are shown "across shot settings," but Sec. 4.1 explicitly excludes the 1-shot setting, and App. C reports that gains there are marginal and unstable. The abstract and intro should be aligned with the actual evidence.
  2. [Sec. 4, model references] In the experimental setup, DMTNet is cited as [13], but reference [13] is the ResNet paper. The DMTNet reference should be corrected to the appropriate DMTNet citation (likely [5] or another entry).
  3. [Sec. 3.2, iteration count] The text says T is kept small ("e.g., five iterations"), but the experiments use eight iterations. Please make the description consistent with the reported configuration.
  4. [Supplementary, Eq. (3)] Equation (3) is typeset in a confusing way: the denominator appears as "YtPN+1 / i=0 Yi". This should be written as a standard fraction, and the index range should be clarified (N+1 classes including background implies sum over i=0..N, not N+1).
  5. [Table 5 caption] The caption mentions "green/gray values" for gains/losses, but the table does not appear to use colors in the provided text. Please clarify the visual encoding.
  6. [Sec. 2.3] Having a subsection titled "Proposed Approach" inside Related Work is structurally odd. Consider moving this to a separate section or folding it into the introduction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported gains are direct empirical comparisons; no fitted parameter or self-citation chain is used to manufacture a prediction.

full rationale

The paper's central contribution is an empirical method, not a derivation: TaP applies LoRA to the encoder using the support set as supervision (Sec. 3.2, Eq. 1) and then reports mean IoU deltas against vanilla versions of six published FSS models on standard benchmarks. There is no step where a quantity is fitted to the target metric and then relabeled as a prediction; the main hyperparameters (rank r=2^6, iterations T=8) are set before the comparisons and are themselves ablated. The LoRA equation is the standard external definition from Hu et al., and the loss in the supplementary (Eq. 2-3) is a standard weighted Focal Loss, so no first-principles result is equivalent to its inputs by construction. The strongest candidate for concern is the claim in Sec. 3.2 that adapting only the encoder 'is compatible with any FSS model that follows the encoder-decoder paradigm'; this is an architectural assumption about decoder robustness, and the supplementary 1-shot results are admittedly 'marginal' and 'unstable,' but these are evidence-quality and generalization-risk issues, not circularity. The only self-citation is the authors' own Label Anything model [23], used as one of six tested base models; since the same positive pattern is reported on five externally developed models, that self-citation is not load-bearing. No uniqueness theorem, no prior ansatz, and no fitted baseline is imported from the authors' earlier work. The paper is therefore self-contained as an empirical evaluation, and no circular step meeting the quote-and-reduction requirement was found.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or architectural entities; LoRA adapters are existing tools. The main burden is on hyperparameters (rank, iterations, learning rate) and on the assumption that a decoder trained with a frozen encoder can tolerate feature shifts from test-time adaptation.

free parameters (5)
  • LoRA rank r = 2^6 = 64 for main results; ablation in Fig. 2
    The rank controls adapter capacity. Fig. 2 shows performance varies strongly with rank and degrades for r=2^9 and r=2^10; the chosen value is selected from the same COCO benchmark.
  • Adaptation iterations T = 8 per episode in experiments (Sec. 4); 'five iterations' mentioned in Sec. 3.2
    T is a hyperparameter chosen by the authors; the inconsistency between text and experiments makes the reported value ambiguous, but the main results use 8.
  • Learning rate = 1e-3 for DCAMA and Label Anything, 1e-4 for BAM, HDMNet, DMTNet
    Per-model learning rates were chosen for the adaptation; no search procedure is described.
  • Focal loss focusing parameter gamma = 2
    Set to 2 following Lin et al. [21]; directly shapes the gradient magnitude during adaptation.
  • Class-weight log smoothing constant = 1.1
    Ad hoc constant in Eq. 3 of the supplement used to stabilize inverse-frequency class weights.
assumptions (4)
  • domain assumption Support-set gradients, computed by treating support images as pseudo-queries, provide a sufficient training signal to adapt the encoder for the actual query image.
    Stated in Sec. 3.2: the substitution approach assumes the support set is representative enough that adapting to it transfers to the query. This is load-bearing and untested theoretically.
  • ad hoc to paper The decoder remains valid when the encoder's features are shifted by LoRA adaptations.
    Sec. 3.2 claims compatibility with any encoder-decoder FSS pipeline, but the decoder was trained on the original frozen encoder's feature distribution; no analysis of feature-space shift is provided.
  • domain assumption Pre-trained ImageNet encoders provide a suitable initialization for low-rank test-time adaptation.
    The method inherits the quality of the backbone; if the backbone is not well pre-trained, the LoRA shift will not help.
  • domain assumption Standard FSS evaluation protocols (fixed splits, 1000 episodes, 5 runs) are treated as ground truth for measuring improvement.
    The paper relies on the benchmark conventions of COCO-20i, Pascal-5i, and CD-FSS datasets without questioning their representativeness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Take a Peek: Efficient Encoder Adaptation for Few-Shot Semantic Segmentation via LoRA." pith.science (2026). https://pith.science/paper/BBRDQWQB

