Pith. sign in

REVIEW 4 major objections 5 minor 77 references

SAMwave: Wavelet-Driven Feature Enrichment for Effective Adaptation of Segment Anything Model

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

Pith's one-line read This paper claims that replacing fixed Fourier-domain masks with wavelet-extracted high-frequency subbands lets SAM and SAM2 adapt to low-level vision tasks more effectively than existing adapter methods.

desk verdict A useful wavelet-adapter recipe for SAM whose headline claim of consistent superiority is undercut by the paper's own Tables 3 and 5; send to review, but require claim revisions and code/error bars. read the letter →

arxiv 2507.20186 v1 pith:YLEKZKCY submitted 2025-07-27 cs.CV eess.IV

classification cs.CVeess.IV
keywords wavelettransformSegmentAnythingModeladapterfine-tuninghigh-frequencyfeaturescomplex-valuedadapterslow-levelvisioncamouflagedobjectdetectionshadow
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 sets out to improve how the Segment Anything Model (SAM) is adapted to low-level vision tasks that need fine-grained boundary detail, such as camouflage and shadow detection. It argues that existing adaptation methods, which feed high-frequency features selected by a fixed mask in the Fourier domain, give only marginal gains. SAMwave instead extracts high-frequency detail subbands with a wavelet transform, combines them into a feature map, and injects that map into SAM's encoder through lightweight adapters. The authors report that this consistently outperforms prior adaptation methods across four low-level vision tasks, with both SAM and SAM2 backbones, and that a complex-valued adapter variant gives further gains on some tasks.

What carries the argument

The load-bearing component is the Wavelet High-Frequency (WHF) module, which upsamples the input image by 2x, applies a 2D discrete wavelet transform, discards the low-frequency approximation band, and sums the three high-frequency detail subbands (LH, HL, HH) into IHF = Ilh + Ihl + Ihh. This map is projected to the embedding dimension and concatenated with SAM's patch embeddings; the concatenation is passed through per-layer adapters (real-valued MLPs, or complex-valued MLPs for the complex variant) whose outputs are added to each transformer layer's features. The paper's argument is that wavelet detail maps carry both spatial and frequency localization, which the Fourier-domain fixed mask lacks, letting the adapters select task-relevant structure.

What would settle it

On the ISTD shadow benchmark, using the Daubechies wavelet with the SAM2 backbone gives a balanced error rate of 1.91, worse than SAM-Adapter's 1.43; on SBU, the Coiflet wavelet with SAM2 gives 5.28, worse than several baselines. If a reader runs the same training setup with a different wavelet family and finds that no single wavelet consistently beats the Fourier-mask baselines across all four tasks, the paper's central claim of consistent superiority fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the choice and combination of high-frequency features used to guide a frozen foundation model matters more than the overall adapter architecture. A fixed Fourier-domain mask, even inverted, yields nearly identical performance, showing that such masks do not select task-relevant detail. Wavelet decomposition provides spatially localized, multi-scale high-frequency subbands (horizontal, vertical, diagonal), and summing these three detail bands into a single high-frequency map gives the encoder a richer and more adaptively usable signal. When this map is fed through per-layer adapters, SAM and SAM2 both improve substantially, and extending the adapters to complex-valued arithmetic with complex wavelets further improves results on shadow detection, defocus blur detection, and polyp detection.

Load-bearing premise

The premise that a fixed combination of three wavelet detail subbands with equal weights, after a fixed 2x upsampling, is the right high-frequency information for every low-level vision task, and that the choice of wavelet family does not change the qualitative conclusion.

Editorial extensions

If this is right

  • On the paper's evidence, SAMwave improves over SAM-Adapter and EVP on camouflage, shadow, blur, and polyp benchmarks with both SAM and SAM2 backbones.
  • The complex-valued adapter variant provides consistent gains on the SAM2 backbone, suggesting that complex wavelet features carry phase and shift-invariance information that plain real wavelets do not.
  • Because the encoder stays frozen and only adapters and the mask decoder are trained, the method is parameter-efficient and could be dropped into other SAM-based pipelines.
  • The wavelet decomposition is interpretable: the combined high-frequency subband is a fixed, explainable feature map, not a learned black-box prompt, which may help debugging and transfer.

