Pith. sign in

REVIEW 4 major objections 4 minor 58 references

ZMIS-SAM: Segment Anything Model Enhanced with Wavelet Transform for Zooplankton Microscopy Image Instance Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read This paper claims that a Segment Anything Model outfitted with two lightweight adapters, a neighboring-feature aggregator, and a wavelet-based high-frequency enhancer achieves state-of-the-art instance segmentation on zooplankton microscopy

desk verdict Dataset is the real contribution; the SOTA claim is inflated by near-duplicate test images and test-set hyperparameter selection. read the letter →

arxiv 2607.27585 v1 pith:UI446OTB submitted 2026-07-30 cs.CV

classification cs.CV
keywords zooplanktoninstancesegmentationSegmentAnythingModelwavelettransformmicroscopyparameter-efficientfine-tuningfeatureaggregationZMIS5K
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

This paper tries to establish that a Segment Anything Model can be adapted to zooplankton microscopy instance segmentation—a domain where generic SAM fails—by adding three targeted modules: two adapters that teach the encoder zooplankton shape and intensity statistics, a feature aggregation module that stitches body and appendage features together, and a wavelet module that restores high-frequency boundary details. On a new 47-species dataset of 5,358 images, the resulting model reaches 73.6 mean average precision, beating all fourteen comparison models. If right, it provides the first SAM-based state-of-the-art baseline for zooplankton segmentation and evidence that small, domain-specific add-ons can close the gap between a generalist foundation model and microscopy imagery.

What carries the argument

The load-bearing machinery is a set of three trainable modules around SAM's frozen image encoder: (1) ZM-ViT—two parameter-efficient adapters, a Shape Adapter using striped depthwise convolutions and an Intensity Adapter using a Conv–GELU–BN parallel branch, inserted into alternate ViT layers from layer 8 onward; (2) NFAM—a feature aggregation block that processes domain-specific, general, and concatenated features in parallel, splits them into n=4 parts, and passes each through multi-scale gated convolutions with receptive fields 3, 5, and 7; (3) WM2FE—a Haar-wavelet decomposition of aggregated features whose LL/LH/HL/HH subbands modulate n feature parts, with learnable fusion weights and c

What would settle it

Retrain ZMIS-SAM exactly as described but with a stricter duplicate-removal step (e.g., dropping any train/test pair with SSIM > 0.90) and with hyperparameters fixed before ever looking at test labels; if the mAP advantage over the second-best method falls from the reported 1.8 points to near zero, the claimed state-of-the-art margin was an artifact of leakage or test-set-driven selection rather than the modules.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that SAM's failure on zooplankton microscopy—misclassification of similar species, fragmented appendages, incomplete transparent boundaries—can be traced to three specific encoder-level deficiencies and that each has a targeted fix. ZM-ViT inserts a Shape Adapter and an Intensity Adapter into selected ViT blocks so the frozen SAM encoder learns zooplankton morphology and microscopy illumination patterns; NFAM fuses those domain-specific features with general frozen-layer features so slender appendages stay connected; WM2FE applies a Haar discrete wavelet decomposition to multi-scale features and dynamically recombines frequency bands, recove

Load-bearing premise

The reported 73.6 mAP is only trustworthy if the near-duplicate images kept in the test split (SSIM above 0.95) really are harmless illumination variants, and if choosing the Haar wavelet and the n=4 split using test-set metrics did not inflate the result.

Editorial extensions

If this is right

  • - On ZMIS5K, ZMIS-SAM reaches 73.6 mAP, 94.6 AP50, and 80.7 AP75, surpassing every comparison model and the best SAM-based baseline by 1.8 mAP.
  • - Each module earns its place: replacing the baseline encoder with ZM-ViT adds 1.3 mAP; adding NFAM adds another 1.3; adding WM2FE adds 1.0; removing any single module lowers performance.
  • - The Haar wavelet outperforms rbio1.1 and bior1.1 for boundary recovery, and n=4 multi-scale blocks is the best split for NFAM.
  • - Retrained on underwater and remote-sensing benchmarks, the model improves mAP over the underwater-adapted baseline (34.1 vs 16.1 on NWPU; 22.5 vs 22.0 on UIIS), supporting the claim of cross-domain generalization.
  • - The new ZMIS5K dataset itself is a contribution: 5,358 pixel-level annotated images, 47 species, 10,228 instances, with category, mask, and box labels supporting detection, semantic, and instance segmentation; the framework stops at whole-instance masks and cannot separately segment heads and antennae.

