Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Regularized Low-Rank Adaptation for Few-Shot Organ Segmentation

T0 review · 3 major / 4 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 ARENA claims that a sparsity penalty on the singular values of a LoRA update lets each few-shot organ-segmentation task discover its own effective rank.

desk verdict The rank-adaptive mechanism is switched off by its own hyperparameters (ρ=0), so the central claim that ARENA learns task-adapted ranks does not follow from the stated algorithm. read the letter →

arxiv 2507.15793 v1 pith:IKBY3U52 submitted 2025-07-21 cs.CV

Ghassen Baklouti , Julio Silva-Rodríguez , Jose Dolz , Houda Bahig , Ismail Ben Ayed This is my paper
classification cs.CV
keywords Few-shotlearningSegmentationLow-rankadaptationParameter-efficientfine-tuningSingularvaluedecompositionProximaloptimizationOrganFoundationmodels
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 fixed rank in Low-Rank Adaptation (LoRA) is a serious limitation for few-shot medical-image segmentation: the best rank differs from organ to organ, and there is too little validation data to search for it. It proposes ARENA, which writes the LoRA update as $W = W_0 + B\,\mathrm{Diag}(v)\,A$ and adds an $\ell^1$ penalty on the singular-value vector $v$; because the number of nonzero entries of $v$ is exactly the decomposition rank, the penalty acts as a rank penalty. The paper reports that this penalty, optimized with a proximal update, gives consistent DICE gains over standard LoRA, notably $+8.9$ and $+11.2$ points on novel-organ tasks at 5 and 10 shots, and that it is robust to suboptimal initial rank. If true, this would make few-shot adaptation of segmentation foundation models less dependent on rank selection.

What carries the argument

The central machinery is the singular-value parameterization $W = W_0 + B\,\mathrm{Diag}(v)\,A$ combined with the $\ell^1$ penalty $\lambda\|v\|_1$ on the singular-value vector $v$. The number of non-zero entries of $v$ equals the rank of the adaptation, so sparsifying $v$ is a direct penalty on rank. Optimization alternates standard gradient steps for $A$ and $B$ with a proximal step for $v$ (the standard way to handle a non-smooth sparsity penalty); the closed form is soft-thresholding, $\mathrm{prox}_{\eta_t\lambda\|\cdot\|_1}(v^t - \rho\nabla_v L)$, which zeroes entries below a threshold and shrinks the rest.

What would settle it

Re-run the novel-organ experiments (MYO, LA, RA, LV, RV at 5 and 10 shots) with $\rho$ set to a positive value, for instance $\rho=1$, and compare DICE with the reported $\rho=0$ results. If the gains come from task-adaptive rank, the $\rho>0$ runs should meet or exceed them; if they do not, the gains are attributable to shrinkage alone. A complementary check is to record the count of non-zero entries of $v$ during training and see whether it changes differently across easy and hard tasks.

Watch

Extended reading notes

Core claim

ARENA's central claim is that the effective rank of a LoRA adaptation can be learned during fine-tuning rather than fixed in advance. The paper represents the incremental weight matrix as $B\,\mathrm{Diag}(v)\,A$, where $v$ holds the singular values, and minimizes $L(A,B,v)+\lambda\|v\|_1$ by alternating gradient steps on $A,B$ with a proximal soft-thresholding step on $v$. Soft-thresholding drives small singular values to zero, so the number of non-zero singular values is reduced and the final rank is determined by the task rather than by the initialization. The paper's experiments in a few-shot volumetric segmentation setting show ARENA outperforming vanilla LoRA, AdaLoRA, linear probing, BitFit, Affine-LN, and full fine-tuning on both novel and base organs, with the largest margins on novel structures.

Load-bearing premise

The load-bearing premise is that the singular-value vector is steered by the segmentation loss during training, but Section 3.1 reports $\rho=0$, which drops the loss-gradient term from the update and leaves only a fixed soft-thresholding operation; the claimed task-adaptive rank mechanism is therefore not actually exercised by the experiments as reported.