Reading between the lines

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

  • The paper's own tables show that the best wavelet family varies by task and backbone; a natural extension would be to let the model learn per-image or per-task wavelet coefficients rather than fixing the combination to equal weights.
  • The same wavelet-enrichment idea could be applied to other vision foundation models beyond SAM, wherever boundary detail matters in dense prediction.
  • The finding that inverting the Fourier mask gives near-identical performance suggests that high-frequency features are only informative when spatially localized; this could be tested directly by comparing wavelet subbands against a learned spatial-frequency mask.
  • A concrete next experiment would be to see whether a learned, image-adaptive Fourier mask (rather than a fixed one) closes the gap with wavelet features, which would isolate whether wavelets' advantage is adaptivity or localization.
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 SAMwave, a parameter-efficient adaptation method for SAM and SAM2 that injects high-frequency information extracted via wavelet transforms through lightweight adapters. A Wavelet High-Frequency (WHF) module computes a composite high-frequency map from the LH, HL, and HH subbands of a resized input, and this map is combined with the encoder's patch embeddings and fed into per-layer adapters. The method is extended to complex-valued wavelets and complex-valued adapters. Experiments are reported on camouflaged object detection, shadow detection, defocus blur detection, and polyp detection, with additional forgery detection results in the supplementary material. The authors claim that SAMwave significantly and consistently outperforms existing SAM adaptation methods across both backbones and across real- and complex-valued variants.

Significance. If the central claim were fully supported, SAMwave would be a simple and valuable recipe: replacing rigid Fourier-mask prompting with wavelet-derived high-frequency features and supporting complex-valued adapters is a natural and interpretable design, and the paper covers four tasks and two backbone families. The paper's strengths include the subband ablation in Table 6, which substantiates the usefulness of combining all three high-frequency subbands for COD, and the demonstration that several wavelet choices improve over baselines on the main COD benchmark. However, the paper's own tables contradict the unqualified claim of consistent superiority, and the absence of a stated model-selection rule for the wavelet family plus the lack of error bars leaves the central claim only partially established. The method is potentially useful, but the reported evidence requires substantial qualification before the abstract's claims can be accepted.

major comments (4)
  1. [Abstract and Section 4.3, Table 3] The abstract's claim that SAMwave 'significantly outperforms existing adaptation methods' and that this 'superior performance is consistent across both the SAM and SAM2 backbones' is contradicted by the paper's own Table 3. On ISTD, the SAM2 backbone with the Daubechies wavelet gives BER 1.91, which is worse than SAM-Adapter's 1.43 and SAM2-Adapter's 1.43. On SBU, the SAM2 backbone with the Coiflet wavelet gives BER 5.28, which is worse than EVP's 4.31 and worse than most of the paper's own variants. Section 4.3 acknowledges only 'relatively smaller improvements' for some wavelets, but these cases are outright degradations, not smaller improvements. The central claim must be qualified to specific configurations, or an a priori selection rule must be stated and validated.
  2. [Section 4.3, Table 5] The polyp detection results also fail to support the claim of consistent superiority. In Table 5, several SAM-backbone real-valued adapter variants underperform the existing SAM-EG method by large margins: on ClinicDB, these variants achieve mDice between 0.652 and 0.866 versus SAM-EG's 0.931; on ColonDB, 0.635-0.743 versus 0.774; and on ETIS, 0.445-0.732 versus 0.757. Since SAM-EG is an existing SAM adaptation method, the blanket statement that SAMwave outperforms existing adaptation methods is empirically false for these configurations. The authors need to either restrict their claim to the best-performing wavelet/adapter choices or provide a clearly specified model-selection procedure, and then compare the selected model, not a set of partially failing variants.
  3. [Section 3.3 and Tables 2-5] The wavelet family is an uncontrolled free parameter that materially changes the results, yet the paper provides no rule for choosing it. For example, Table 3 shows SAM2/ISTD BER ranging from 0.92 (Symmetric-b) to 1.91 (Daubechies), and SAM2/SBU BER ranging from 2.99 (Haar) to 5.28 (Coiflet). The text in Section 4.4 states that 'using any wavelet provides better finer detail information than previous methods,' which is false for the shadow detection cases noted above. If the best wavelet is selected per task and per backbone after seeing the test set, the evaluation is post-hoc selection; if the wavelet is meant to be a fixed design choice, the paper must say which one is the recommended configuration. A robustness claim requires a validation-based selection protocol, not a post-hoc reading of test tables.
  4. [Section 4.3 and Tables 2-4] The paper repeatedly uses the word 'significant' without any statistical support. Many reported differences are extremely small, such as Table 2 CHAMELEON Sm values of 0.922 vs 0.923, or weighted F-measure values of 0.890 for both the Coiflet variant and a competing method. No error bars, multiple seeds, or significance tests are reported. Given that the claimed consistent superiority hinges on differences of this size, the authors should either provide variance estimates across at least three runs or soften the significance language to 'reported improvement' on the specific metrics.
minor comments (5)
  1. [Section 3.2 and Figure 2] Section 3.2 says the image encoder is frozen and 'we train only the parameters of the mask decoder,' but the method also trains the adapters and the Embedding Tune and WHF Tune modules. Figure 2's caption says 'we only updated the decoder for our training.' This is inconsistent with the architecture description in Section 3.4 and should be reworded to say that the pre-trained encoder backbone is frozen while the added adapters, embedding projection, WHF projection, and mask decoder are trained.
  2. [Section 4.2] The implementation details are too sparse for reproducibility: the paper omits learning rate, optimizer, batch size, number of epochs, input resolution details beyond image size, adapter hidden dimension (Dembed), and the number of trainable parameters. Since the paper claims efficiency, a parameter-count comparison against SAM-Adapter and EVP would be valuable.
  3. [Figure 1] Figure 1 is difficult to interpret. The caption says 'Learned weights for previous high frequency extraction method, compared with SAMwave's real and complex adapter weights,' but it does not specify which previous method the weights come from, how the weights are pooled or visualized, or why the concentration near zero is claimed to be disadvantageous. A precise description of the weight statistics and the comparison protocol should be added.
  4. [Table 4] For defocus blur detection, Table 4 compares against EVP but not against SAM-Adapter or SAM2-Adapter, and the text says 'we primarily compare our method with EVP' without explaining why the other two adapter baselines are absent. Since the abstract claims superiority over existing adaptation methods generally, the absence of two of the three main adapter baselines on this task weakens the comparison and should be addressed.
  5. [References] References [64] and [65] are the same paper (DeFusionNet) and should be merged or one removed. There are also minor typographical errors in the abstract ('However, Our analysis') and in Section 1 ('low-level tvision').

