Pith. sign in

REVIEW 4 major objections 7 minor 55 references

InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic Perspective

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

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

Pith's one-line read InfoSAM claims that adding two mutual-information losses to SAM fine-tuning preserves domain-invariant pre-trained knowledge and improves segmentation on specialized benchmarks.

desk verdict InfoSAM is a broadly tested, modestly effective plug-in loss for SAM fine-tuning, but its information-theoretic 'domain-invariant' mechanism is not established; the gains likely come from generic regularization and Gram-matrix alignment. read the letter →

arxiv 2505.21920 v2 pith:S4NPWMNX submitted 2025-05-28 cs.CV

Yuanhong Zhang , Muyao Yuan , Weizhan Zhang , Tieliang Gong , Wen Wen , Jiangyong Ying , Weijie Shi This is my paper
classification cs.CV
keywords SegmentAnythingModelparameter-efficientfine-tuningknowledgedistillationmutualinformationRényientropydomain-invariantrepresentationimagesegmentationbottleneck
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

InfoSAM claims that the main thing lost when fine-tuning the Segment Anything Model (SAM) to specialized domains is not parameters but relationships: the stable, domain-invariant structural links between SAM's image encoder and its mask decoder. To keep those links, the paper adds two mutual-information-based losses to ordinary adapter or LoRA fine-tuning: one that compresses a learned relation representation $r^T$ so it keeps only domain-invariant information, and one that maximizes the mutual information between the teacher's relation $r^T$ and the student's relation $r^S$. On medical, agricultural, remote-sensing, and natural-image benchmarks, adding these losses raises segmentation quality (measured by IoU) by roughly 0.5 to 1 point over the same PEFT backbone, and the recipe transfers from SAM to SAM2. A sympathetic reader would care because it turns a vague worry about forgetting pre-trained knowledge into a concrete, measurable objective.

What carries the argument

The load-bearing object is an attention-based relation module $f(z_i^T,z_m^T;\theta)$ that maps SAM's image-encoder output $z_i^T$ and mask-decoder token $z_m^T$ to a normalized relation map $r^T$, with the teacher and student versions sharing parameters. Around it, the paper builds two losses: $L_r$ compresses the relation by minimizing $I_\alpha(z_i^T,z_m^T;r^T)$, and $L_d$ distills it by maximizing $I_\alpha(r^T;r^S)$. Both are evaluated with matrix-based Rényi's $\alpha$-entropy at $\alpha=2$, which turns the information quantities into Frobenius-norm terms on Gram matrices, avoiding eigenvalue decomposition. The shared parameterization lets the same information bottleneck be computed on the frozen teacher and the fine-tuning student.

What would settle it

Shuffle the teacher relation module's inputs during training, for example by permuting the order of mask tokens or image embeddings across a batch. If InfoSAM's improvement over plain Adapter or LoRA training persists unchanged, the relation module is not encoding pair-specific relations, and the gains come from the regularization terms alone rather than from preserved domain-invariant knowledge.

Watch

Extended reading notes

Core claim

The paper's central claim is that PEFT of SAM can be improved by explicitly preserving domain-invariant relational knowledge from the pre-trained model. Formally, InfoSAM solves a constrained optimization: minimize $I_\alpha(z_i^T,z_m^T;r^T)$ to compress the relation extracted from the teacher's image embeddings and mask tokens, while maximizing $I_\alpha(r^T;r^S)$ so the student's relation matches the teacher's. The two objectives are combined as $\max_\omega I_\alpha(r^T;r^S)-\beta I_\alpha(z_i^T,z_m^T;r^T)$, implemented with matrix-based Rényi's $\alpha$-entropy at $\alpha=2$, where the entropy terms reduce to Frobenius norms of Gram matrices and are cheap to compute. The author claims this information-bottleneck formulation selectively filters out pseudo-invariant features such as color statistics while keeping structural patterns like edges, and that this is why InfoSAM outperforms both standard PEFT methods and existing distillation methods, which align only paired features and can hurt when the teacher is weak.