Editorial extensions

If this is right

  • Rank selection for LoRA can be removed from the few-shot model-selection loop: the regularizer automatically favors sparser or denser adaptations per task.
  • On novel organs, the realistic transfer target for a segmentation foundation model, ARENA's reported margins over vanilla LoRA are large enough to change deployment choices in low-data clinical settings.
  • ARENA reaches or exceeds full fine-tuning on several few-shot tasks while training only the low-rank matrices, so the accuracy gain does not come at the cost of parameter efficiency.
  • Because the construction is defined directly on the LoRA factorization, it is not tied to a particular task-specific architecture and applies to any LoRA-style adaptation of a pre-trained network.

Reading between the lines

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

  • A distinction the paper leaves untested: the proximal-update scaling is set to $\rho=0$ in the reported experiments, which removes the loss-gradient term from the singular-value update and leaves pure soft-thresholding; the gains may come from shrinkage of the singular values rather than from gradient-driven rank discovery, and a direct comparison with $\rho>0$ would settle which mechanism matters
  • A natural extension is to compare ARENA against a post-hoc baseline that trains fixed-rank LoRA and then prunes small singular values, isolating the benefit of interleaving sparsity with training from the benefit of merely shrinking the final weights.
  • The $\ell^1$-on-singular-values formulation is a continuous proxy for rank minimization; connecting it to nuclear-norm or matrix-rank theory could give a principled account of why the learned rank varies across organs and shots.
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

3 major / 4 minor

Summary. This manuscript proposes ARENA, a parameter-efficient fine-tuning method for few-shot volumetric organ segmentation. It reparameterizes LoRA updates as W = W0 + B Diag(v) A, adds an l1 penalty on the singular-value vector v, and optimizes with block coordinate descent in which A and B receive standard gradient updates while v is updated via a proximal soft-thresholding step. The central claim is that this procedure automatically finds task-adapted ranks. Experiments compare ARENA to LoRA, AdaLoRA, BitFit, affine-LN, and full fine-tuning on TotalSegmentator and FLARE'22 for both base and novel organs, reporting average gains over LoRA of +8.9 and +11.2 DICE in the novel-organ 5-shot and 10-shot settings.

Significance. The paper targets a relevant problem: rank selection in LoRA under validation-free few-shot adaptation. The evaluation setting is realistic, the code is publicly available, and the comparison covers several PEFT baselines. If the rank-adaptation mechanism worked as described, ARENA would be a useful practical contribution. However, the paper's own implementation sets rho = 0 in the proximal update, which removes the dependence on the segmentation loss from the v-update. The central mechanism is therefore not operational in the reported experiments, and the empirical gains cannot be attributed to the advertised adaptive rank selection. As written, the core contribution is not established.