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SAMwave is evaluated on external benchmarks with standard train/test splits; self-citations are contextual, not load-bearing.

full rationale

The paper's central claim is empirical: SAMwave adapts SAM/SAM2 to four low-level vision tasks and reports test-set metrics against prior methods. The adapters and projection layers are trained on the task-specific training splits and evaluated on held-out test splits, which is standard supervised learning rather than a fitted input being renamed as a prediction. The WHF module combines fixed wavelet subbands (IHF = Ilh + Ihl + Ihh) with learnable adapters; no load-bearing theoretical derivation is claimed, so there is no self-definitional or ansatz-smuggling step that reduces the result to its inputs. Table 1's HFreq-versus-LFreq comparison is a motivation ablation, not a prediction, and its conclusion (fixed Fourier masks give small gains) is an empirical observation about their own setup, not a circular argument. The paper cites prior work by the same authors ([19], [74], and the self-reference [75]), but these citations are contextual (related work on camouflaged object detection and complex-valued networks) and are not used to justify the central adaptation claim or to forbid alternatives. No uniqueness theorem is imported from the authors' prior work. The main weakness is internal consistency: Tables 3 and 5 show several SAMwave wavelet/backbone variants underperforming SAM-Adapter, EVP, and SAM-EG, which contradicts the abstract's claim of consistent superiority across backbones and adapter variants. That is a correctness or overclaiming issue, not circularity, because the reported numbers are independent benchmark results rather than consequences of the method's definitions. Overall, the derivation chain is self-contained against external benchmarks, and no circular step is exhibited.

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

The central claim rests on standard DWT mathematics and on several domain assumptions about feature complementarity, none of which are derived. The main free choice is the wavelet family, which materially changes results. No new physical entities are introduced.

free parameters (3)
  • Wavelet family = varies: db, cf, ha, sym, sy-a, sy-b; no single default
    The method's performance depends on the chosen wavelet (e.g., ISTD BER ranges 0.92 to 2.60 across variants in Table 3). The paper does not fix one wavelet, making the 'method' a family of configurations.
  • Upsampling factor for WHF = 2
    Input is resized to 2x before DWT to compensate for downsampling; chosen by hand and never ablated.
  • Adapter hidden dimension / Dembed = not stated
    The projection dimension is mentioned as Dembed/gamma but its value is not reported; presumably borrowed from prior work.
assumptions (5)
  • domain assumption Wavelet high-frequency subbands capture useful boundary/detail cues for dense prediction
    The entire motivation relies on this; not proven analytically.
  • domain assumption Equal-weight sum of LH, HL, HH subbands is a suitable high-frequency representation
    IHF = Ilh + Ihl + Ihh; no learned weighting or ablation of different combination weights.
  • domain assumption The adapter architecture and training pipeline from SAM-Adapter/EVP transfer correctly to all four tasks
    They 'follow [7,37]' for training; no per-task tuning of the base pipeline is detailed.
  • domain assumption SAM/SAM2 encoder features are frozen and remain directly compatible with injected wavelet features
    The authors assume the pre-trained feature space can be additively modified by P(i) without re-training.
  • domain assumption Complex-valued neural network operations are correctly implemented and trained
    Complex adapters rely on [66,74]; details like split GeLU are referenced, not fully specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAMwave: Wavelet-Driven Feature Enrichment for Effective Adaptation of Segment Anything Model." pith.science (2026). https://pith.science/paper/YLEKZKCY

