Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Probabilistic Prototype Calibration of Vision-Language Models for Generalized Few-shot Semantic Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FewCLIP claims that modeling calibration prototypes as probability distributions, rather than fixed vectors, lets a frozen CLIP segmenter register novel classes from one or five annotated images while keeping base-class performance nearly…

desk verdict Strong GFSS method with real gains, but the probabilistic benefit is confounded with ensembling and needs an M=1 ablation to back the abstract's claim. read the letter →

arxiv 2506.22979 v1 pith:QV5VQSF3 submitted 2025-06-28 cs.CV

Jie Liu , Jiayi Shen , Pan Zhou , Jan-Jakob Sonke , Efstratios Gavves This is my paper
classification cs.CV
keywords generalizedfew-shotsemanticsegmentationprobabilisticprototypecalibrationvision-languagemodelsCLIPclass-incrementaluncertainty-awarelearningmulti-modalprototypes
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

Generalized few-shot semantic segmentation asks a model to keep segmenting the classes it was trained on while also picking up new classes from one to five annotated images. FewCLIP's proposal is to keep a pretrained CLIP frozen and, for each class, calibrate its text prototype with a learnable visual prototype that is sampled from a probability distribution instead of being a fixed vector. The intended effect is that scarce novel-class data shapes a distribution of plausible prototype corrections rather than a single overfitted one, so new classes can be registered by optimizing only their calibration prototypes while base classes are untouched. The paper reports roughly ten percentage-point gains in novel-class mIoU over the strongest existing methods on PASCAL-5i and COCO-20i, and shows the same mechanism carries over to class-incremental few-shot segmentation.

What carries the argument

The load-bearing component is the probabilistic multi-modal encoder, which turns each class-name text prototype and the image's class token into a class-conditional Gaussian via a multi-head cross-attention layer and two MLPs that output a mean and a variance vector. Calibration prototypes are obtained by reparameterized sampling from this distribution, added to a deterministic learnable prototype, and used as $M$ different calibrations whose mask predictions are averaged pixel-wise. This distribution-shaped correction is what is supposed to convert scarce supervision into uncertainty-aware adaptation instead of a single overfitted vector, with a KL term to a standard Gaussian prior keeping the distributions well structured.

What would settle it

A decisive test would register a novel class whose appearance overlaps heavily with a base class while holding the decoder fixed and removing CLIP's textual separation, and check whether novel-class mIoU collapses; a second test is ablating the KL regularization and setting the number of sampled prototypes $M=1$, which would reveal how much of the gain truly comes from probabilistic sampling.

Watch

Extended reading notes

Core claim

The central claim is that probabilistic prototype calibration alone, without retraining CLIP or the mask decoder, is enough to adapt a vision-language model to generalized few-shot semantic segmentation. For each class, the frozen text encoder produces a textual prototype $P_t$, the frozen image encoder produces a class token and patch embeddings, and a lightweight probabilistic encoder maps the pair to a Gaussian distribution over calibration prototypes. Sampled prototypes are combined with a learnable deterministic prototype, and the result calibrates the text prototype through $P_t \odot P_c + P_t$ before the mask decoder sees it. During base-class training the visual prompts, probabilistic encoder, calibration prototypes, background prototype, and mask decoder are learned; during novel-class registration everything is frozen except the new class's calibration prototypes. The paper argues this decoupling is what lets novel classes be added without degrading base classes, and supports the argument with experiments on PASCAL-5i and COCO-20i in both the 1-shot and 5-shot settings.

Load-bearing premise

The load-bearing assumption is that a mask decoder trained only on base classes, with all parameters frozen during novel-class registration, can correctly segment novel classes once only their calibration prototypes are inserted.

Editorial extensions

If this is right

  • Novel classes are registered by optimizing only their calibration prototypes; CLIP encoders, the mask decoder, base-class prototypes, and the background prototype all stay frozen.
  • In the 1-shot setting, FewCLIP reports roughly 10-point hIoU gains over the strongest prior methods on both PASCAL-5i and COCO-20i.
  • In the class-incremental setting, adding a session reduces to concatenating new text and calibration prototypes, and the reported hIoU stays well above prior methods in every session.
  • The ablation shows the calibration formula $P_t \odot P_c + P_t$ outperforms addition, subtraction, dot product, and concatenation alternatives.
  • Both visual and textual information are needed during novel-class registration; ablations with only images or only class names are clearly worse.