Reading between the lines

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

  • - If the modules are as transferable as the cross-domain tables suggest, the same recipe—adapter-based domain knowledge plus wavelet high-frequency recovery—offers a template for adapting SAM to other microscopy domains where transparency and fine protrusions matter.
  • - The paper's own comparison shows WM2FE alone improves a prior underwater SAM baseline by 0.8 mAP; a natural editorial extension is that WM2FE could be extracted and reused as a standalone boundary-refinement component in other SAM-based segmenters, independent of the zooplankton dataset.
  • - Because deployment is offline at 3.9 FPS, the immediate practical use is automated batch analysis in the lab; a faster backbone or distilled variant would be needed before this could run on an in-situ plankton imager.
  • - The explicit limitation—no separation of head and antennae—suggests the next testable step: adding text or taxonomic labels to prompt the decoder for part-level masks, which the current single-class instance segmentation cannot express.
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 / 4 minor

Summary. The paper proposes ZMIS-SAM, a SAM-based instance segmentation model for zooplankton microscopy images, and introduces ZMIS5K, a new dataset of 5,358 images across 47 species with 10,228 instances. The method adds three components to SAM: ZM-ViT (Shape and Intensity Adapters for domain adaptation), NFAM (neighboring feature aggregation for slender appendages), and WM2FE (wavelet-based multi-scale/multi-directional enhancement for boundary details). Experiments report state-of-the-art performance on ZMIS5K (Table 1: mAP 73.6, AP50 94.6, AP75 80.7) and cross-domain results on UIIS and NWPU (Table 10). The paper also includes ablations in Tables 2, 3, 5, 6, and C.2.

Significance. If the reported results are sound, this is a useful contribution to domain adaptation of SAM and provides the first large-scale zooplankton microscopy instance segmentation dataset. The cross-domain evaluations on UIIS and NWPU are a strength, and the paper is generally clearly written. However, the validity of the central SOTA claim depends on the cleanliness of the test set and on the model-selection protocol, both of which are currently questionable. The paper does not include machine-checked proofs or released code/dataset links, but the empirical comparisons and ablations are extensive.

major comments (4)
  1. [Appendix A.2] The SSIM-based deduplication described in A.2 appears to identify near-duplicate train/test pairs but does not remove them. The sentence 'Samples with SSIM above 0.95 ... conform to practical real-world imaging conditions' suggests such pairs are retained. If near-duplicate images remain in both training and test sets, the 400-epoch trained model may memorize them, inflating the reported mAP. The SOTA claim (Table 1) requires a clean, unseen test set. Please re-split with strict removal of near-duplicates, report results on a cleaned subset, or quantitatively show that retaining these pairs does not affect mAP.
  2. [Tables 3 and C.2] The wavelet basis (Table 3) and the NFAM split count n (Table C.2) are selected using ZMIS5K test-set metrics. No validation split is described in Section 4.1. This makes the reported 73.6 mAP an optimistic, test-set-driven estimate. Model selection should be performed on a held-out validation split, with the test set used only once for final evaluation. Please also report test metrics for all hyperparameter choices so readers can see the selection bias.
  3. [Eq. (9), Section 3.3] The gating formula in Eq. (9) is x_gate_k = ψ_k * x + (1 - ψ_k) * x, which simplifies to x. Thus the gating mechanism described in MConvBlock is the identity operation and does not dynamically weight features across receptive fields. This undermines the architectural claim for NFAM and the ablation in Table 2. Please correct the equation to the intended gating operation, re-run the ablations, and state whether the corrected gating changes the results.
  4. [Tables 2, 5, 6] No error bars or multiple-seed runs are reported. The incremental improvements are small (e.g., +1.3, +1.0, +0.8 mAP in Tables 2, 5, 6), and the SOTA margin over RSPrompter is only 1.8 mAP in Table 1. Without variance estimates or significance tests, the claimed improvements and the SOTA margin may be within run-to-run noise. Please report at least three seeds with mean±std, or a paired significance test.
minor comments (4)
  1. [Abstract] The line 'Code: ZMIS-SAM' is not a usable URL. Please provide a public link to code and dataset, or state clearly how they will be released.
  2. [Eq. (1), Section 3.2] In Eq. (1), the output uses 'Fin' but the residual term is not defined; it should likely be 'x'. Please clarify. Also, Eq. (2) uses the notation 'sum_{max;avg}' which is unclear; define the operation explicitly.
  3. [Section C.4 / Table 8] The text says each SA/IA/NFAM has 1.65M/1.23M/4.65M parameters, while Table 8 reports 21.4M/32.0M/60.0M. These are consistent if the table gives totals across all inserted modules (13 SA, 26 IA, 13 NFAM), but the table should say so. Also, the text after Table 8 mentions 'AP70' which is likely a typo for AP75.
  4. [Appendix C.2] The statement that n=6 'falls within the range of normal variation' acknowledges run-to-run variability, but no variance is reported. This reinforces the need for multiple seeds (see Major Comment 4).