major comments (3)
  1. [Section 3.1, Eq. (5)] The implementation states that 'The gating vector parameters are adjusted to lambda = 0.5 and rho = 0 according to the updating rule in Eq. 5.' With rho = 0, the term -rho * grad_v L(A,B,v) in Eq. (5) vanishes, so v(t+1) = prox_{eta_t lambda ||.||_1}(v(t)). The singular-value vector is then updated by a loss-independent soft-thresholding schedule, and the effective rank becomes a deterministic function of the initial v and the threshold schedule. This contradicts the central claim in the Abstract and Section 2 that minimizing the l1-regularized loss 'enables to find task-adapted ranks automatically.' The mechanism that is supposed to make the rank task-adaptive is switched off in the reported configuration.
  2. [Section 3.2, Table 1] The reported gains of +8.9 and +11.2 DICE over LoRA are presented as evidence for the proposed adaptive rank selection. Since the v-update is independent of the segmentation loss under the chosen rho = 0, these gains cannot be explained by loss-coupled rank adaptation. They might arise from the fixed shrinkage schedule, from the SVD reparameterization, or from the interaction with decoder fine-tuning. The paper needs either experiments with rho > 0 or an explicit alternative mechanism before the headline claim can be credited.
  3. [Section 3.1, hyperparameters] With rho = 0, the entire rank-adaptation behavior in the current implementation is governed by lambda and the threshold schedule eta_t * lambda. The paper reports a single value lambda = 0.5 with no sensitivity analysis and no reporting of the final effective ranks per task. Given that the central claim is automatic task-adapted rank selection, at least a sensitivity study over lambda and a report of the learned rank for each organ are necessary to substantiate the claim.
minor comments (4)
  1. [Fig. 1] The figure caption refers to 'Adaptive-LoRA' while the method is called ARENA; the terminology should be unified.
  2. [Section 3.1, Evaluation protocol] Results are averaged over 3 seeds but no standard deviations or confidence intervals are reported; adding variance estimates would strengthen the comparisons.
  3. [Section 2] The acronym expansion 'Adaptive Rank sEgmeNtAtion' uses an unusual capitalization pattern; consider using the conventional form 'Adaptive Rank Segmentation' consistently.
  4. [Section 3.1, Adaptation training details] Early stopping is performed on the support-set loss, which is a form of validation; the 'validation-free' claim should be qualified or the stopping rule should be described as part of the adaptation protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: ARENA's l1-regularized SVD objective is not defined in terms of the reported Dice gains, and the empirical evaluation is anchored to external benchmarks; the rho=0 implementation detail is an internal-consistency concern, not a circularity.

full rationale

The paper's central claim—that the l1 sparsity regularizer on singular values enables automatic rank selection—is a standard proximal-gradient construction (Eqs. 2-6). Nothing in that construction is defined in terms of the test-set Dice scores or the reported +8.9/+11.2 gains; the regularizer is a fixed penalty with lambda=0.5, and the experiments compare against external baselines on TotalSegmentator and FLARE'22. The paper does rely on the authors' prior foundation model and few-shot framework [27] for the model, data splits, and decoder-freezing choices, but that reliance is infrastructural rather than argumentative: [27] is a published, externally evaluated prior work, and no uniqueness claim or theorem from [27] is invoked to forbid alternative methods. Reference [34] (Zanella & Ben Ayed) is cited only as CLIP-LoRA context and is not load-bearing. The most significant manuscript-level issue is internal consistency, not circularity: Section 3.1 states 'The gating vector parameters are adjusted to lambda = 0.5 and rho = 0 according to the updating rule in Eq. 5.' Substituting rho=0 into Eq. 5 removes the -rho * nabla_v L term, so the proximal update for v reduces to a loss-independent soft-thresholding schedule; this weakens the advertised 'task-adapted ranks' mechanism. However, this is a correctness/falsifiability concern about whether the algorithm implements the stated mechanism, not a case of a prediction being equivalent to its inputs by construction. The Dice results are externally grounded, so no circularity score is warranted.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or architectural entities. The free parameters and domain assumptions are standard for a PEFT method, with the notable exception that the rho=0 setting undermines the claimed adaptive mechanism.

free parameters (3)
  • Regularization coefficient lambda = 0.5
    Chosen by hand in Section 3.1 to control the sparsity versus fidelity trade-off; no sensitivity analysis is reported.
  • Gradient step-size rho for v-update = 0
    Set to 0 in Section 3.1, which removes the gradient term from the proximal update in Eq. (5); this choice defeats the claimed task-adaptive rank learning.
  • Initial rank r = 8
    Fixed initial rank for ARENA across experiments (Section 3.1); robustness to this choice is claimed but the initial value is still a manual selection.
assumptions (4)
  • standard math Soft-thresholding is the proximal operator of the l1 norm.
    Used to derive Eq. (5)-(6); standard convex analysis result.
  • standard math The proximal gradient algorithm converges to a minimizer of the regularized loss L + lambda||v||1.
    Assumed for optimizing Eq. (2); cites Beck and Teboulle (2009).
  • domain assumption The unconstrained factor matrices A and B in BDiag(v)A preserve the rank interpretation of v.
    The paper calls Eq. (1) an SVD, but A and B are not restricted to be orthogonal; the rank of BDiag(v)A is not exactly the number of nonzero entries of v unless A and B have full row and column rank. Section 2.
  • domain assumption A support set of 5-10 annotated volumes is sufficient to fine-tune the model without a validation set.
    The few-shot protocol in Section 3.1 assumes this realistic clinical setting for early stopping and model selection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Regularized Low-Rank Adaptation for Few-Shot Organ Segmentation." pith.science (2026). https://pith.science/paper/IKBY3U52