Load-bearing premise

The load-bearing premise is that the shared relation module actually learns to encode the teacher's encoder-decoder relations rather than collapsing to a constant output, with the Frobenius-norm regularization terms as the only empirical safeguard against that collapse.

Editorial extensions

If this is right

  • Adding InfoSAM's $L_r$ and $L_d$ losses to Adapter or LoRA fine-tuning improves SAM's segmentation IoU by roughly 0.5 to 1 point across medical, agricultural, remote-sensing, and natural-image benchmarks, and the gains carry over to SAM2.
  • Because the losses are computed on relation maps rather than raw paired features, InfoSAM avoids the degradation that the paper observes for most distillation methods during PEFT, including when the teacher is weak.
  • The same relation module can be attached to different teacher sizes: InfoSAM improves a ViT-B student using ViT-L and ViT-H teachers as well, suggesting the recipe is not tied to a particular backbone.
  • The relation module's learned information appears transferable across domains: a module trained on one domain still helps when frozen and applied to another domain, which the paper reads as evidence that it captures domain-invariant structure.
  • The Frobenius-norm regularization terms in the losses are needed to keep the relation module from collapsing to trivial solutions, so the reported gains are the combined effect of distillation and regularization.

Reading between the lines

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

  • If the relation module truly encodes domain-invariant structure, a natural test is to attach the same module and losses to other large encoder-decoder vision models: the recipe should improve their fine-tuning without architectural changes.
  • A natural follow-up is to test whether the same relation module transfers to video or 3D segmentation, where the notion of structural invariance is richer.
  • The cross-domain transfer result suggests a practical pipeline: pre-train one relation module once, freeze it, and reuse it across many downstream tasks to cut per-task training cost.
  • Replacing the shared teacher-student relation module with two independently initialized modules would reveal how much of the gain depends on parameter sharing versus genuine relational alignment.
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 / 7 minor

Summary. The manuscript proposes InfoSAM, a parameter-efficient fine-tuning (PEFT) method for SAM that adds two information-theoretic losses derived from matrix-based Rényi α-entropy: a relation-compression loss Lr that minimizes Iα(z_i^T, z_m^T; r_T) and a distillation loss Ld that maximizes Iα(r_T; r_S), where r_T and r_S are outputs of an attention-based relation module applied to teacher and student image-encoder and mask-decoder features. With α=2, both losses reduce to Frobenius-norm terms. The method is evaluated on SAM and SAM2 with Adapter and LoRA across eight datasets in natural, medical, agricultural, and remote-sensing domains, and is compared with a broad set of PEFT and distillation baselines. Ablations address the contributions of Lr, Ld, the relation module, and its regularization terms.

Significance. The empirical scope is a strength: two backbones, two PEFT instantiations, eight datasets, three seeds, and comparisons with many baselines. The algebraic reduction of the Rényi losses to Frobenius norms (Eqs. 11 and 13) is transparent and computationally efficient. If the mechanism were established, the contribution would be a simple, architecture-agnostic plug-in loss for SAM fine-tuning. However, the paper's central mechanistic claim—that the losses preserve domain-invariant relational knowledge—is currently supported only by indirect ablations; the objective as written admits low-entropy, non-semantic solutions, and the reported gains may be attributable to generic spectral alignment. The cross-domain transfer experiment (Table 6) is suggestive but does not isolate the proposed information-theoretic mechanism.