@misc{pith2026251210521,
  author       = {Pith},
  title        = {Pith review of: Take a Peek: Efficient Encoder Adaptation for Few-Shot Semantic Segmentation via LoRA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BBRDQWQB}},
  note         = {Machine review of arXiv:2512.10521}
}
abstract

Few-shot semantic segmentation (FSS) aims to segment novel classes in query images using only a small annotated support set. While prior research has mainly focused on improving decoders, the encoder's limited ability to extract meaningful features for unseen classes remains a key bottleneck. In this work, we introduce \textit{Take a Peek} (TaP), a simple yet effective method that enhances encoder adaptability for both FSS and cross-domain FSS \rev{by inducing a lightweight \textit{feature-space shift} conditioned on the support set}. TaP leverages Low-Rank Adaptation to fine-tune the encoder on the support set with minimal computational overhead, enabling fast adaptation to novel classes while mitigating catastrophic forgetting. Our method is model-agnostic and can be seamlessly integrated into existing FSS pipelines. Extensive experiments across multiple benchmarks--including COCO $20^i$, Pascal $5^i$, and cross-domain datasets such as DeepGlobe, ISIC, and Chest X-ray--demonstrate that TaP consistently improves segmentation performance across diverse models and shot settings. Notably, TaP delivers significant gains in complex multi-class scenarios, highlighting its practical effectiveness in realistic settings. A rank sensitivity analysis also shows that strong performance can be achieved even with low-rank adaptations, thereby ensuring computational efficiency. By addressing a critical limitation in FSS--the encoder's generalization to novel classes--TaP paves the way toward more robust, efficient, and generalizable segmentation systems. The code is available at https://github.com/pasqualedem/TakeAPeek.

Figures

Figures reproduced from arXiv: 2512.10521 by the authors.

Figure 1
Figure 1. Overview of Take a Peek. Given a support set containing examples of novel classes, the encoder is fine-tuned to improve its [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Impact of LoRA rank on model performance over adap [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Top: t-SNE visualizations of the query image and support images’ class-specific feature embeddings across different stages of the TaP adaptation process. Middle: Segmentation predictions at each adaptation step, showing progressive refinement. Bottom: Input data with overlaid ground truth masks (blue for train, red for sofa); from left to right: query image, support images for the train class (shots 1 and 2), and su… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance comparison of different adaptation strate [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Performance (mIoU %) on COCO-20i using DCAMA in 1-shot and 2-shot settings over different folds and iterations. main experiments. Future research could investigate alter￾native strategies, such as leveraging data augmentation on the support image to enrich the support …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 7 linked inside Pith

  1. [1]

    Lung Segmentation in Chest Radiographs Using Anatomical Atlases With Nonrigid Registration. 5

  2. [2]

    A Novel Benchmark for Few-Shot Semantic Seg- mentation in the Era of Foundation Models.Transactions on Machine Learning Research, 2025

    Reda Bensaid, Vincent Gripon, Franc ¸ois Leduc-Primeau, Lukas Mauch, Ghouthi BOUKLI Hacene, and Fabien Car- dinaux. A Novel Benchmark for Few-Shot Semantic Seg- mentation in the Era of Foundation Models.Transactions on Machine Learning Research, 2025. 2

    2025

  3. [3]

    Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-Shot Seg- mentation Without Meta-Learning: A Good Transductive In- ference Is All You Need? In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13974–13983, Nashville, TN, USA, 2021. IEEE. 2

    2021

  4. [4]

    Pixel Matching Network for Cross-Domain Few-Shot Segmentation

    Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel Matching Network for Cross-Domain Few-Shot Segmentation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 978–987, 2024. 2