@misc{pith2026250715793,
  author       = {Pith},
  title        = {Pith review of: Regularized Low-Rank Adaptation for Few-Shot Organ Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IKBY3U52}},
  note         = {Machine review of arXiv:2507.15793}
}
read the original abstract

Parameter-efficient fine-tuning (PEFT) of pre-trained foundation models is increasingly attracting interest in medical imaging due to its effectiveness and computational efficiency. Among these methods, Low-Rank Adaptation (LoRA) is a notable approach based on the assumption that the adaptation inherently occurs in a low-dimensional subspace. While it has shown good performance, its implementation requires a fixed and unalterable rank, which might be challenging to select given the unique complexities and requirements of each medical imaging downstream task. Inspired by advancements in natural image processing, we introduce a novel approach for medical image segmentation that dynamically adjusts the intrinsic rank during adaptation. Viewing the low-rank representation of the trainable weight matrices as a singular value decomposition, we introduce an l_1 sparsity regularizer to the loss function, and tackle it with a proximal optimizer. The regularizer could be viewed as a penalty on the decomposition rank. Hence, its minimization enables to find task-adapted ranks automatically. Our method is evaluated in a realistic few-shot fine-tuning setting, where we compare it first to the standard LoRA and then to several other PEFT methods across two distinguishable tasks: base organs and novel organs. Our extensive experiments demonstrate the significant performance improvements driven by our method, highlighting its efficiency and robustness against suboptimal rank initialization. Our code is publicly available: https://github.com/ghassenbaklouti/ARENA

Figures

Figures reproduced from arXiv: 2507.15793 by the authors.

Figure 1
Figure 1. Adaptive-LoRA. We introduce a novel few-shot PEFT technique for Adaptive Rank Segmentation (ARENA) that is (a) robust to rank initializa￾tion and (b) enhances the parameter efficiency vs. performance trade-off. the input text prompts, could be beneficial. Given the growing success of PEFT in vision and NLP, there is an increasing interest in its application to medi￾cal domains and in the few-shot learning settings. … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 37 canonical work pages

  1. [1]

    In: AAAI

    Basu, S., et al.: Strong baselines for parameter efficient few-shot fine-tuning. In: AAAI. vol. 38, pp. 11024–11031 (2024) 2, 6, 7, 8

    work page 2024

  2. [2]

    SIAM journal on imaging sciences2(1), 183–202 (2009) 5

    Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences2(1), 183–202 (2009) 5

    work page 2009

  3. [3]

    Neural Information Processing Systems (NeurIPS)33, 2445–2457 (2020) 2

    Boudiaf, M., Ziko, I., et al.: Information maximization for few-shot learning. Neural Information Processing Systems (NeurIPS)33, 2445–2457 (2020) 2

    work page 2020

  4. [4]

    Boudiaf, M., et al.: Few-shot segmentation without meta-learning: A good trans- ductive inference is all you need? In: Computer Vision and Pattern Recognition (CVPR). pp. 13979–13988 (2021) 2

    work page 2021

  5. [5]

    arXiv preprint arXiv:2306.07967 (2023) 3

    Chavan, A., Liu, Z., Gupta, D., Xing, E., Shen, Z.: One-for-all: Generalized lora for parameter-efficient fine-tuning. arXiv preprint arXiv:2306.07967 (2023) 3

    arXiv 2023

  6. [6]

    In: International Conference on Learning Representations (ICLR) (2019) 2

    Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B.: A closer look at few-shot classification. In: International Conference on Learning Representations (ICLR) (2019) 2

    work page 2019

  7. [7]