@misc{pith2026250720186,
  author       = {Pith},
  title        = {Pith review of: SAMwave: Wavelet-Driven Feature Enrichment for Effective Adaptation of Segment Anything Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLEKZKCY}},
  note         = {Machine review of arXiv:2507.20186}
}
read the original abstract

The emergence of large foundation models has propelled significant advances in various domains. The Segment Anything Model (SAM), a leading model for image segmentation, exemplifies these advances, outperforming traditional methods. However, such foundation models often suffer from performance degradation when applied to complex tasks for which they are not trained. Existing methods typically employ adapter-based fine-tuning strategies to adapt SAM for tasks and leverage high-frequency features extracted from the Fourier domain. However, Our analysis reveals that these approaches offer limited benefits due to constraints in their feature extraction techniques. To overcome this, we propose \textbf{\textit{SAMwave}}, a novel and interpretable approach that utilizes the wavelet transform to extract richer, multi-scale high-frequency features from input data. Extending this, we introduce complex-valued adapters capable of capturing complex-valued spatial-frequency information via complex wavelet transforms. By adaptively integrating these wavelet coefficients, SAMwave enables SAM's encoder to capture information more relevant for dense prediction. Empirical evaluations on four challenging low-level vision tasks demonstrate that SAMwave significantly outperforms existing adaptation methods. This superior performance is consistent across both the SAM and SAM2 backbones and holds for both real and complex-valued adapter variants, highlighting the efficiency, flexibility, and interpretability of our proposed method for adapting segment anything models.

Figures

Figures reproduced from arXiv: 2507.20186 by the authors.

Figure 1
Figure 1. Learned weights for previ￾ous high frequency extraction method, compared with SAMwave’s real and complex adapter weights. Unlike pre￾vious method, SAMwave’s weights are not concentrated around 0. In this work, we investigate the optimal adaptation of large models to related tasks, focusing particularly on SAM, which comprises of a powerful encoder and a prompt-conditioned decoder. We hypothesize that the encoder can… view at source ↗
Figure 2
Figure 2. Overview of SAMwave for low-level vision tasks. Right shows that the image is passed through SAM and our WHF (wavelet high frequency) module, where the high-frequency features are extracted. The extracted details are then combined with patch embeddings obtained from the trans￾former encoder. Note that the encoder is frozen; we only updated the decoder for our training. Left shows the overview of a single adapter use… view at source ↗
Figure 3
Figure 3. An overview of the WHF module, we first use the wavelet transform to obtain high- and low-frequency features. Then, we combine the high-frequency features to create a feature map, which is then used to provide finer details in the image. The Wavelet High-Frequency (WHF) mod￾ule extracts rich high-frequency spatial fea￾tures using the Discrete Wavelet Transform (DWT). Given an input image I ∈ R H×W×3 , we first resiz… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of our proposed SAMwave with existing methods for the COD task. Shadow Detection: We compare the performance of our approach against existing methods across two benchmark datasets. In both cases, our method yields substantial performance improvements.…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 60 canonical work pages

  1. [1]

    A data-free approach to mitigate catastrophic forgetting in federated class incremental learning for vision tasks

    Sara Babakniya, Zalan Fabian, Chaoyang He, Mahdi Soltanolkotabi, and Salman Avestimehr. A data-free approach to mitigate catastrophic forgetting in federated class incremental learning for vision tasks. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Advances in Neural Informa- tion Processing Systems 3...

  2. [2]

    WM-DOV A maps for accurate polyp highlighting in colonoscopy: Validation vs

    Jorge Bernal, Francisco Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez de Miguel, and Fernando Vilariño. WM-DOV A maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians.Com- put. Medical Imaging Graph. , 43:99–111, 2015. doi: 10.1016/J.COMPMEDIMAG. 2015.02.007. URL https://doi.org/10.1...