Reading between the lines

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

  • An implication the paper leaves implicit is that the frozen-decoder-plus-calibration recipe could generalize to open-vocabulary segmentation, where the bottleneck is also registering unseen classes without retraining; the paper's own limitation note flags domain shift as the risk.
  • One testable extension is replacing the Gaussian with a heavier-tailed prior or making the KL weight class-dependent, which would clarify whether the gain comes from regularization strength or from capturing true intra-class variation.
  • The same calibration mechanism could plausibly transfer to other dense prediction tasks that use CLIP text embeddings, such as referring segmentation or open-vocabulary detection, because it only needs a class-name prototype and a few annotated examples.
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 / 5 minor

Summary. The paper proposes FewCLIP, a framework for generalized few-shot semantic segmentation (GFSS) that calibrates frozen CLIP textual prototypes with learnable visual calibration prototypes, then models these calibration prototypes probabilistically. In the novel-class registration phase, only the novel calibration prototypes are updated, with all other parameters frozen. The method is evaluated on PASCAL-5i and COCO-20i under GFSS and class-incremental (CIFSS) settings, reporting substantial improvements over prior methods such as VPT and BCM, and the authors provide code. The central claimed contribution is that probabilistic prototype calibration provides uncertainty-aware, regularized prototype learning that mitigates overfitting to scarce novel-class data.

Significance. If the probabilistic calibration claim is valid, the paper makes a useful contribution: it introduces uncertainty-aware prototype learning into few-shot segmentation with a parameter-efficient, frozen-CLIP design, and the reported leaderboard gains over strong baselines are large (around 10 hIoU points over VPT in the 1-shot setting). The paper also demonstrates efficiency advantages (fewer learnable parameters, higher FPS) and provides code. However, the key mechanistic claim is currently confounded with ensembling, as detailed below, so the significance of the specific probabilistic contribution is not yet established. The empirical leaderboard result could hold even if the probabilistic mechanism is not the source of the gain.