major comments (4)
  1. [Section 4.2 / Algorithm 1] The claim that minimizing Eq. (5) and maximizing Eq. (6) preserves domain-invariant relations is asserted rather than derived. Because the relation modules are shared (f_s = f_t in Algorithm 1) and r_T is deterministic in (z_i^T, z_m^T), the objective can be satisfied by a low-entropy, nearly input-independent code: minimizing I_α(z_i^T, z_m^T; r_T) encourages low entropy of r_T, and maximizing I_α(r_T; r_S) then rewards similarity of two low-entropy codes. The Frobenius-norm terms in Eqs. (11) and (13) prevent collapse empirically (Appendix F.2), but they are generic spectral regularizers; nothing in the derivation shows they encode semantic, domain-invariant structure. Please add a direct test of the mechanism, e.g., measuring the task-relevant information in r_T (mutual information with labels) or comparing against a control with a frozen relation module and an equivalent spectral regularizer but no teacher.
  2. [Section 5.4 / Table 4] The ablation attributes most of the gain to Ld (e.g., +1.0 Sα on Kvasir for Ld vs +0.2 for Lr). Since Ld's third term is a Frobenius-norm alignment between teacher and student Gram matrices, the observed improvements are consistent with generic feature alignment or regularization rather than with transfer of pre-trained relational knowledge. A control experiment that replaces r_T/r_S with random projections, or that uses a non-information-theoretic alignment loss with the same spectral terms, is needed to support the specific interpretation in the title and abstract.
  3. [Appendix E.1 / Fig. 6] The hyperparameters λ1 and λ2 are selected by accuracy heatmaps on the same datasets (Kvasir, Leaf, Road) that appear in the main results. This is a form of test-set tuning and may inflate the reported gains relative to baselines. Please select λ1 and λ2 on held-out validation splits, or report results for fixed values across all datasets, together with the sensitivity range.
  4. [Section 5 / Tables 1-3] The claim of consistent superiority is based on mean differences of roughly 0.1–1.0 points with three seeds, and several differences are within one reported standard error of the baseline (e.g., CAMO Sα 88.6±0.09 vs SU-SAM 88.3±0.21; Leaf IoU 77.3±0.14 vs 76.8±0.56 for SAM2). Please report paired significance tests across seeds, confidence intervals, or additional seeds to establish that the improvements are not noise.
minor comments (7)
  1. [Section 3.1 / Eq. (9)] The symbol Sα is used both for Rényi entropy (Eq. (1)) and for the attention score (Eq. (9)); please use a distinct symbol to avoid confusion.
  2. [Section 4.2, after Eq. (9)] The sentence 'α is flattened and normalized' appears to be a typo; the attention map (or its normalized version) should be flattened and normalized.
  3. [Appendix A.2, Eq. (21) vicinity] The identity 'G^S_imr = G^T_i ◦ G^T_m ◦ G^T_r' appears to be a typo; the superscript S should likely be T.
  4. [Algorithm 1] The variable 'lamda_1' should be 'lambda_1'.
  5. [Section 5.3] The text says 'InfoSAM is better suited for PFET'; this should be 'PEFT'.
  6. [Table 3(b)] The method name 'InfoSAM2' is inconsistent with 'InfoSAM' used elsewhere; please harmonize the naming.
  7. [Section 4.2, Eq. (13)] The phrase 'log2 improving robustness to relations' is unclear; please clarify what quantity is robust to what variation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: InfoSAM's losses are derived algebraically from restated entropy definitions and validated on external benchmarks.

full rationale

InfoSAM's derivation chain is not circular. The mutual-information objectives in Eqs. (5)-(7) are defined through the matrix-based Rényi entropy restated in Definitions 1-3, and Eqs. (10)-(13) are algebraic substitutions (α=2, Frobenius norm identity ||A||_F^2 = Σ λ_i^2) applied to those definitions. No fitted parameter is renamed as a prediction, and the empirical IoU gains are measured against external benchmarks with standard PEFT and distillation baselines. The self-citations to Gong et al. (2022) and Dong et al. (2023), both co-authored by Tieliang Gong, are not load-bearing: the entropy definition and the Frobenius-norm equivalence are reproduced in the paper and are externally checkable mathematical facts. The shared relation module (f_s = f_t in Algorithm 1) and the possibility of low-entropy collapse are explicitly examined in Appendix F.2, where the authors show that the Frobenius regularization terms prevent trivial solutions; this is a mechanism-robustness concern, not an equivalence of inputs and outputs. Hyperparameters λ1, λ2 and α are chosen with sensitivity analysis on the evaluation datasets, which is a selection burden, but it does not make the central claim circular because the reported improvements are not constructed from those choices. Overall, the paper's derivation is self-contained and its central empirical claim is externally falsifiable.