    Medical Image Analysis79, 4 (2022) 1

    Chen, X., Wang, X., Zhang, K., Fung, K.M., Thai, T.C., Moore, K., Mannel, R.S., Liu, H., Zheng, B., Qiu, Y.: Recent advances and clinical applications of deep learning in medical image analysis. Medical Image Analysis79, 4 (2022) 1

    work page 2022

  8. [8]

    Neural Infor- mation Processing Systems (NeurIPS)36, 10088–10115 (2023) 3

    Dettmers, T., et al.: Qlora: Efficient finetuning of quantized LLMs. Neural Infor- mation Processing Systems (NeurIPS)36, 10088–10115 (2023) 3

    work page 2023

Show all 38 references
  1. [9]

    In: EMNLP (2023) 3, 5

    Ding, N., et al.: Sparse low-rank adaptation of pre-trained language models. In: EMNLP (2023) 3, 5

    2023

  2. [10]

    In: MICCAI (2024) 2, 3

    Fereshteh Shakeri, Y., et al.: Few-shot adaptation of medical vision-language mod- els. In: MICCAI (2024) 2, 3

    2024

  3. [11]

    International Journal of Computer Vision132(2), 581–595 (2024) 2 10 G

    Gao, P., et al.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision132(2), 581–595 (2024) 2 10 G. Baklouti et al

    2024

  4. [12]

    In: Computer Vision and Pattern Recognition (CVPR)

    Hajimiri, S., Boudiaf, M., et al.: A strong baseline for generalized few-shot semantic segmentation. In: Computer Vision and Pattern Recognition (CVPR). pp. 11269– 11278 (2023) 2

    2023

  5. [13]

    Hatamizadeh, A., et al.: Swin unetr: Swin transformers for semantic segmentation ofbraintumorsinmriimages.In:InterventionBrainlesionMICCAIW.pp.272–284 (2022) 5

    2022

  6. [14]

    In: International Conference on Learning Representations (ICLR) (2022) 2

    He, J., Zhou, C., Ma, X., Berg-Kirkpatrick, T., Neubig, G.: Towards a unified view of parameter-efficient transfer learning. In: International Conference on Learning Representations (ICLR) (2022) 2

    2022

  7. [15]

    International Confer- ence on Learning Representations (ICLR)1(2), 3 (2022) 2, 3, 6, 7, 8

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. International Confer- ence on Learning Representations (ICLR)1(2), 3 (2022) 2, 3, 6, 7, 8

    2022

  8. [16]

    In: Computer Vision and Pattern Recognition (CVPR) (2024) 2

    Huang, Y., et al.: LP++: A surprisingly strong linear probe for few-shot CLIP. In: Computer Vision and Pattern Recognition (CVPR) (2024) 2

    2024

  9. [17]

    In: International Conference on Computer Vision (ICCV)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: International Conference on Computer Vision (ICCV). pp. 4015–4026 (2023) 2

    2023

  10. [18]

    Li, W., Yuille, A., Zhou, Z.: How well do supervised models transfer to 3d image segmentation? In: International Conference on Learning Representations (ICLR). pp. 1–16 (2024) 2

    2024

  11. [19]

    Medical image analysis42, 60–88 (2017) 1

    Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., Van Der Laak, J.A., Van Ginneken, B., Sánchez, C.I.: A survey on deep learning in medical image analysis. Medical image analysis42, 60–88 (2017) 1

    2017

  12. [20]

    In: International Conference on Computer Vision (ICCV)

    Liu, J., Zhang, Y., Chen, J.N., Xiao, J., Lu, Y., Landman, B.A., Yuan, Y., Yuille, A., Tang, Y., Zhou, Z.: CLIP-Driven universal model for organ segmentation and tumor detection. In: International Conference on Computer Vision (ICCV). pp. 21152–21164 (2023) 2

    2023

  13. [21]

    Nature Medicine30, 863–874 (2024) 2

    Lu, M.Y., et al.: A visual-language foundation model for computational pathology. Nature Medicine30, 863–874 (2024) 2