major comments (4)
  1. [Sec. 4.3, Table 2, Fig. 5] The central claim that probabilistic prototype calibration drives the reported gains is not isolated. The Probabilistic PC variant differs from Deterministic PC in two ways at once: it samples M latent prototypes and averages M mask predictions ('we adopt pixel-wise mean of these predictions', Sec. 4.3), whereas Deterministic PC uses a single prototype. Fig. 5 shows performance increasing with M, which is the standard signature of ensembling or Monte Carlo averaging. The observed ~1.2 hIoU gain in Table 2 (1-shot and 5-shot) is therefore consistent with test-time averaging alone, and the KL term and distributional model may contribute nothing. Please add (i) Probabilistic PC with M=1 (no averaging), (ii) Probabilistic PC with the KL term removed at both M=1 and the chosen M, and (iii) an appropriate deterministic control with equivalent aggregation if meaningful. Without these ablations, the 'probabilistic prototype calibration framework' claimed in the abstract is unsupported as the mechanism behind the improvements.
  2. [Appendix Tables 7, 8, 10, 11] The appendix reports results for 'PPCN' and 'DPCN' without defining these names or relating them to FewCLIP as described in the main paper. For example, Table 11 reports DPCN in the 5-shot CIFSS setting, and Table 10 reports PPCN-Res50 and PPCN-ViT. If DPCN and PPCN are the deterministic and probabilistic variants of FewCLIP, this should be stated explicitly; if they are different models, the appendix numbers cannot be attributed to the method described in Sec. 4. The notation must be unified so the reader can map each table to the exact configuration of the model.
  3. [Sec. 4.3, Eqs. (6)-(8)] The probabilistic model is ambiguous. Eq. (6) defines a per-class Gaussian p(z|P_i^t, g), while Eq. (8) defines a mixture over all classes, and the text then says 'we sample M latent prototypes for each class' from p(z|Pt, g). If the per-class component is used, Eq. (8) is unused; if the mixture is sampled, then for a given class the model can draw from another class's component. The paper should specify the exact sampling distribution per class, or justify why cross-class components are desirable. This is central to what the probabilistic multi-modal encoder computes and to the interpretation of the ablation results.
  4. [Table 2, Appendix Tables 7-8] No error bars or standard deviations are reported. The probabilistic-versus-deterministic difference in Table 2 is small (about 1.2 hIoU in both settings), while per-fold results in Tables 7-8 vary substantially (e.g., PPCN 1-shot hIoU ranges from 56.40 to 69.50 on PASCAL-5i). Without per-fold variance or repeated-seed error bars, the reader cannot assess whether the 1.2-point gain is statistically reliable. Please report mean +/- std over folds (or seeds) at least for the Table 2 ablation and ideally for the main comparisons against VPT and O3S.
minor comments (5)
  1. [Sec. 3] There are several typos and grammar issues, e.g., 'toimproving ... whilemaintaining' in the GFSS preliminaries paragraph, and 'sesson' for 'session' in Table 3 and Table 11 captions. Please proofread the manuscript.
  2. [Fig. 3] The notation 'w/o Z!' and 'w/ Z!' in the figure is unclear; please define what Z! denotes (presumably the sampled latent prototypes) and what the red rectangles indicate in the caption.
  3. [Table 2, Table 6] The 'w/o PC' and 'w/o FT' rows report identical values for 1-shot and 5-shot, which is presumably because there is no novel-class registration in these baselines. This should be stated explicitly so the reader does not mistake it for a repeated result or an error.
  4. [Sec. 4.4, Eq. (10)] The KL balancing weight lambda is set to 0.001 with no sensitivity analysis. Please include an ablation over lambda or at least a brief justification for this value.
  5. [Sec. 4.5, Eq. (11)] In the class-incremental extension, it is not specified how the mixture in Eq. (8) is updated when new sessions are added, given that previous calibration prototypes are frozen. Please clarify whether the mixture components, the sampling procedure, or both change when new classes arrive.

Circularity Check

0 steps flagged · score 1.0 of 10

No derivation-level circularity: the central claim is an external benchmark comparison, though the probabilistic-vs-deterministic ablation is confounded with ensembling and some design choices are selected on the same evaluation folds.

full rationale

The paper's central contribution is an empirical architecture, not a formal derivation: FewCLIP's performance is measured against external baselines in Tables 1 and 3 on PASCAL-5i and COCO-20i, and no equation in Sec. 4 reduces to its own input. Equations 4-5 define a prototype-calibration layer, and Eqs. 6-10 define a conditional VAE-style objective with a KL regularizer; these are architectural assumptions, not circular predictions. The KL term is standard and parameter-free apart from the chosen weight lambda=0.001, and the paper does not invoke a uniqueness theorem or use a load-bearing self-citation: references [13,19] support the KL regularization, while self-citations [23,24] are not used to justify the core claim. The main caveats are concerns about evidence, not circularity. First, the 'Probabilistic PC' row in Table 2 differs from 'Deterministic PC' by both latent sampling and pixel-wise averaging over M mask predictions (Sec. 4.3: 'we adopt pixel-wise mean of these predictions as final mask'), and Fig. 5 shows performance varies with M, so the reported gain is confounded with ensembling and an M=1 ablation is missing. Second, the calibration format in Table 4 and the number of latent prototypes M in Fig. 5 are selected on the same evaluation folds that are later reported as the main results, so those choices are not independently predicted; this is model-selection bias rather than equation-level circularity. The appendix labels the method PPCN/DPCN in Tables 7/8/10/11 without explaining the relationship to FewCLIP, but that is a reporting gap. Overall, the central leaderboard claim is externally falsifiable and does not reduce to the paper's own assumptions; score 1 reflects the confounded ablation rather than derivation circularity.