  3. [4]

    Solving the catastrophic forgetting problem in generalized category discovery

    Xinzi Cao, Xiawu Zheng, Guanhong Wang, Weijiang Yu, Yunhang Shen, Ke Li, Yutong Lu, and Yonghong Tian. Solving the catastrophic forgetting problem in generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16880–16889, June 2024

  4. [5]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kir- illov, and Sergey Zagoruyko. End-to-end object detection with transformers. In An- drea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Com- puter Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23- 28, 2020, Proceedings, Part I , vol...

  5. [6]

    Saving 100x storage: Prototype replay for reconstructing training sample distribution in class-incremental semantic segmentation

    Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho-Shing Ip, and Sam Kwong. Saving 100x storage: Prototype replay for reconstructing training sample distribution in class-incremental semantic segmentation. In Alice Oh, Tristan Naumann, Amir Glober- son, Kate Saenko, Moritz Hardt, and Sergey Levine, editors,Advances in Neural Infor- mation Processing Systems...

  6. [7]

    Sam fails to segment any- thing? – sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more, 2023

    Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Shangzhan Zhang, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment any- thing? – sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more, 2023

  7. [8]

    Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image seg- mentation, and more, 2024

    Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image seg- mentation, and more, 2024. URL https://arxiv.org/abs/2408.04579

  8. [9]

    Vision transformer adapter for dense predictions

    Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. InThe Eleventh International Confer- ence on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. Open- Review.net, 2023. URL https://openreview.net/pdf?id=plKu2GByCNW

Show all 77 references
  1. [10]

    A multi-task mean teacher for semi-supervised shadow detection

    Zhihao Chen, Lei Zhu, Liang Wan, Song Wang, Wei Feng, and Pheng-Ann Heng. A multi-task mean teacher for semi-supervised shadow detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pages 5610–5619. Co...

  2. [11]

    Schwing, and Alexander Kirillov

    Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. 2021

  3. [12]

    Image splicing localization via semi-global network and fully connected conditional random fields

    Xiaodong Cun and Chi-Man Pun. Image splicing localization via semi-global network and fully connected conditional random fields. In Laura Leal-Taixé and Stefan Roth, editors, Computer Vision - ECCV 2018 Workshops - Munich, Germany, September 8- 14, 2018, Proceedings, Part II, ...

  4. [13]

    Kroese, Shie Mannor, and Reuven Y

    Pieter-Tjerk de Boer, Dirk P. Kroese, Shie Mannor, and Reuven Y . Rubin- stein. A tutorial on the cross-entropy method. Ann. Oper. Res. , 134(1):19–67,

  5. [14]

    Casia image tampering detection evaluation database

    Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In 2013 IEEE China Summit and International Conference on Signal and Information Processing, pages 422–426, 2013. doi: 10.1109/ChinaSIP.2013.6625374. 12 Y ADA V , GUPTA, JERRIPOTHULA: SAMW A VE

  6. [15]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Trans- formers for image recognitio...

  7. [16]

    Structure-measure: A new way to evaluate foreground maps

    Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017

  8. [17]

    Camouflaged object detection

    Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

  9. [18]

    Fridrich and Jan Kodovský

    Jessica J. Fridrich and Jan Kodovský. Rich models for steganalysis of digital im- ages. IEEE Trans. Inf. Forensics Secur., 7(3):868–882, 2012. doi: 10.1109/TIFS.2012. 2190402. URL https://doi.org/10.1109/TIFS.2012.2190402

  10. [19]

    CIRCOD: co-saliency inspired referring camouflaged object discovery

    Avi Gupta, Koteswar Rao Jerripothula, and Tammam Tillo. CIRCOD: co-saliency inspired referring camouflaged object discovery. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025, pages 8313–8323. IEEE, 2025. ...

  11. [21]

    Selective amnesia: A continual learning approach to forgetting in deep generative models

    Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Advances in Neural Informa- tion Processing Systems 36: Ann...

  12. [22]

    Parameter- efficient transfer learning for NLP

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter- efficient transfer learning for NLP. In Kamalika Chaudhuri and Ruslan Salakhutdi- nov, editors, Proceedings of the 36th In...

  13. [23]

    Direction- aware spatial context features for shadow detection

    Xiaowei Hu, Lei Zhu, Chi-Wing Fu, Jing Qin, and Pheng-Ann Heng. Direction- aware spatial context features for shadow detection. In 2018 IEEE Confer- ence on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake Y ADA V , GUPTA, JERRIPOTHULA: SAMW A VE 13 City, UT, USA,...

  14. [24]

    SPAN: spatial pyramid attention network for image manipulation localization

    Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. SPAN: spatial pyramid attention network for image manipulation localization. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 ...