    2024

  14. [22]

    Lancet Digit Health (2024) 6

    Ma, J., et al.: Unleashing the strengths of unlabeled data in pan-cancer abdominal organ quantification: the flare22 challenge. Lancet Digit Health (2024) 6

    2024

  15. [23]

    Nature616(7956), 259–265 (2023) 2

    Moor, M., Banerjee, O., et al.: Foundation models for generalist medical artificial intelligence. Nature616(7956), 259–265 (2023) 2

    2023

  16. [24]

    Nature 616, 259–265 (4 2023) 2

    Moor, M., et al.: Foundation models for generalist medical artificial intelligence. Nature 616, 259–265 (4 2023) 2

    2023

  17. [25]

    In: ICML

    Radford, A., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763. PMLR (2021) 2

    2021

  18. [26]

    Medical Image Analysis99, 103357 (2025) 2

    Silva-Rodriguez, J., et al.: A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision. Medical Image Analysis99, 103357 (2025) 2

    2025

  19. [27]

    Medical Image Analysis103, 103596 (2025) 2, 3, 4, 5, 6

    Silva-Rodríguez, J., Dolz, J., Ben Ayed, I.: Towards foundation models and few- shot parameter-efficient fine-tuning for volumetric organ segmentation. Medical Image Analysis103, 103596 (2025) 2, 3, 4, 5, 6

    2025

  20. [28]

    In: MICCAI

    Ulrich, C., et al.: Multitalent: A multi-dataset approach to medical image segmen- tation. In: MICCAI. pp. 508–518 (2023) 2

    2023

  21. [29]

    Valipour, M., et al.: Dylora: Parameter-efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. In: ACL. pp. 3274–3287 (2023) 3, 8

    2023

  22. [30]

    In: Proceedings of EMNLP

    Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from un- paired medical images and text. In: Proceedings of EMNLP. vol. 2022, p. 3876 (2022) 2 Regularized Low-Rank Adaptation for Few-Shot Organ Segmentation 11

    2022

  23. [31]

    Radiology: Artificial Intelligence5(5), e230024 (2023) 2, 6

    Wasserthal, J., et al.: Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence5(5), e230024 (2023) 2, 6

    2023

  24. [32]

    EGOVIS13429, 32–46 (2022) 2

    Wójcik, M.A.: Foundation models in healthcare: Opportunities, biases and regu- latory prospects in europe. EGOVIS13429, 32–46 (2022) 2

    2022

  25. [33]

    In: ACL (2022) 2, 6, 7, 8

    Zaken, E., et al.: Bitfit: Simple parameter-efficient fine-tuning for transformer- based masked language models. In: ACL (2022) 2, 6, 7, 8

    2022

  26. [34]

    In: Computer Vision and Pattern Recognition (CVPR)

    Zanella, M., Ben Ayed, I.: Low-rank few-shot adaptation of vision-language models. In: Computer Vision and Pattern Recognition (CVPR). pp. 1593–1603 (2024) 2, 3

    2024

  27. [35]

    In: International Conference on Learning Representations (ICLR) (2023) 3, 6, 7, 8

    Zhang, Q., et al.: Adalora: Adaptive budget allocation for parameter-efficient fine- tuning. In: International Conference on Learning Representations (ICLR) (2023) 3, 6, 7, 8

    2023

  28. [36]

    In: European Conference on Computer Vision (ECCV)

    Zhang, R., et al.: Tip-adapter: Training-free adaption of clip for few-shot classifi- cation. In: European Conference on Computer Vision (ECCV). pp. 493–510 (2022) 2

    2022

  29. [37]

    Medical Image Analysis91, 102996 (2024) 2

    Zhang, S., Metaxas, D.: On the challenges and perspectives of foundation models for medical image analysis. Medical Image Analysis91, 102996 (2024) 2

    2024

  30. [38]

    International Journal of Computer Vision130(9), 2337–2348 (2022) 2

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision130(9), 2337–2348 (2022) 2

    2022

Pith tools

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