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

The method rests on the frozen CLIP encoders and a frozen mask decoder, on a Gaussian assumption for the calibration distribution, and on several hyperparameters chosen without sensitivity analysis. The model itself introduces no new physical entities.

free parameters (3)
  • KL regularization weight lambda = 0.001
    Set during training with no sensitivity analysis; it controls the strength of the distribution regularization, one of the two proposed components.
  • Number of sampled latent prototypes M = not reported in main text (see Fig. 5)
    Ablation shows performance varies with M; without the chosen value, exact reproduction is impossible.
  • Prototype calibration format = Pt*Pc+Pt
    Selected among six formats in Table 4 based on mIoU on the same benchmarks used for final evaluation, a post-hoc model selection.
assumptions (3)
  • domain assumption Frozen CLIP image and text encoders provide sufficiently discriminative per-pixel features for segmentation when combined with calibrated prototypes.
    The method freezes both encoders and only learns prototypes and a mask decoder; if CLIP's patch features are not aligned with the segmentation task, the approach fails. Invoked in Section 4.1-4.2.
  • domain assumption The mask decoder trained on base classes generalizes to an extended label space that includes novel classes without retraining.
    During novel registration, all base-trained parameters including the mask decoder are frozen (Section 4.4); the decoder must handle a larger prototype set.
  • ad hoc to paper The calibration prototype distribution is well approximated by a Gaussian (or mixture of Gaussians) with diagonal covariance.
    Equation (6) assumes Gaussian form without empirical justification; the KL term to N(0,I) is a convenience for tractability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Prototype Calibration of Vision-Language Models for Generalized Few-shot Semantic Segmentation." pith.science (2026). https://pith.science/paper/QV5VQSF3

@misc{pith2026250622979,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Prototype Calibration of Vision-Language Models for Generalized Few-shot Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QV5VQSF3}},
  note         = {Machine review of arXiv:2506.22979}
}
abstract

Generalized Few-Shot Semantic Segmentation (GFSS) aims to extend a segmentation model to novel classes with only a few annotated examples while maintaining performance on base classes. Recently, pretrained vision-language models (VLMs) such as CLIP have been leveraged in GFSS to improve generalization on novel classes through multi-modal prototypes learning. However, existing prototype-based methods are inherently deterministic, limiting the adaptability of learned prototypes to diverse samples, particularly for novel classes with scarce annotations. To address this, we propose FewCLIP, a probabilistic prototype calibration framework over multi-modal prototypes from the pretrained CLIP, thus providing more adaptive prototype learning for GFSS. Specifically, FewCLIP first introduces a prototype calibration mechanism, which refines frozen textual prototypes with learnable visual calibration prototypes, leading to a more discriminative and adaptive representation. Furthermore, unlike deterministic prototype learning techniques, FewCLIP introduces distribution regularization over these calibration prototypes. This probabilistic formulation ensures structured and uncertainty-aware prototype learning, effectively mitigating overfitting to limited novel class data while enhancing generalization. Extensive experimental results on PASCAL-5$^i$ and COCO-20$^i$ datasets demonstrate that our proposed FewCLIP significantly outperforms state-of-the-art approaches across both GFSS and class-incremental setting. The code is available at https://github.com/jliu4ai/FewCLIP.

Figures

Figures reproduced from arXiv: 2506.22979 by the authors.