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

The central claim rests on several domain assumptions: that frozen SAM contains transferable domain-invariant relations, that the compression loss isolates them, and that the shared relation module does not collapse. The math of the Rényi entropy reduction is standard, but the semantic claims are not established by the equations. No new physical or formal entities are introduced; the relation module is a standard attention module, and 'pseudo-invariant information' is borrowed from the domain generalization literature.

free parameters (5)
  • λ1 = 1
    Weight for relation compression loss Lr; selected from accuracy heatmaps on Kvasir, Leaf, and Road datasets in Appendix E.1, the same datasets used for main results.
  • λ2 = 0.5
    Weight for distillation loss Ld; selected from the same accuracy heatmaps in Appendix E.1.
  • α (Rényi entropy order) = 2
    Set to 2 to enable Frobenius norm computation; chosen after comparing 1.01, 2, and 3 in Table 11.
  • Adapter bottleneck ratio r/D = 0.25
    Standard adapter hyperparameter used for all adapter-based methods, stated in Appendix B.2.
  • LoRA rank r = 4
    Standard LoRA rank used for all LoRA-based methods, stated in Appendix B.2.
assumptions (5)
  • standard math Matrix-based Rényi α-entropy as defined in Eq. (1), and joint entropy via Hadamard products in Eq. (2), is a valid estimator of mutual information.
    Taken from Yu et al. (2019) and Gong et al. (2022), including works co-authored by the present authors. It is an established estimator but not universally accepted as a true mutual information.
  • standard math For α=2, the entropy terms equal -log2 of squared Frobenius norms of trace-normalized Gram matrices.
    Algebraic consequence of the eigenvalue expansion; stated in Appendix A.2.
  • domain assumption The frozen pre-trained SAM encodes domain-invariant relations between image encoder and mask decoder that are beneficial to transfer.
    This is the central motivation of Section 4.2; no direct evidence is given beyond the downstream transfer experiments.
  • ad hoc to paper Minimizing I(z_i,z_m;r_T) removes pseudo-invariant information and preserves domain-invariant information.
    The loss Lr has no term that identifies domain-specific semantics; the interpretation that it filters pseudo-invariant information is asserted, not derived.
  • domain assumption A single shared relation module can represent teacher and student relations without a trivial collapse.
    The shared parameters in Algorithm 1 create a shortcut for maximizing I(r_T;r_S); the method assumes the regularization terms prevent degenerate solutions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic Perspective." pith.science (2026). https://pith.science/paper/S4NPWMNX

@misc{pith2026250521920,
  author       = {Pith},
  title        = {Pith review of: InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4NPWMNX}},
  note         = {Machine review of arXiv:2505.21920}
}
read the original abstract

The Segment Anything Model (SAM), a vision foundation model, exhibits impressive zero-shot capabilities in general tasks but struggles in specialized domains. Parameter-efficient fine-tuning (PEFT) is a promising approach to unleash the potential of SAM in novel scenarios. However, existing PEFT methods for SAM neglect the domain-invariant relations encoded in the pre-trained model. To bridge this gap, we propose InfoSAM, an information-theoretic approach that enhances SAM fine-tuning by distilling and preserving its pre-trained segmentation knowledge. Specifically, we formulate the knowledge transfer process as two novel mutual information-based objectives: (i) to compress the domain-invariant relation extracted from pre-trained SAM, excluding pseudo-invariant information as possible, and (ii) to maximize mutual information between the relational knowledge learned by the teacher (pre-trained SAM) and the student (fine-tuned model). The proposed InfoSAM establishes a robust distillation framework for PEFT of SAM. Extensive experiments across diverse benchmarks validate InfoSAM's effectiveness in improving SAM family's performance on real-world tasks, demonstrating its adaptability and superiority in handling specialized scenarios.

Figures

Figures reproduced from arXiv: 2505.21920 by the authors.