  15. [25]

    Feature shrinkage pyramid for camouflaged object detection with trans- formers

    Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with trans- formers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17...

  16. [26]

    Smedsrud, Michael A

    Debesh Jha, Pia H. Smedsrud, Michael A. Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D. Johansen. Kvasir-seg: A segmented polyp dataset. In Yong Man Ro, Wen-Huang Cheng, Junmo Kim, Wei-Ta Chu, Peng Cui, Jung- Woo Choi, Min-Chun Hu, and Wesley De Neve, edit...

  17. [27]

    Ali Karaali and Cláudio R. Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Trans. Image Process., 27(3):1126–1137, 2018. doi: 10.1109/ TIP.2017.2771563. URL https://doi.org/10.1109/TIP.2017.2771563

  18. [29]

    Murat Tekalp, and Zafer Dogan

    Cansu Korkmaz, A. Murat Tekalp, and Zafer Dogan. Training generative image super- resolution models by wavelet-domain losses enables better control of artifacts. CoRR, abs/2402.19215, 2024. doi: 10.48550/ARXIV .2402.19215. URL https://doi. org/10.48550/arXiv.2402.19215

  19. [30]

    Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugi- moto

    Trung-Nghia Le, Tam V . Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugi- moto. Anabranch network for camouflaged object segmentation. Journal of Computer Vision and Image Understanding, 184:45–56, 2019. 14 Y ADA V , GUPTA, JERRIPOTHULA: SAMW A VE

  20. [32]

    Uncertainty-aware joint salient object and camouflaged object detection

    Aixuan Li, Jing Zhang, Yunqiu Lyu, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

  21. [33]

    Ailurus: A scalable vit framework for dense prediction

    Jin Li, Yaoming Wang, Xiaopeng Zhang, Bowen Shi, Dongsheng Jiang, Chenglin Li, Wenrui Dai, Hongkai Xiong, and Qi Tian. Ailurus: A scalable vit framework for dense prediction. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, A...

  22. [35]

    Girshick, and Kaiming He

    Yanghao Li, Hanzi Mao, Ross B. Girshick, and Kaiming He. Exploring plain vi- sion transformer backbones for object detection. In Shai Avidan, Gabriel J. Bros- tow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors, Com- puter Vision - ECCV 2022 - 17th Europea...

  23. [36]

    Jiaying Lin, Xin Tan, Ke Xu, Lizhuang Ma, and Rynson W. H. Lau. Frequency-aware camouflaged object detection. ACM Trans. Multim. Comput. Commun. Appl. , 19(2): 61:1–61:16, 2023. doi: 10.1145/3545609. URL https://doi.org/10.1145/ 3545609

  24. [37]

    Explicit visual prompt- ing for low-level structure segmentations

    Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompt- ing for low-level structure segmentations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19434–19445. IEEE, 2023. doi: 1...

  25. [38]

    Pscc-net: Progressive spatio- channel correlation network for image manipulation detection and localization

    Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pscc-net: Progressive spatio- channel correlation network for image manipulation detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 2022. Y ADA V , GUPTA, JERRIPOTHULA: SAMW A VE 15

  26. [39]

    Darenerf: Direction-aware representation for dynamic scenes

    Ange Lou, Benjamin Planche, Zhongpai Gao, Yamin Li, Tianyu Luan, Hao Ding, Ter- rence Chen, Jack Noble, and Ziyan Wu. Darenerf: Direction-aware representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages...

  27. [40]

    Prompt guided transformer for multi-task dense prediction

    Yuxiang Lu, Shalayiding Sirejiding, Yue Ding, Chunlin Wang, and Hongtao Lu. Prompt guided transformer for multi-task dense prediction. IEEE Trans. Multim., 26: 6375–6385, 2024. doi: 10.1109/TMM.2024.3349865. URL https://doi.org/ 10.1109/TMM.2024.3349865

  28. [42]

    Simultaneously localize, segment and rank the camouflaged objects

    Yunqiu Lyu, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng- Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

  29. [43]

    How to evaluate foreground maps? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2014

    Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2014

  30. [44]

    Camouflaged object segmentation with distraction mining

    Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, vir- tual, June 19-25, 2021 , pages 8772–8781. Computer Vision Found...

  31. [45]

    Camouflaged object segmentation with distraction mining

    Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 8772–8781, June 2021

  32. [46]

    Imd2020: A large-scale annotated dataset tailored for detecting manipulated images

    Adam Novozámský, Babak Mahdian, and Stanislav Saic. Imd2020: A large-scale annotated dataset tailored for detecting manipulated images. In 2020 IEEE Winter Applications of Computer Vision Workshops (WACVW), pages 71–80, 2020. doi: 10. 1109/W ACVW50321.2020.9096940

  33. [47]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Ra...