    2024

  5. [5]

    Cross-Domain Few- Shot Semantic Segmentation via Doubly Matching Trans- formation

    Jiayi Chen, Rong Quan, and Jie Qin. Cross-Domain Few- Shot Semantic Segmentation via Doubly Matching Trans- formation. InProceedings of the Thirty-ThirdInternational Joint Conference on Artificial Intelligence, pages 641–649, Jeju, South Korea, 2024. International Joint Conferences on Artificial Intelligence Organization. 1, 2

    2024

  6. [6]

    A Transformer-Based Adaptive Prototype Matching Network for Few-Shot Semantic Segmentation

    Sihan Chen, Yadang Chen, Yuhui Zheng, Zhi-Xin Yang, and Enhua Wu. A Transformer-Based Adaptive Prototype Matching Network for Few-Shot Semantic Segmentation. In Proceedings of the Thirty-ThirdInternational Joint Confer- ence on Artificial Intelligence, pages 659–667, Jeju, South Korea, 2024. International Joint Conferences on Artificial In- telligence Org...

    2024

  7. [7]

    Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, Harald Kittler, and Allan Halpern

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M. Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, Harald Kittler, and Allan Halpern. Skin Le- sion Analysis Toward Melanoma Detection 2018: A Chal- lenge Hosted by the International Skin Imaging Collabora- tion (ISIC), 2019. arXiv:1902.03368 [cs]. 5

    arXiv 2018

  8. [8]

    DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images

    Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 172–17209, Salt Lake City, UT, USA, 2018. IEEE. 4

    2018

Show all 36 references
  1. [9]

    Few-shot semantic segmen- tation with prototype learning

    Nanqing Dong and Eric P Xing. Few-shot semantic segmen- tation with prototype learning. InBMVC, 2018. 1, 2 7

    2018

  2. [11]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    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.arXiv preprint ar...

    2010 arXiv

  3. [12]

    CLIP-Driven Prototype Net- work for Few-Shot Semantic Segmentation.Entropy, 25(9): 1353, 2023

    Shi-Cheng Guo, Shang-Kun Liu, Jing-Yu Wang, Wei-Min Zheng, and Cheng-Yu Jiang. CLIP-Driven Prototype Net- work for Few-Shot Semantic Segmentation.Entropy, 25(9): 1353, 2023. 2

    2023

  4. [13]

    Deep Residual Learning for Image Recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. pages 770– 778, 2016. 1, 3

    2016

  5. [14]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    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,

  6. [15]

    Fed- Para: Low-rank Hadamard Product for Communication- Efficient Federated Learning

    Nam Hyeon-Woo, Moon Ye-Bin, and Tae-Hyun Oh. Fed- Para: Low-rank Hadamard Product for Communication- Efficient Federated Learning. 2021. 1

    2021

  7. [16]

    Im- ageNet Classification with Deep Convolutional Neural Net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Im- ageNet Classification with Deep Convolutional Neural Net- works. InAdvances in Neural Information Processing Sys- tems. Curran Associates, Inc., 2012. 3

    2012

  8. [17]

    Learning What Not to Segment: A New Perspective on Few- Shot Segmentation

    Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning What Not to Segment: A New Perspective on Few- Shot Segmentation. In2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 8047– 8057, New Orleans, LA, USA, 2022. IEEE. 1, 3

    2022

  9. [18]

    Cross-Domain Few-Shot Semantic Segmentation

    Shuo Lei, Xuchao Zhang, Jianfeng He, Fanglan Chen, Bowen Du, and Chang-Tien Lu. Cross-Domain Few-Shot Semantic Segmentation. InComputer Vision – ECCV 2022, pages 73–90. Springer Nature Switzerland, Cham, 2022. Se- ries Title: Lecture Notes in Computer Science. 2

    2022

  10. [19]

    Label-Efficient Few-Shot Seman- tic Segmentation with Unsupervised Meta-Training.Pro- ceedings of the AAAI Conference on Artificial Intelligence, 38(4):3109–3117, 2024

    Jianwu Li, Kaiyue Shi, Guo-Sen Xie, Xiaofeng Liu, Jian Zhang, and Tianfei Zhou. Label-Efficient Few-Shot Seman- tic Segmentation with Unsupervised Meta-Training.Pro- ceedings of the AAAI Conference on Artificial Intelligence, 38(4):3109–3117, 2024. Number: 4. 2