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation-level circularity; reported SOTA is an empirical benchmark result, though test-set-driven module selection and SSIM-based near-duplicate handling are benchmark-hygiene concerns.

full rationale

The paper's central claim — that ZMIS-SAM achieves 73.6 mAP on ZMIS5K — is an evaluated model comparison, not a quantity derived from its own assumptions. The dataset split (4,262 training / 1,096 test images) is stated, the model is trained on the training split and evaluated on the held-out test split, and cross-domain experiments on UIIS and NWPU provide external empirical anchors. Ablations, feature visualizations, and comparisons against 14 baselines are all empirical rather than claimed as first-principles derivations. The only self-citation of note (PlanktonTNet, ref. [58]) is used contextually for intelligent identification approaches and is not load-bearing for the architecture or the SOTA claim. The two genuinely weak points are benchmark-hygiene issues, not circularity: Appendix A.2 reports that train/test near-duplicates with SSIM > 0.95 were treated as acceptable illumination/focal-plane variants without stating that they were removed, and Tables 3 and C.2 select the Haar wavelet and the NFAM split count n=4 using ZMIS5K test-set metrics, so the reported 73.6 mAP may be optimistically biased for unseen data. However, this is a soundness/leakage concern, not a reduction of the result to its inputs by construction. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. The derivation chain is therefore not circular; the empirical claim has independent content despite test-selection caveats.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central results rest on a new private dataset and on hyperparameters tuned on the test set. The model itself is a parameter-efficient adaptation of SAM with standard building blocks; no new physical/mathematical entity is introduced. ZMIS5K is listed as an invented entity because it is the load-bearing resource and is not publicly released.

free parameters (4)
  • Wavelet basis φ = Haar
    Selected on the ZMIS5K test set (Table 3) among rbio1.1, bior1.1, and haar; test-set selection means the reported 73.6 mAP is optimistic for unseen data.
  • NFAM split count n = 4
    Chosen from a test-set ablation (Appendix C.2), where n=4 gives 73.0 mAP vs 72.2-72.7 for other values.
  • Loss weights α, β1, β2, θ1, θ2, λ = all 1.0
    Set by hand in Appendix B.3 with no sensitivity analysis; the central result depends on this unexamined weighting.
  • ZM-ViT placement = replace one frozen ViT layer every 2 layers starting at layer 8
    Chosen empirically ('to ensure fair comparison' in §3.2); no ablation shows this placement is optimal.
assumptions (6)
  • domain assumption ZMIS5K annotations are accurate enough to support claimed AP differences.
    Appendix A.1 reports expert review and pairwise IoU 0.9008, but all metrics are computed against these labels; label errors propagate into every comparison.
  • domain assumption Near-duplicate images with SSIM>0.95 between train and test are acceptable and do not materially inflate performance.
    Appendix A.2 asserts these are due to illumination/focal-plane variation, but retains them in the test split, a load-bearing assumption for the SOTA claim.
  • domain assumption Parameter-efficient fine-tuning of SAM's ViT-H with lightweight adapters closes the zooplankton domain gap.
    Section 3.2 assumes frozen SAM features plus adapters preserve general representations while adding domain-specific knowledge; only ablation evidence supports this.
  • domain assumption Wavelet high-frequency subbands compensate for FPN upsampling loss and improve boundary segmentation.
    Section 3.4 assumes the mechanism of detail recovery; ablation (Table 2) supports it empirically but no derivation or separate measurement isolates the effect.
  • domain assumption Cross-domain results on UIIS and NWPU are meaningful evidence of generalization for zooplankton microscopy.
    Appendix D compares against only USIS-SAM; on UIIS, ZMIS-SAM's AP50/AP75 are lower, so the 'strong generalization' inference is weak.
  • domain assumption The ZMIS5K train/test split is i.i.d. except for the retained SSIM>0.95 pairs.
    Section 4.1 describes an 8:2 split; the only leakage control mentioned is the SSIM check, which explicitly permits high-similarity pairs.
invented entities (1)
  • ZMIS5K dataset
    purpose: Benchmark and training resource for zooplankton microscopy instance segmentation; central to all reported results.
    The dataset is not publicly released, so its composition, labels, and difficulty cannot be independently checked; it provides no falsifiable handle outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZMIS-SAM: Segment Anything Model Enhanced with Wavelet Transform for Zooplankton Microscopy Image Instance Segmentation." pith.science (2026). https://pith.science/paper/UI446OTB

@misc{pith2026260727585,
  author       = {Pith},
  title        = {Pith review of: ZMIS-SAM: Segment Anything Model Enhanced with Wavelet Transform for Zooplankton Microscopy Image Instance Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UI446OTB}},
  note         = {Machine review of arXiv:2607.27585}
}
read the original abstract

As primary consumers in the marine food chain, zooplankton play a crucial role in maintaining marine ecological balance. However, the Segment Anything Model (SAM) exhibits limited performance in microscopic image instance segmentation due to its lack of zooplankton-specific domain knowledge. To address these challenges, we propose a novel instance segmentation model based on SAM and wavelet transform (ZMIS-SAM), effectively tackling issues such as inaccurate classification, discontinuous segmentation of slender appendages, and incomplete boundary segmentation. Our framework incorporates three core innovations: ZM-ViT enhances SAM's capability to model zooplankton morphology and image intensity distributions through two lightweight adapters, the Neighboring Feature Aggregation Module (NFAM) improves continuous segmentation of semi-transparent slender appendages by integrating general-purpose and domain-specific features, and the Wavelet-based Multi-scale Multi-directional Feature Enhancement (WM2FE) module effectively recovers high-frequency details to refine boundary segmentation completeness. Extensive experiments demonstrate that ZMIS-SAM achieves state-of-the-art instance segmentation performance on the zooplankton dataset and exhibits strong generalization capability across multiple public cross-domain datasets.