  34. [49]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila an...

  35. [52]

    Zoom in and out: A mixed-scale triplet network for camouflaged object detection

    Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2160–2170, June 2022

  36. [53]

    Gir- shick, Piotr Dollár, and Christoph Feichtenhofer

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloé Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Gir- shick, Piotr Dollár, and Christoph Feic...

  37. [54]

    URL http://proceedings.mlr.press/v139/radford21a.html

  38. [55]

    Das, and Ulas Bagci

    Vanshali Sharma, Abhishek Kumar, Debesh Jha, Manas Kamal Bhuyan, Pradip K. Das, and Ulas Bagci. Controlpolypnet: Towards controlled colon polyp synthesis for improved polyp segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, ...

  39. [56]

    Discriminative blur detection features

    Jianping Shi, Li Xu, and Jiaya Jia. Discriminative blur detection features. In2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014 , pages 2965–2972. IEEE Computer Society, 2014. doi: 10. 1109/CVPR.2014.379. URL https:...

  40. [57]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenh...

  41. [58]

    Continual learning with deep generative replay

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V . N. Vishwanathan, and Roman Garnett, editors, Advances in Neural Information Processin...

  42. [59]

    Baraniuk

    Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veer- araghavan, and Richard G. Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18507–18516, J...

  43. [60]

    Animal camouflage analysis: Chameleon database

    Przemysław Skurowski, Hassan Abdulameer, J Błaszczyk, Tomasz Depta, Adam Kor- nacki, and P Kozieł. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018

  44. [61]

    Over- coming catastrophic forgetting for multi-label class-incremental learning

    Xiang Song, Kuang Shu, Songlin Dong, Jie Cheng, Xing Wei, and Yihong Gong. Over- coming catastrophic forgetting for multi-label class-incremental learning. InIEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2024, Waikoloa, HI, USA, January 3-8, 2024, pages 2...

  45. [62]

    Just noticeable defocus blur detection and estima- tion

    Jianping Shi, Li Xu, and Jiaya Jia. Just noticeable defocus blur detection and estima- tion. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015 , pages 657–665. IEEE Computer Society, 2015. doi: 10.1109/CVPR.2015.7298665....

  46. [63]

    Gurudu, and Jianming Liang

    Nima Tajbakhsh, Suryakanth R. Gurudu, and Jianming Liang. Automated polyp de- tection in colonoscopy videos using shape and context information. IEEE Trans. Medical Imaging , 35(2):630–644, 2016. doi: 10.1109/TMI.2015.2487997. URL https://doi.org/10.1109/TMI.2015.2487997. 18 Y...

  47. [64]

    Toward embedded detection of polyps in wce images for early di- agnosis of colorectal cancer

    Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early di- agnosis of colorectal cancer. International journal of computer assisted radi- ology and surgery , 9(2):283—293, March 2014. ISSN 186...

  48. [65]

    Chang Tang, Xinzhong Zhu, Xinwang Liu, Lizhe Wang, and Albert Y . Zomaya. Defusionnet: Defocus blur detection via recurrently fusing and refining multi-scale deep features. In IEEE Conference on Computer Vision and Pattern Recogni- tion, CVPR 2019, Long Beach, CA, USA, June 16...

  49. [66]

    Chiheb Trabelsi, Olexa Bilaniuk, Ying Zhang, Dmitriy Serdyuk, Sandeep Subrama- nian, João Felipe Santos, Soroush Mehri, Negar Rostamzadeh, Yoshua Bengio, and Christopher J. Pal. Deep complex networks. In 6th International Conference on Learn- ing Representations, ICLR 2018, Va...

  50. [67]

    Gilbert Strang and Truong Q. Nguyen. Wavelets and filter banks. Wellesley-Cambridge Press, 1997. ISBN 978-0-9614088-7-9

  51. [68]

    Sam-eg: Segment anything model with egde guidance framework for efficient polyp segmentation

    Quoc-Huy Trinh, Hai-Dang Nguyen, Nguyen Ngoc Bao Tram, Debesh Jha, Ulas Bagci, and Minh-Triet Tran. Sam-eg: Segment anything model with egde guidance framework for efficient polyp segmentation. In 35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25...

  52. [70]

    T. F. Y . Vicente, Le Hou, Chen-Ping Yu, Minh Hoai, and Dimitris Samaras. Large-scale training of shadow detectors with noisily-annotated shadow examples. In European Conference on Computer Vision, 2016. URL https://api.semanticscholar. org/CorpusID:17623309. Y ADA V , GUPTA, ...