Figure 1
Figure 1. Comparing traditional PEFT and distillation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The Flowchart of InfoSAM. To leverage the domain-invariant relationships within modules from a well-trained foundation model (i.e., SAM) for enhancing PEFT. InfoSAM pioneers an information-theoretic framework for parameter￾efficient SAM adaptation through two synergistic components: 1) Strategic compression of task-irrelevant dependencies while preserving domain-invariant feature relationships through optimized inte… view at source ↗
Figure 3
Figure 3. The architecture of attention-based relation mod [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Performance of InfoSAM with larger teacher models (i.e., ViT-L, ViT-H), while the student remains based on a ViT-B backbone. Each subfigure shows IoU metrics for different distillation methods on a specific dataset. dent’s features to fully mimic the teacher at every s…
Figure 5
Figure 5. Figure 5: The architecture of InfoSAM with Adapter and LoRA [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Hyper-parameter sensitivity study of λ1 and λ2 in balancing Lr and Ld, with Kvasir, Leaf, and Road datasets (Best viewed in color). 19 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Boundary quality measured by the F1-score at various thresholds (0, 1, 2, and 3 pixels) on the Leaf dataset. The thresholds represent the allowable pixel distance: if the predicted boundary is within the threshold distance from the ground truth boundary, it is regarded…
Figure 8
Figure 8. Figure 8: Relation maps evolve from early to late epochs. During training, the relation module gradually captures key [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Evolution of relation maps and their statistical distributions over epochs, without and with the regularization term. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Visualization results on camouflaged object segmentation. [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Visualization results on leaf disease segmentation. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Visualization results on remote sensing road segmentation. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 46 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    X., Damianou, A., Lawrence, N

    Ahn, S., Hu, S. X., Damianou, A., Lawrence, N. D., and Dai, Z. Variational information distillation for knowledge transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9163--9171, 2019

    work page 2019

  3. [3]

    B it F it: Simple parameter-efficient fine-tuning for transformer-based masked language-models

    Ben Zaken, E., Goldberg, Y., and Ravfogel, S. B it F it: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pp.\ 1--9, 2022

    work page 2022

  4. [4]

    J., Fern \'a ndez-Esparrach, G., Gil, D., Rodr \' guez, C., and Vilari \ n o, F

    Bernal, J., S \'a nchez, F. J., Fern \'a ndez-Esparrach, G., Gil, D., Rodr \' guez, C., and Vilari \ n o, F. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics, 43: 0 99--111, 2015

    2015

  5. [5]

    Infinitely divisible matrices

    Bhatia, R. Infinitely divisible matrices. The American Mathematical Monthly, 113 0 (3): 0 221--235, 2006

    work page 2006

  6. [6]

    Convex optimization

    Boyd, S. Convex optimization. Cambridge UP, 2004

    work page 2004

  7. [7]

    Pkd: General distillation framework for object detectors via pearson correlation coefficient

    Cao, W., Zhang, Y., Gao, J., Cheng, A., Cheng, K., and Cheng, J. Pkd: General distillation framework for object detectors via pearson correlation coefficient. Advances in Neural Information Processing Systems, 35: 0 15394--15406, 2022

    work page 2022

  8. [8]

    All about structure: Adapting structural information across domains for boosting semantic segmentation

    Chang, W.-L., Wang, H.-P., Peng, W.-H., and Chiu, W.-C. All about structure: Adapting structural information across domains for boosting semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1900--1909, 2019

    work page 1900

Show all 55 references
  1. [9]

    Cross-layer distillation with semantic calibration

    Chen, D., Mei, J.-P., Zhang, Y., Wang, C., Wang, Z., Feng, Y., and Chen, C. Cross-layer distillation with semantic calibration. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 7028--7036, 2021 a

    2021

  2. [10]

    Distilling knowledge via knowledge review

    Chen, P., Liu, S., Zhao, H., and Jia, J. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 5008--5017, 2021 b

    2021

  3. [11]

    Adaptformer: Adapting vision transformers for scalable visual recognition

    Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., and Luo, P. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35: 0 16664--16678, 2022

    2022

  4. [12]

    Sam-adapter: Adapting segment anything in underperformed scenes

    Chen, T., Zhu, L., Deng, C., Cao, R., Wang, Y., Zhang, S., Li, Z., Sun, L., Zang, Y., and Mao, P. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 3367--3375, 2023

    2023

  5. [13]

    C., Gutman, D., Celebi, M

    Codella, N. C., Gutman, D., Celebi, M. E., Helba, B., Marchetti, M. A., Dusza, S. W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted ...