Figures

Figures reproduced from arXiv: 2607.27585 by the authors.

Figure 1
Figure 1. Comparison of ZMIS-SAM with other state-of-the-art models on the ZMIS5K dataset. Challenge 1, Challenge 2, and Challenge 3 represent inaccurate classification, discontinuous segmentation of slender appendages, and incomplete boundary segmen￾tation, respectively. intelligent identification techniques based on deep learning methods, such as instance segmentation, offer a promising alternative for achieving efficient a… view at source ↗
Figure 2
Figure 2. Overview of ZMIS-SAM architecture. ZMIS-SAM includes ZMIS-SAM En￾coder, ZMIS-SAM Prompter, and Mask Decoder. 3 Methodology In this section, we will introduce our proposed ZMIS-SAM model, a SAM-based framework that is designed specifically for instance segmentation of zooplankton microscopic images. The subsections are as follows: 3.1 Overall Framework of ZMIS-SAM, 3.2 Zooplankton Micrograph Adaptive Vision Transform… view at source ↗
Figure 3
Figure 3. Overview of Neighboring Feature Aggregation Module (NFAM) architecture. 3.3 NFAM: Neighboring Feature Aggregation Module To achieve automatic instance segmentation, existing approaches usually treat the representations extracted from certain ViT layers of the SAM image en￾coder as pseudo-prompts. However, when transferring SAM to domain-specific instance segmentation tasks, current methods either rely solely on the … view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Overview of Wavelet-based Multi-scale and Multi-directional Feature Enhance￾ment (WM2FE) architecture. design a Wavelet-based Multi-scale Feature Enhancement (WM2FE) module to compensate for the missing details during upsampling and improve the instance segmentation pe…
Figure 5
Figure 5. Figure 5: Quantitative comparison between the ZMIS-SAM model and other state-of-the￾art models on the ZMIS5K dataset. (a), (b), and (c) denote inaccurate classification, discontinuous segmentation of slender appendages, and incomplete boundary segmen￾tation, respectively. Higher…
Figure 6
Figure 6. Figure 6: Evaluation results of dataset construction quality. A.2 Dataset Characteristics and Statistics In this subsection, we will introduce the basic information, characteristics of ZMIS5K, and the challenges encountered in instance segmentation. Some statis￾tical information…
Figure 7
Figure 7. Figure 7: Examples of annotations for various instances in ZMIS5K. (a) is Fish larva. (b) includes Sagitta crassa, Penilia avirostris, and Evadne tergestina. (c) includes Eirene menoni and Pleurobrachia globosa. (d) includes Turritopsis nutricula and Oikopleura longicauda. magni…
Figure 8
Figure 8. Figure 8: Essential characteristics of the ZMIS5K dataset. (a) Distribution of the number of instances per image in the ZMIS5K dataset. (b) The number of instances per category in the ZMIS5K dataset. Number and Size of Instances. There are many individual images con￾taining mult…
Figure 9
Figure 9. Figure 9: Average channel intensity in the case of different datasets with proportion. (a) Average channel intensity of ZMIS5K under dark-field conditions. (b) Average channel intensity of ZMIS5K under flight-field conditions. (c) Average channel intensity of un￾derwater images.…
Figure 10
Figure 10. Figure 10: Wavelet Transform Process. (a) The Multi-Scale Wavelet Transform Process. (b) The 2-Dimensional Discrete Wavelet Transform Process for Zooplankton. cA, cH, cV, and cD stand for Approximation Coefficients, Horizontal Detail Coefficients, Ver￾tical Detail Coefficients, …
Figure 11
Figure 11. Figure 11: Visualize the feature map of ZM-ViT. The ZM-ViT enables SAM to learn knowledge of zooplankton. adaptively adjust its response to channel-wise intensity distributions, more ac￾curately capturing illumination variations under microscopic imaging conditions. Working syne…
Figure 12
Figure 12. Figure 12: Visualize the feature map of NFAM. The NFAM enhances the model’s ability to continuously segment slender appendages [PITH_FULL_IMAGE:figures/full_fig_p029_12.png]
Figure 13
Figure 13. Figure 13: Visualize the feature map of WM2FE. The WM2FE focuses more on species boundaries, thus enhancing the model’s segmentation ability for boundaries. C.7 Clear comparative visualization results In this subsection, we present clearer visualizations compared to those shown …
Figure 14
Figure 14. Figure 14: Visualization results of inaccurate classification cases. D Generalization experiments This section evaluates the cross-domain instance segmentation performance of the proposed model. Due to the absence of publicly available instance segmen￾tation datasets for zooplan…
Figure 15
Figure 15. Figure 15: Visualization results of discontinuous segmentation of slender appendages cases. Ground Truth Mask2Former Mask R-CNN RSPrompter SAM+Bbox SAM+Mask USIS-SAM ZMIS-SAM (Our) [PITH_FULL_IMAGE:figures/full_fig_p031_15.png]
Figure 16
Figure 16. Figure 16: Visualization results of incomplete boundary segmentation cases [PITH_FULL_IMAGE:figures/full_fig_p031_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 6 canonical work pages