  53. [71]

    Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal

    Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

  54. [72]

    M2unet: Metaformer multi- scale upsampling network for polyp segmentation

    Quoc-Huy Trinh, Nhat-Tan Bui, Trong-Hieu Nguyen Mau, Minh-Van Nguyen, Hai- Minh Phan, Minh-Triet Tran, and Hai-Dang Nguyen. M2unet: Metaformer multi- scale upsampling network for polyp segmentation. In31st European Signal Processing Conference, EUSIPCO 2023, Helsinki, Finland,...

  55. [73]

    Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anoma- lous features

    Yue Wu, Wael AbdAlmageed, and Premkumar Natarajan. Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anoma- lous features. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2...

  56. [74]

    Fccns: Fully complex-valued convo- lutional networks using complex-valued color model and loss function

    Saurabh Yadav and Koteswar Rao Jerripothula. Fccns: Fully complex-valued convo- lutional networks using complex-valued color model and loss function. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 10655–10664. IEEE, ...

  57. [75]

    Multi-task learning for dense prediction tasks: A survey

    Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE Trans. Pattern Anal. Mach. Intell. , 44(7):3614–3633, 2022. doi: 10.1109/TPAMI.2021.3054719. URL https://...

  58. [76]

    Camoformer: Masked separable attention for camouflaged object detection

    Bowen Yin, Xuying Zhang, Qibin Hou, Bo-Yuan Sun, Deng-Ping Fan, and Luc Van Gool. Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570, 2022

  59. [77]

    Defocus blur detec- tion via multi-stream bottom-top-bottom fully convolutional network

    Wenda Zhao, Fan Zhao, Dong Wang, and Huchuan Lu. Defocus blur detec- tion via multi-stream bottom-top-bottom fully convolutional network. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 , pages 3080–3088....

  60. [78]

    Objectformer for image manipulation detection and localization

    Junke Wang, Zuxuan Wu, Jingjing Chen, Xintong Han, Abhinav Shrivastava, Ser- Nam Lim, and Yu-Gang Jiang. Objectformer for image manipulation detection and localization. In IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, CVPR 2022, New Orleans, LA, USA, June 1...

  61. [79]

    Defocus blur detection via boosting diversity of deep ensemble networks

    Wenda Zhao, Xueqing Hou, You He, and Huchuan Lu. Defocus blur detection via boosting diversity of deep ensemble networks. IEEE Trans. Image Process. , 30: 5426–5438, 2021. doi: 10.1109/TIP.2021.3084101. URL https://doi.org/10. 1109/TIP.2021.3084101

  62. [80]

    Self-generated defocus blur detection via dual adversarial discriminators

    Wenda Zhao, Cai Shang, and Huchuan Lu. Self-generated defocus blur detection via dual adversarial discriminators. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021 , pages 6933–6942. Computer Vision Foundation / IEEE, 2021. doi...

  63. [81]

    SAMwave: wavelet- driven feature enrichment for effective adaptation of segment anything model

    Saurabh Yadav, Avi Gupta, and Koteswar Rao Jerripothula. SAMwave: wavelet- driven feature enrichment for effective adaptation of segment anything model. In 36th British Machine Vision Conference, BMVC 2025, Sheffield, UK, November 24- 27, 2025. BMV A Press, 2025

  64. [83]

    Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson W. H. Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 4682–4691. IEEE...

  65. [84]

    Enhancing di- versity of defocus blur detectors via cross-ensemble network

    Wenda Zhao, Bowen Zheng, Qiuhua Lin, and Huchuan Lu. Enhancing di- versity of defocus blur detectors via cross-ensemble network. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pages 8905–8913. Computer Vi- sion...

  66. [87]

    Quanlong Zheng, Xiaotian Qiao, Ying Cao, and Rynson W. H. Lau. Distraction-aware shadow detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pages 5167–5176. Computer Vision Foundation / IEEE, 2019. doi: 1...

  67. [1119]

    doi: 10.23919/EUSIPCO58844.2023.10290110

    IEEE, 2023. doi: 10.23919/EUSIPCO58844.2023.10290110. URL https: //doi.org/10.23919/EUSIPCO58844.2023.10290110

  68. [2005]

    URL https://doi.org/10.1007/ s10479-005-5724-z

    doi: 10.1007/S10479-005-5724-Z. URL https://doi.org/10.1007/ s10479-005-5724-z

  69. [2021]

    URL https://openreview.net/forum?id=YicbFdNTTy

Pith tools

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