Figure 1
Figure 1. Concept diagram comparing deterministic and prob￾abilistic prototype learning with pre-trained CLIP in GFSS task. (a) Existing methods integrate a learnable or precomputed visual prototype with a frozen textual prototype for each class to perform pixel-wise deterministic classification. (b) Our pro￾posed FewCLIP models the class-wise visual calibration prototype as a distribution, capturing intra-class variations an… view at source ↗
Figure 2
Figure 2. Overview of the proposed FewCLIP with probabilistic prototype calibration. Given an image and a set of class names, we first extract textual prototypes Pt and visual embeddings [g, H] using CLIP’s text and image encoders, respectively. For each class, we learn a sample-specific distribution p(zi|P i t , g) and a calibration prototype P i c , which are further integrated into a probabilistic calibration prototype Pˆi… view at source ↗
Figure 3
Figure 3. Benefit of introducing latent prototypes zm from the prototype distribution when dealing with large object varia￾tions. The image masked with red rectangles denotes the training sample. Left triplet: Person , right triplet: Car . of 0.5 to train the calibration prototypes for 100 epochs by using novel classes samples. See more details in Appendix. 5.1. Comparison with state-of-the-art GFSS setting [PITH_FULL_IMAGE:… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative examples of O3S [43] and our method on PASCAL-5i . While O3S ’forget’ base classes (2-th column) or fail to generalize to novel classes (7-th column), our model effectively generalizes to novel classes without compromising base classes [PITH_FULL_IMAGE:fig…
Figure 5
Figure 5. Figure 5: Effects of different number M of latent prototypes. and the dining table. Likewise, as depicted in [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Qualitative examples of generalized few-shot semantic segmentation on PASCAL-5i . Results from the simple baseline and our method are reported. Examples with large object variations are exhibited [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Qualitative examples of generalized few-shot semantic segmentation on COCO-20i . Results from the simple baseline and our method are reported. Examples with large object variations are exhibited [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multi-Cache Enhanced Prototype Learning for Test-Time Generalization of Vision-Language Models

    cs.CV 2025-08 unverdicted novelty 5.0 of 10

    The submitted full text does not match the abstract, so the manuscript cannot be assessed as a coherent preprint.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ex- ploiting a joint embedding space for generalized zero-shot semantic segmentation

    Donghyeon Baek, Youngmin Oh, and Bumsub Ham. Ex- ploiting a joint embedding space for generalized zero-shot semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9536– 9545, 2021. 5

    work page 2021

  2. [2]

    Effective conditioned and composed im- age retrieval combining clip-based features

    Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Al- berto Del Bimbo. Effective conditioned and composed im- age retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 21466–21474, 2022. 2

    work page 2022

  3. [3]

    Prototype-based incre- mental few-shot semantic segmentation

    Fabio Cermelli, Massimiliano Mancini, Yongqin Xian, Zeynep Akata, and Barbara Caputo. Prototype-based incre- mental few-shot semantic segmentation. In British Machine Vision Conference, 2021. 1, 7, 14

    work page 2021

  4. [4]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017. 1

    2017

  5. [5]

    Masked-attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1

    work page 2022

  6. [6]

    Bayesian prompt learn- ing for image-language model generalization

    Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learn- ing for image-language model generalization. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 15237–15246, 2023. 3

    work page 2023

  7. [7]

    Few-shot semantic segmen- tation with prototype learning

    Nanqing Dong and Eric P Xing. Few-shot semantic segmen- tation with prototype learning. In British Machine Vision Conference, 2018. 1

    work page 2018

  8. [8]

    The pascal visual object classes (voc) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. In IJCV, 2010. 5

    work page 2010

Show all 50 references
  1. [9]

    A strong baseline for generalized few-shot semantic segmentation

    Sina Hajimiri, Malik Boudiaf, Ismail Ben Ayed, and Jose Dolz. A strong baseline for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11269– 11278, 2023. 1, 2, 3, 5, 6, 12, 13

    2023

  2. [10]

    Visual prompting for generalized few- shot segmentation: A multi-scale approach

    Mir Rayat Imtiaz Hossain, Mennatullah Siam, Leonid Sigal, and James J Little. Visual prompting for generalized few- shot segmentation: A multi-scale approach. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23470–23480, 2024. 1, 2

    2024

  3. [11]

    Semivl: semi- supervised semantic segmentation with vision-language guidance

    Lukas Hoyer, David Joseph Tan, Muhammad Ferjad Naeem, Luc Van Gool, and Federico Tombari. Semivl: semi- supervised semantic segmentation with vision-language guidance. In European Conference on Computer Vision , pages 257–275. Springer, 2024. 3

    2024

  4. [12]

    Prototyp- ical kernel learning and open-set foreground perception for generalized few-shot semantic segmentation