  1. [1]

    arXiv (2016),http://arxiv.org/abs/1609.025834

    Arnab, A., Torr, P.H.S.: Bottom-up instance segmentation using deep higher-order crfs. arXiv (2016),http://arxiv.org/abs/1609.025834

  2. [2]

    IEEE Transactions on Circuits and Sys- tems for Video Technology16(2), 220–230 (2006).https://doi.org/10.1109/ TCSVT.2005.8578165

    Balster, E., Zheng, Y., Ewing, R.: Combined spatial and temporal domain wavelet shrinkage algorithm for video denoising. IEEE Transactions on Circuits and Sys- tems for Video Technology16(2), 220–230 (2006).https://doi.org/10.1109/ TCSVT.2005.8578165

  3. [3]

    In: Global Oceans 2020: Singapore – U.S

    Bergum,S.,Saad,A.,Stahl,A.:Automaticin-situinstanceandsemanticsegmenta- tion of planktonic organisms using mask r-cnn. In: Global Oceans 2020: Singapore – U.S. Gulf Coast. pp. 1–8 (2020).https://doi.org/10.1109/IEEECONF38699. 2020.938937722

  4. [4]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Bi, Q., You, S., Gevers, T.: Learning generalized segmentation for foggy-scenes by bi-directional wavelet guidance. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 801–809 (2024).https://doi.org/10.1609/ aaai.v38i2.278385

  5. [5]

    In: 2019 IEEE International Conference on Computer Vision (ICCV) (2019)

    Bolya, D., Zhou, C., Xiao, F., Lee, Y.J.: Yolact: Real-time instance segmenta- tion. In: 2019 IEEE International Conference on Computer Vision (ICCV) (2019). https://doi.org/10.1109/ICCV.2019.009254

  6. [6]

    In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6154–6162 (2018).https://doi.org/10.1109/CVPR.2018.0064412

  7. [7]

    In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Chen, H., Sun, K., Tian, Z., Shen, C., Huang, Y., Yan, Y.: BlendMask: Top-down meets bottom-up for instance segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8570–8578 (2020).https: //doi.org/10.1109/CVPR42600.2020.008604

  8. [8]

    arXiv (2019),https://arxiv

    Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv (2019),https://arxiv. org/abs/1906.0715511

Show all 58 references
  1. [9]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–17 (2024)

    Chen, K., Liu, C., Chen, H., Zhang, H., Li, W., Zou, Z., Shi, Z.: Rsprompter: Learn- ing to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing62, 1–17 (2024). https://doi.org/10.1109/TGRS.2024.33...

  2. [10]

    In: 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)

    Chen, T., Zhu, L., Ding, C., Cao, R., Wang, Y., Zhang, S., Li, Z., Sun, L., Zang, Y., Mao, P.: Sam-adapter: Adapting segment anything in underperformed scenes. In: 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). pp. 3359–3367 (2023).https://doi.org...

  3. [11]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Chen, W., Vong, Y.J., Kuo, S.Y., Ma, S., Wang, J.: Robustsam: Segment anything robustly on degraded images. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4081–4091 (2024).https://doi.org/10. 1109/CVPR52733.2024.003912, 5

  4. [13]

    In: Proceedings of the 35th International Conference on Neural Information Processing Systems (NeurIPS) (2021) 4

    Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: Proceedings of the 35th International Conference on Neural Information Processing Systems (NeurIPS) (2021) 4

  5. [14]

    IEEE Trans- actions on Geoscience and Remote Sensing54(12), 7405–7415 (2016).https: //doi.org/10.1109/TGRS.2016.260162222, 32

    Cheng, G., Zhou, P., Han, J.: Learning rotation-invariant convolutional neural networks for object detection in vhr optical remote sensing images. IEEE Trans- actions on Geoscience and Remote Sensing54(12), 7405–7415 (2016).https: //doi.org/10.1109/TGRS.2016.260162222, 32

  6. [15]

    Ding, S., Wang, Q., Guo, L., Li, X., Ding, L., Wu, X.: Wavelet and adaptive coor- dinate attention guided fine-grained residual network for image denoising. vol. 34, pp. 6156–6166 (2024).https://doi.org/10.1109/TCSVT.2023.33488045

  7. [16]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 38(2), 295–307 (2016).https://doi.org/10.1109/TPAMI.2015.24392813

    Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolu- tional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38(2), 295–307 (2016).https://doi.org/10.1109/TPAMI.2015.24392813

  8. [17]

    In: 9th International Conference on Learning Representations (ICLR) (2021) 3, 23

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16 x 16 words: Transformers for image recognition at scale. In: 9th International Conferenc...