    2017

  6. [14]

    An efficient segment anything model for the segmentation of medical images

    Dong, G., Wang, Z., Chen, Y., Sun, Y., Song, H., Liu, L., and Cui, H. An efficient segment anything model for the segmentation of medical images. Scientific Reports, 14 0 (1): 0 19425, 2024

    2024

  7. [15]

    Optimal randomized approximations for matrix-based r \'e nyi’s entropy

    Dong, Y., Gong, T., Yu, S., and Li, C. Optimal randomized approximations for matrix-based r \'e nyi’s entropy. IEEE Transactions on Information Theory, 69 0 (7): 0 4218--4234, 2023

    2023

  8. [16]

    Camouflaged object detection

    Fan, D.-P., Ji, G.-P., Sun, G., Cheng, M.-M., Shen, J., and Shao, L. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 2777--2787, 2020 a

    2020

  9. [17]

    Pranet: Parallel reverse attention network for polyp segmentation

    Fan, D.-P., Ji, G.-P., Zhou, T., Chen, G., Fu, H., Shen, J., and Shao, L. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pp.\ 263--273. Springer, 2020 b

    2020

  10. [18]

    Computationally efficient approximations for matrix-based r \'e nyi's entropy

    Gong, T., Dong, Y., Yu, S., and Dong, B. Computationally efficient approximations for matrix-based r \'e nyi's entropy. IEEE Transactions on Signal Processing, 70: 0 6170--6184, 2022

    2022

  11. [19]

    J., and Tao, D

    Gou, J., Yu, B., Maybank, S. J., and Tao, D. Knowledge distillation: A survey. International Journal of Computer Vision, 129 0 (6): 0 1789--1819, 2021

    2021

  12. [20]

    Cycada: Cycle-consistent adversarial domain adaptation

    Hoffman, J., Tzeng, E., Park, T., Zhu, J.-Y., Isola, P., Saenko, K., Efros, A., and Darrell, T. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pp.\ 1989--1998. Pmlr, 2018

    1989

  13. [21]

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

    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. Proceedings of the International Conference on Learning Representations, pp.\ 1--20, 2022

    2022

  14. [22]

    Multi-level adversarial network for domain adaptive semantic segmentation

    Huang, J., Guan, D., Xiao, A., and Lu, S. Multi-level adversarial network for domain adaptive semantic segmentation. Pattern Recognition, 123: 0 108384, 2022

    2022

  15. [23]

    H., Riegler, M

    Jha, D., Smedsrud, P. H., Riegler, M. A., Halvorsen, P., De Lange, T., Johansen, D., and Johansen, H. D. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5--8, 2020, proceedings, part II 26, p...

    2020

  16. [24]

    Segment anything is not always perfect: An investigation of sam on different real-world applications, 2024

    Ji, W., Li, J., Bi, Q., Liu, T., Li, W., and Cheng, L. Segment anything is not always perfect: An investigation of sam on different real-world applications, 2024

    2024

  17. [25]

    Segment anything in high quality

    Ke, L., Ye, M., Danelljan, M., Tai, Y.-W., Tang, C.-K., Yu, F., et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024

    2024

  18. [26]

    Qr decomposition on gpus

    Kerr, A., Campbell, D., and Richards, M. Qr decomposition on gpus. In Proceedings of 2nd Workshop on General Purpose Processing on Graphics Processing Units, pp.\ 71--78, 2009

    2009

  19. [27]

    C., Lo, W.-Y., et al

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 4015--4026, 2023

    2023

  20. [28]

    Improving adversarial robustness via information bottleneck distillation

    Kuang, H., Liu, H., Wu, Y., Satoh, S., and Ji, R. Improving adversarial robustness via information bottleneck distillation. Advances in Neural Information Processing Systems, 36: 0 10796--10813, 2023