    Kai Huang, Feigege Wang, Ye Xi, and Yutao Gao. Prototyp- ical kernel learning and open-set foreground perception for generalized few-shot semantic segmentation. arXiv preprint arXiv:2308.04952, 2023. 1, 2, 3, 5, 6, 7, 12, 13, 14

    2023 arXiv

  5. [13]

    Clap4clip: Contin- ual learning with probabilistic finetuning for vision-language models

    Saurav Jha, Dong Gong, and Lina Yao. Clap4clip: Contin- ual learning with probabilistic finetuning for vision-language models. arXiv preprint arXiv:2403.19137, 2024. 2, 3, 5

    2024 arXiv

  6. [14]

    Scaling up visual and vision-language representa- tion learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representa- tion learning with noisy text supervision. In International conference on machine learning, pages 4904–4916. PMLR,

  7. [15]

    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. In European Conference on Computer Vision, pages 709–727. Springer, 2022. 2, 3, 6

    2022

  8. [16]

    Finetuning pretrained vision-language models with correlation informa- tion bottleneck for robust visual question answering

    Jingjing Jiang, Ziyi Liu, and Nanning Zheng. Finetuning pretrained vision-language models with correlation informa- tion bottleneck for robust visual question answering. arXiv preprint arXiv:2209.06954, 2022. 2

    2022 arXiv

  9. [17]

    Varia- tional dropout and the local reparameterization trick

    Durk P Kingma, Tim Salimans, and Max Welling. Varia- tional dropout and the local reparameterization trick. Ad- vances in neural information processing systems , 28, 2015. 5

    2015

  10. [18]

    Fine-tuning can distort pretrained fea- tures and underperform out-of-distribution

    Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained fea- tures and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022. 2

    2022 arXiv

  11. [19]

    Probabilistic prompt learning for dense prediction, 2023

    Hyeongjun Kwon, Taeyong Song, Somi Jeong, Jin Kim, Jinhyun Jang, and Kwanghoon Sohn. Probabilistic prompt learning for dense prediction, 2023. 3, 5

    2023

  12. [20]

    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. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 8057–8067, 2022. 6, 7, 13, 14

    2022

  13. [21]

    Visualbert: A simple and perfor- mant baseline for vision and language

    Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perfor- mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2

    1908 arXiv

  14. [22]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vi- sion, pages 740–755, 2014. 5

    2014

  15. [23]

    Dynamic prototype convolu- tion network for few-shot semantic segmentation

    Jie Liu, Yanqi Bao, Guo-Sen Xie, Huan Xiong, Jan-Jakob Sonke, and Efstratios Gavves. Dynamic prototype convolu- tion network for few-shot semantic segmentation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553–11562, 2022. 1

    2022

  16. [24]

    Clip-driven universal model for organ segmentation and tumor detection

    Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Compute...

    2023

  17. [25]

    Learning orthogonal pro- totypes for generalized few-shot semantic segmentation

    Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal pro- totypes for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 11319–11328, 2023. 1...

    2023

  18. [26]

    Inter- mediate prototype mining transformer for few-shot semantic segmentation

    Yuanwei Liu, Nian Liu, Xiwen Yao, and Junwei Han. Inter- mediate prototype mining transformer for few-shot semantic segmentation. Advances in Neural Information Processing Systems, 35:38020–38031, 2022. 1

    2022

  19. [27]

    Image retrieval on real-life images with pre- trained vision-and-language models

    Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre- trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 2125–2134, 2021. 2

    2021

  20. [28]

    Hypercorrela- tion squeeze for few-shot segmentation

    Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrela- tion squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941–6952, 2021. 1

    2021

  21. [29]

    Generalized few-shot semantic seg- mentation: All you need is fine-tuning

    Josh Myers-Dean, Yinan Zhao, Brian Price, Scott Cohen, and Danna Gurari. Generalized few-shot semantic seg- mentation: All you need is fine-tuning. arXiv preprint arXiv:2112.10982, 2021. 1, 2, 3

    2021 arXiv

  22. [30]

    ifs-rcnn: An incre- mental few-shot instance segmenter

    Khoi Nguyen and Sinisa Todorovic. ifs-rcnn: An incre- mental few-shot instance segmenter. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010–7019, 2022. 7, 14

    2022

  23. [31]

    Hierarchical dense cor- relation distillation for few-shot segmentation

    Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense cor- relation distillation for few-shot segmentation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641–23651, 2023. 1

    2023

  24. [32]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  25. [33]

    Denseclip: Language-guided dense prediction with context- aware prompting

    Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context- aware prompting. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 1808...