  9. [18]

    IEEE Transactions on Circuits and Systems for Video Technology35(4), 2963–2974 (2025).https://doi.org/10.1109/TCSVT

    Fang, H., Zhang, T., Zhou, X., Zhang, X.: Learning better video query with sam for video instance segmentation. IEEE Transactions on Circuits and Systems for Video Technology35(4), 2963–2974 (2025).https://doi.org/10.1109/TCSVT. 2024.33610765

  10. [19]

    In: Computer Vision – ECCV 2024

    Finder, S.E., Amoyal, R., Treister, E., Freifeld, O.: Wavelet convolutions for large receptive fields. In: Computer Vision – ECCV 2024. pp. 363–380. Springer Nature Switzerland, Cham (2025).https://doi.org/10.1007/978-3-031-72949-2_215

  11. [20]

    In: 2017 IEEE Inter- national Conference on Computer Vision (ICCV)

    He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: 2017 IEEE Inter- national Conference on Computer Vision (ICCV). pp. 2980–2988 (2017).https: //doi.org/10.1109/ICCV.2017.3222, 4, 12

  12. [21]

    IEEE Transactions on Image Processing33, 2895–2907 (2024)

    Hu, J., Lu, Y., Zhang, S., Cao, L.: Istr: Mask-embedding-based instance segmenta- tion transformer. IEEE Transactions on Image Processing33, 2895–2907 (2024). https://doi.org/10.1109/TIP.2024.33859804

  13. [22]

    In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Huang, Z., Huang, L., Gong, Y., Huang, C., Wang, X.: Mask scoring r-cnn. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6402–6411 (2019).https://doi.org/10.1109/CVPR.2019.0065712

  14. [23]

    IEEE Trans- actions on Image Processing35, 4012–4026 (2026).https://doi.org/10.1109/ TIP.2026.368212612

    Jia, Y., Lin, Q., Li, H., Li, Y., Kwong, S., Cong, R.: Vit-uwa: Vision transformer underwater-adapter for dense predictions beneath the water surface. IEEE Trans- actions on Image Processing35, 4012–4026 (2026).https://doi.org/10.1109/ TIP.2026.368212612

  15. [25]

    In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR)

    Kirillov, A., Wu, Y., He, K., Girshick, R.: Pointrend: Image segmentation as ren- dering. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 9796–9805 (2020).https://doi.org/10.1109/CVPR42600. 2020.009824

  16. [26]

    IEEE Transactions on Multimedia27, 2795–2808 (2025)

    Lan, X., Lyu, J., Jiang, H., Dong, K., Niu, Z., Zhang, Y., Xue, J.: Foodsam: Any food segmentation. IEEE Transactions on Multimedia27, 2795–2808 (2025). https://doi.org/10.1109/TMM.2023.33300472, 5

  17. [28]

    ICES Journal of Marine Science77(4), 1427–1439 (09 2019)

    Li, Q., Sun, X., Dong, J., Song, S., Zhang, T., Liu, D., Zhang, H., Han, S.: Develop- ing a microscopic image dataset in support of intelligent phytoplankton detection using deep learning. ICES Journal of Marine Science77(4), 1427–1439 (09 2019). https://doi.org/10.1093/icesjm...

  18. [29]

    In: Pattern Recognition and Computer Vision

    Li, Q., Shen, L.: Wavesnet: Wavelet integrated deep networks for image segmenta- tion. In: Pattern Recognition and Computer Vision. pp. 325–337. Springer Nature Switzerland, Cham (2022).https://doi.org/10.1007/978-3-031-18916-6_275

  19. [30]

    In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV)

    Lian, S., Li, H., Cong, R., Li, S., Zhang, W., Kwong, S.: Watermask: Instance seg- mentation for underwater imagery. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1305–1315 (2023).https://doi.org/10.1109/ ICCV51070.2023.001264, 12, 32

  20. [31]

    In: Proceedings of the 41st International Conference on Machine Learning (ICML)

    Lian, S., Zhang, Z., Li, H., Li, W., Yang, L.T., Kwong, S., Cong, R.: Diving into underwater:Segmentanythingmodelguidedunderwatersalientinstancesegmenta- tion and a large-scale dataset. In: Proceedings of the 41st International Conference on Machine Learning (ICML). pp. 29545–...