    2024

  11. [20]

    Lawrence Zitnick, and Piotr Doll ´ar

    Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Doll ´ar. Microsoft COCO: Common Objects in Context, 2015. arXiv:1405.0312 [cs]. 3

    2015 arXiv

  12. [21]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection. InPro- ceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. 3, 4

    2017

  13. [22]

    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,

  14. [23]

    Label Anything: Multi-Class Few- Shot Semantic Segmentation with Visual Prompts, 2024

    Pasquale De Marinis, Nicola Fanelli, Raffaele Scaringi, Emanuele Colonna, Giuseppe Fiameni, Gennaro Vessio, and Giovanna Castellano. Label Anything: Multi-Class Few- Shot Semantic Segmentation with Visual Prompts, 2024. arXiv:2407.02075. 1, 2

    2024 arXiv

  15. [24]

    arXiv:2103.14030 [cs]. 3

    arXiv

  16. [25]

    Hierarchical Dense Correlation Distillation for Few-Shot Segmentation

    Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical Dense Correlation Distillation for Few-Shot Segmentation. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 23641–23651, Vancouver, BC, ...

    2023

  17. [26]

    Hypercorrela- tion squeeze for few-shot segmentation

    Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrela- tion squeeze for few-shot segmentation. InProceedings of the IEEE/CVF international conference on computer vision, pages 6941–6952, 2021. 2

    2021

  18. [27]

    Dense cross-query-and-support attention weighted mask aggrega- tion for few-shot segmentation

    Xinyu Shi, Dong Wei, Yu Zhang, Donghuan Lu, Munan Ning, Jiashun Chen, Kai Ma, and Yefeng Zheng. Dense cross-query-and-support attention weighted mask aggrega- tion for few-shot segmentation. InEuropean Conference on Computer Vision, pages 151–168. Springer, 2022. 1, 2

    2022

  19. [28]

    One-Shot Learning for Semantic Segmenta- tion.British Machine Vision Conference (BMVC), 2017

    Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-Shot Learning for Semantic Segmenta- tion.British Machine Vision Conference (BMVC), 2017. 1

    2017

  20. [29]

    Visualizing data using t-SNE.Journal of machine learning research, 9 (11), 2008

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE.Journal of machine learning research, 9 (11), 2008. 5

    2008

  21. [30]

    Singular Value Fine-tuning: Few- shot Segmentation requires Few-parameters Fine-tuning

    Yanpeng Sun, Qiang Chen, Xiangyu He, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Jian Cheng, Zechao Li, and Jingdong Wang. Singular Value Fine-tuning: Few- shot Segmentation requires Few-parameters Fine-tuning. 2

  22. [31]

    Panet: Few-shot image semantic seg- mentation with prototype alignment

    Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic seg- mentation with prototype alignment. Inproceedings of the IEEE/CVF international conference on computer vision, pages 9197–9206, 2019. 1, 2

    2019

  23. [32]

    Adaptive FSS: A Novel Few- Shot Segmentation Framework via Prototype Enhancement,

    Jing Wang, Jinagyun Li, Chen Chen, Yisi Zhang, Haoran Shen, and Tianxiang Zhang. Adaptive FSS: A Novel Few- Shot Segmentation Framework via Prototype Enhancement,

  24. [33]

    Feature- proxy transformer for few-shot segmentation.Advances in Neural Information Processing Systems, 35:6575–6588,

    Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Feature- proxy transformer for few-shot segmentation.Advances in Neural Information Processing Systems, 35:6575–6588,

  25. [34]

    MFNet: Multiclass Few-Shot Segmentation Network With Pixel-Wise Metric Learning.IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8586–8598, 2022

    Miao Zhang, Miaojing Shi, and Li Li. MFNet: Multiclass Few-Shot Segmentation Network With Pixel-Wise Metric Learning.IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8586–8598, 2022. 1, 2 8 Supplementary Material A. Loss Function We use the Focal Loss as ...

    2022

  26. [35]

    Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard B. W. Yang, Giyeong Oh, and Yanmin Gong. Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation. 2023. 1

    2023

  27. [2021]

    arXiv:2106.09685. 1, 3

    arXiv

  28. [2024]

    arXiv:2312.15731 [cs]. 2, 3, 1

    arXiv

Pith tools

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