    2023

  21. [29]

    V., Nie, Z., Tran, M.-T., and Sugimoto, A

    Le, T.-N., Nguyen, T. V., Nie, Z., Tran, M.-T., and Sugimoto, A. Anabranch network for camouflaged object segmentation. Computer vision and image understanding, 184: 0 45--56, 2019

    2019

  22. [30]

    Invariant information bottleneck for domain generalization

    Li, B., Shen, Y., Wang, Y., Zhu, W., Li, D., Keutzer, K., and Zhao, H. Invariant information bottleneck for domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 7399--7407, 2022 a

    2022

  23. [31]

    A stepwise domain adaptive segmentation network with covariate shift alleviation for remote sensing imagery

    Li, J., Zi, S., Song, R., Li, Y., Hu, Y., and Du, Q. A stepwise domain adaptive segmentation network with covariate shift alleviation for remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 60: 0 1--15, 2022 b

    2022

  24. [32]

    Decomposition-based unsupervised domain adaptation for remote sensing image semantic segmentation

    Ma, X., Zhang, X., Ding, X., Pun, M.-O., and Ma, S. Decomposition-based unsupervised domain adaptation for remote sensing image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  25. [33]

    K., Manganelli, B., and Sa \`a -Garriga, A

    Miles, R., Yucel, M. K., Manganelli, B., and Sa \`a -Garriga, A. Mobilevos: Real-time video object segmentation contrastive learning meets knowledge distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10480--10490, 2023

    2023

  26. [34]

    Machine learning for aerial image labeling

    Mnih, V. Machine learning for aerial image labeling. University of Toronto (Canada), 2013

    2013

  27. [35]

    Probabilistic knowledge transfer for lightweight deep representation learning

    Passalis, N., Tzelepi, M., and Tefas, A. Probabilistic knowledge transfer for lightweight deep representation learning. IEEE Transactions on Neural Networks and Learning Systems, 32 0 (5): 0 2030--2039, 2020

    2020

  28. [36]

    Learning to adapt sam for segmenting cross-domain point clouds

    Peng, X., Chen, R., Qiao, F., Kong, L., Liu, Y., Sun, Y., Wang, T., Zhu, X., and Ma, Y. Learning to adapt sam for segmenting cross-domain point clouds. In European Conference on Computer Vision, pp.\ 54--71. Springer, 2024 a

    2024

  29. [37]

    Parameter efficient fine-tuning via cross block orchestration for segment anything model

    Peng, Z., Xu, Z., Zeng, Z., Xie, L., Tian, Q., and Shen, W. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3743--3752, 2024 b

    2024

  30. [38]

    Rath, S. R. Leaf disease segmentation dataset, 2023. URL https://www.kaggle.com/datasets/sovitrath/ leaf-disease-segmentation- Accessed: January 18, 2025

    2023

  31. [39]

    Sam 2: Segment anything in images and videos

    Ravi, N., Gabeur, V., Hu, Y.-T., Hu, R., Ryali, C., Ma, T., Khedr, H., R \"a dle, R., Rolland, C., Gustafson, L., et al. Sam 2: Segment anything in images and videos. Proceedings of the International Conference on Learning Representations, pp.\ 1--20, 2022

    2022

  32. [40]

    Fastsam3d: An efficient segment anything model for 3d volumetric medical images

    Shen, Y., Li, J., Shao, X., Inigo Romillo, B., Jindal, A., Dreizin, D., and Unberath, M. Fastsam3d: An efficient segment anything model for 3d volumetric medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp.\ 542--552. S...