  21. [32]

    In: 2017 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Lin, T.Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Fea- ture pyramid networks for object detection. In: 2017 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 936–944 (2017).https: //doi.org/10.1109/CVPR.2017.1063, 6

  22. [33]

    arXiv (2023),https://arxiv.org/abs/2304.097285

    Liu, S., Ye, J., Wang, X.: Any-to-any style transfer: Making picasso and da vinci collaborate. arXiv (2023),https://arxiv.org/abs/2304.097285

  23. [34]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR)

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 11966–11976 (2022).https://doi.org/10.1109/CVPR52688. 2022.0116712

  24. [35]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Lu, Y., Liu, J., Zhang, Y., Liu, Y., Tian, X.: Prompt distribution learning. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5196–5205 (2022).https://doi.org/10.1109/CVPR52688.2022.005143

  25. [36]

    IEEE Journal of Selected Topics in Applied Earth Obser- vations and Remote Sensing17, 17117–17133 (2024).https://doi.org/10.1109/ JSTARS.2024.34244981

    Masoudi, M., Giering, S.L., Eftekhari, N., Massot-Campos, M., Irisson, J.O., Thornton, B.: Optimizing plankton image classification with metadata-enhanced representation learning. IEEE Journal of Selected Topics in Applied Earth Obser- vations and Remote Sensing17, 17117–17133...

  26. [37]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3523–3542 (2022).https://doi.org/10

    Minaee, S., Boykov, Y., Porikli, F., Plaza, A., Kehtarnavaz, N., Terzopoulos, D.: Image segmentation using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3523–3542 (2022).https://doi.org/10. 1109/TPAMI.2021.30599682

  27. [38]

    In: Proceedings of the IEEE/CVF International Conference on 18 Dekun.Y et al

    Pu, Y., Feng, Z., Wang, Z., Yang, Z., Li, J.: Anomaly detection for in situ marine plankton images. In: Proceedings of the IEEE/CVF International Conference on 18 Dekun.Y et al. Computer Vision (ICCV) Workshops. pp. 3661–3671 (2021).https://doi.org/ 10.1109/ICCVW54120.2021.004091

  28. [39]

    In: Proceedings of the 38th International Conference on Machine Learning (ICML)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transfer- able visual models from natural language supervision. In: Proceedings of the 38th International Conference on M...

  29. [40]

    Nature Communications14(2023).https://doi.org/10

    Ratnarajah, Lavenia, A.A., Rana, A., Angus, B., Sonia, B., Nicholas J, B., Kim S, C., Gabrielle, C., Astrid, E., Jason D, G., Maria, I., Nurul Huda Ahmad, J., David, L., Fabien, M., Erik, O., Clare, P., Sophie, R., Anthony J, S., Katrin, S., Lars, S., Kerrie M, Y., Guang, Y., ...

  30. [41]

    arXiv (2024),https://arxiv.org/abs/2408.007142

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: Sam 2: Segment anything in images and videos. arXiv (2024),https:/...

  31. [42]

    In: Computer Analysis of Images and Patterns (CAIP)

    Rossi, L., Karimi, A., Prati, A.: Recursively refined r-cnn: Instance segmentation with self-roi rebalancing. In: Computer Analysis of Images and Patterns (CAIP). pp. 476–486 (2021).https://doi.org/10.1007/978-3-030-89128-2_464, 12

  32. [43]

    In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Shan, Z., Liu, Y., Zhou, L., Yan, C., Wang, H., Xie, X.: Ros-sam: High-quality interactive segmentation for remote sensing moving object. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3625–3635 (2025).https://doi.org/10.1109/CVPR52734.2025.003435

  33. [44]

    In: Computer Vision – ECCV 2020

    Tian, Z., Shen, C., Chen, H.: Conditional convolutions for instance segmentation. In: Computer Vision – ECCV 2020. pp. 282–298 (2020).https://doi.org/10. 1007/978-3-030-58452-8_1712

  34. [45]

    IEEE Trans- actions on Image Processing35, 3283–3298 (2026).https://doi.org/10.1109/ TIP.2026.367550212

    Wang, C., Li, H., Li, C., Liu, H., Tang, X., Kwong, S.: Expose camouflage in the water: Underwater camouflaged instance segmentation and dataset. IEEE Trans- actions on Image Processing35, 3283–3298 (2026).https://doi.org/10.1109/ TIP.2026.367550212

  35. [46]

    Wang, D., Zhang, J., Du, B., Xu, M., Liu, L., Tao, D., Zhang, L.: Samrs: Scaling-up remotesensingsegmentationdatasetwithsegmentanythingmodel.In:Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS). vol. 36, pp. 8815–8827 (2023) 5

  36. [47]

    arXiv (2024), https://arxiv.org/abs/2410.198787

    Wang, L., Chen, S., Jiang, L., Pan, S., Cai, R., Yang, S., Yang, F.: Parameter- efficient fine-tuning in large models: A survey of methodologies. arXiv (2024), https://arxiv.org/abs/2410.198787

  37. [48]

    In: Computer Vision – ECCV 2020

    Wang, X., Kong, T., Shen, C., Jiang, Y., Li, L.: SOLO: Segmenting objects by locations. In: Computer Vision – ECCV 2020. pp. 649–665 (2020).https://doi. org/10.1007/978-3-030-58523-5_384

  38. [49]

    IEEE Trans- actions on Intelligent Transportation Systems26(6), 8794–8809 (2025).https: //doi.org/10.1109/TITS.2025.35439255