    2022

  26. [34]

    A sur- prisingly simple approach to generalized few-shot semantic segmentation

    Tomoya Sakai, Haoxiang Qiu, Takayuki Katsuki, Daiki Kimura, Takayuki Osogami, and Tadanobu Inoue. A sur- prisingly simple approach to generalized few-shot semantic segmentation. Advances in Neural Information Processing Systems, 37:27005–27023, 2025. 1, 2, 6

    2025

  27. [35]

    One-shot learning for semantic segmentation

    Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. In BMVC, 2017. 5

    2017

  28. [36]

    Prototypical networks for few-shot learning

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 13

    2017

  29. [37]

    Prior guided feature enrich- ment network for few-shot segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(2):1050– 1065, 2020

    Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrich- ment network for few-shot segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(2):1050– 1065, 2020. 1, 7, 14

    2020

  30. [38]

    Generalized few-shot se- mantic segmentation

    Zhuotao Tian, Xin Lai, Li Jiang, Shu Liu, Michelle Shu, Hengshuang Zhao, and Jiaya Jia. Generalized few-shot se- mantic segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 11563–11572, 2022. 1, 2, 3, 5, 6, 7, 12, 13, 14

    2022

  31. [39]

    Rethinking prior information genera- tion with clip for few-shot segmentation

    Jin Wang, Bingfeng Zhang, Jian Pang, Honglong Chen, and Weifeng Liu. Rethinking prior information genera- tion with clip for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941–3951, 2024. 1, 2

    2024

  32. [40]

    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. In proceedings of the IEEE/CVF International Conference on Computer Vision , pages 9197–9206, 2019. 1

    2019

  33. [41]

    Cris: Clip- driven referring image segmentation

    Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip- driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686–11695, 2022. 2

    2022

  34. [42]

    Prototype mixture models for few-shot semantic segmentation

    Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qix- iang Ye. Prototype mixture models for few-shot semantic segmentation. In Computer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part VIII 16, pages 763–778. Springer, 2020. 1

    2020

  35. [43]

    Multi-modal prototypes for open-set semantic segmentation

    Yuhuan Yang, Chaofan Ma, Chen Ju, Ya Zhang, and Yan- feng Wang. Multi-modal prototypes for open-set semantic segmentation. arXiv preprint arXiv:2307.02003, 2023. 2, 3, 4, 6, 7, 8

    2023 arXiv

  36. [44]

    Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation

    Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 9587–9595,

  37. [45]

    Canet: Class-agnostic segmentation networks with it- erative refinement and attentive few-shot learning

    Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with it- erative refinement and attentive few-shot learning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5217–5226, 2019. 1

    2019

  38. [46]

    Exploring regional clues in clip for zero-shot semantic seg- mentation

    Yi Zhang, Meng-Hao Guo, Miao Wang, and Shi-Min Hu. Exploring regional clues in clip for zero-shot semantic seg- mentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3270– 3280, 2024. 3

    2024

  39. [47]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2881–2890, 2017. 1

    2017

  40. [48]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In- ternational Journal of Computer Vision, 130(9):2337–2348,

  41. [49]

    Zegclip: Towards adapting clip for zero-shot se- mantic segmentation

    Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot se- mantic segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 11175–11185, 2023. 2, 3, 4

    2023

  42. [50]

    Unlocking the potential of pre-trained vision transform- ers for few-shot semantic segmentation through relationship descriptors

    Ziqin Zhou, Hai-Ming Xu, Yangyang Shu, and Lingqiao Liu. Unlocking the potential of pre-trained vision transform- ers for few-shot semantic segmentation through relationship descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pag...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.