    2024

  33. [41]

    Tinysam: Pushing the envelope for efficient segment anything model

    Shu, H., Li, W., Tang, Y., Zhang, Y., Chen, Y., Li, H., Wang, Y., and Chen, X. Tinysam: Pushing the envelope for efficient segment anything model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 20470--20478, 2025

    2025

  34. [42]

    Animal camouflage analysis: Chameleon database

    Skurowski, P., Abdulameer, H., B aszczyk, J., Depta, T., Kornacki, A., and Kozie , P. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2 0 (6): 0 7, 2018

    2018

  35. [43]

    Simada: A simple unified framework for adapting segment anything model in underperformed scenes

    Song, Y., Zhou, Q., Lu, X., Shao, Z., and Ma, L. Simada: A simple unified framework for adapting segment anything model in underperformed scenes. arXiv preprint arXiv:2401.17803, 2024

    2024 arXiv

  36. [44]

    and Zaslavsky, N

    Tishby, N. and Zaslavsky, N. Deep learning and the information bottleneck principle. In 2015 ieee information theory workshop (itw), pp.\ 1--5. IEEE, 2015

    2015

  37. [45]

    Samcl: Empowering sam to continually learn from dynamic domains

    Wang, Z., Ji, K., Wang, D., and Cheng, F. Samcl: Empowering sam to continually learn from dynamic domains. arXiv preprint arXiv:2412.05012, 2024

    2024

  38. [46]

    Medical sam adapter: Adapting segment anything model for medical image segmentation

    Wu, J., Wang, Z., Hong, M., Ji, W., Fu, H., Xu, Y., Xu, M., and Jin, Y. Medical sam adapter: Adapting segment anything model for medical image segmentation. Medical image analysis, 102: 0 103547, 2025

    2025

  39. [47]

    Cat-sam: Conditional tuning for few-shot adaptation of segment anything model

    Xiao, A., Xuan, W., Qi, H., Xing, Y., Ren, R., Zhang, X., Shao, L., and Lu, S. Cat-sam: Conditional tuning for few-shot adaptation of segment anything model. In European Conference on Computer Vision, pp.\ 189--206. Springer, 2025

    2025

  40. [48]

    Dirl: Domain-invariant representation learning for generalizable semantic segmentation

    Xu, Q., Yao, L., Jiang, Z., Jiang, G., Chu, W., Han, W., Zhang, W., Wang, C., and Tai, Y. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pp.\ 2884--2892, 2022

    2022

  41. [49]

    Yu, S., Giraldo, L. G. S., Jenssen, R., and Principe, J. C. Multivariate extension of matrix-based r \'e nyi's -order entropy functional. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42 0 (11): 0 2960--2966, 2019

    2019

  42. [50]

    U., Bae, S.-H., Lee, S., and Hong, C

    Zhang, C., Han, D., Qiao, Y., Kim, J. U., Bae, S.-H., Lee, S., and Hong, C. S. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023 a

    2023 arXiv

  43. [51]

    Blo-sam: Bi-level optimization based finetuning of the segment anything model for overfitting-preventing semantic segmentation

    Zhang, L., Liang, Y., Zhang, R., Javadi, A., and Xie, P. Blo-sam: Bi-level optimization based finetuning of the segment anything model for overfitting-preventing semantic segmentation. In Forty-first International Conference on Machine Learning, 2024 a

    2024

  44. [52]

    Distilling semantic priors from sam to efficient image restoration models

    Zhang, Q., Liu, X., Li, W., Chen, H., Liu, J., Hu, J., Xiong, Z., Yuan, C., and Wang, Y. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 25409--25419, 2024 b

    2024

  45. [53]

    Learning shape-invariant representation for generalizable semantic segmentation

    Zhang, Y., Tian, S., Liao, M., Hua, G., Zou, W., and Xu, C. Learning shape-invariant representation for generalizable semantic segmentation. IEEE Transactions on Image Processing, 32: 0 5031--5045, 2023 b

    2023

  46. [54]

    Convolution meets lo RA : Parameter efficient finetuning for segment anything model

    Zhong, Z., Tang, Z., He, T., Fang, H., and Yuan, C. Convolution meets lo RA : Parameter efficient finetuning for segment anything model. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=ezscMer8L0

    2024

  47. [55]

    Knowledge distillation by on-the-fly native ensemble

    Zhu, X., Gong, S., et al. Knowledge distillation by on-the-fly native ensemble. Advances in neural information processing systems, 31, 2018

    2018

Pith tools

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