    Wang, Z., Liao, Z., Wang, P., Chen, P., Luo, W.: Wavecrnet: Wavelet transform- guided learning for semantic segmentation in adverse railway scenes. IEEE Trans- actions on Intelligent Transportation Systems26(6), 8794–8809 (2025).https: //doi.org/10.1109/TITS.2025.35439255

  39. [50]

    In: Computer Vision – ECCV 2018

    Woo, S., Park, J., Lee, J.Y., Kweon, I.S.: Cbam: Convolutional block attention module. In: Computer Vision – ECCV 2018. pp. 3–19. Springer International Pub- lishing, Cham (2018).https://doi.org/10.1007/978-3-030-01234-2_110

  40. [51]

    ZMIS-SAM 19 Pattern Recognition143, 109819 (2023).https://doi.org/10.1016/j.patcog

    Xu, G., Liao, W., Zhang, X., Li, C., He, X., Wu, X.: Haar wavelet downsam- pling: A simple but effective downsampling module for semantic segmentation. ZMIS-SAM 19 Pattern Recognition143, 109819 (2023).https://doi.org/10.1016/j.patcog. 2023.1098195

  41. [52]

    IEEE Transactions on Image Processing33, 6204–6215 (2024).https://doi.org/10

    Xu, Y., Tang, J., Men, A., Chen, Q.: Eviprompt: A training-free evidential prompt generation method for adapting segment anything model in medical images. IEEE Transactions on Image Processing33, 6204–6215 (2024).https://doi.org/10. 1109/TIP.2024.34821755

  42. [53]

    IEEE Transactions on Circuits and Systems for Video Technology33(4), 1899–1910 (2023).https: //doi.org/10.1109/TCSVT.2022.32187355

    Yang, Y., Jiao, L., Liu, X., Liu, F., Yang, S., Li, L., Chen, P., Li, X., Huang, Z.: Dual wavelet attention networks for image classification. IEEE Transactions on Circuits and Systems for Video Technology33(4), 1899–1910 (2023).https: //doi.org/10.1109/TCSVT.2022.32187355

  43. [54]

    IEEE Trans- actions on Pattern Analysis and Machine Intelligence47(3), 1431–1447 (2025)

    Ye, M., Zhang, J., Liu, J., Liu, C., Yin, B., Liu, C., Du, B., Tao, D.: Hi-sam: Mar- rying segment anything model for hierarchical text segmentation. IEEE Trans- actions on Pattern Analysis and Machine Intelligence47(3), 1431–1447 (2025). https://doi.org/10.1109/TPAMI.2024.34958315

  44. [55]

    Engineering Applications of Ar- tificial Intelligence163, 112719 (2026).https://doi.org/10.1016/j.engappai

    Ye, X., Wang, H., Yao, J.: A lightweight deep learning network for the precise detection and classification of variable plankton. Engineering Applications of Ar- tificial Intelligence163, 112719 (2026).https://doi.org/10.1016/j.engappai. 2025.1127191

  45. [56]

    arXiv (2023),https://arxiv.org/abs/ 2304.067902, 5

    Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv (2023),https://arxiv.org/abs/ 2304.067902, 5

  46. [57]

    IEEE Transactions on Circuits and Systems for Video Technology35(3), 2932–2944 (2025).https://doi.org/10.1109/TCSVT.2024.34983491

    Yu, Y., Lv, Q., Li, Y., Wei, Z., Dong, J.: Phytracker: An online tracker for phyto- plankton. IEEE Transactions on Circuits and Systems for Video Technology35(3), 2932–2944 (2025).https://doi.org/10.1109/TCSVT.2024.34983491

  47. [58]

    In: OCEANS 2025 Brest

    Yuan,D.,Qi,Y.,Zhang,J.,Li,Z.:Planktontnet:Rethinkingplanktonclassification from a global view with swin-transformer. In: OCEANS 2025 Brest. pp. 1–5 (2025). https://doi.org/10.1109/OCEANS58557.2025.111047971

  48. [59]

    arXiv (2023),https: //arxiv.org/abs/2305.081963

    Zhang, C., Liu, L., Cui, Y., Huang, G., Lin, W., Yang, Y., Hu, Y.: A comprehensive survey on segment anything model for vision and beyond. arXiv (2023),https: //arxiv.org/abs/2305.081963

  49. [60]

    IEEE Trans- actions on Circuits and Systems for Video Technology34(4), 2469–2483 (2024)

    Zhang, W., Zhou, L., Zhuang, P., Li, G., Pan, X., Zhao, W., Li, C.: Underwater image enhancement via weighted wavelet visual perception fusion. IEEE Trans- actions on Circuits and Systems for Video Technology34(4), 2469–2483 (2024). https://doi.org/10.1109/TCSVT.2023.32993145

  50. [61]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Zou, W., Gao, H., Yang, W., Liu, T.: Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement. In: Proceedings of the 32nd ACM International Conference on Multimedia. p. 1534–1543 (2024).https://doi. org/10.1145/3664647.36815805 20 Dekun.Y et...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.