Pith. sign in

REVIEW 2 major objections 5 minor 48 references

Whether a frozen vision backbone applies global attention at high resolution decides if raising input size improves fine-grained segmentation under low-rank adaptation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 07:34 UTC pith:LPPJXKLT

load-bearing objection Clean controlled study: under fixed LoRA, global attention on a high-res token grid predicts whether resolution keeps buying accuracy; the claim is real and correctly scoped to attention-level adapters. the 2 major comments →

arxiv 2607.02708 v1 pith:LPPJXKLT submitted 2026-07-02 cs.CV

When Does Resolution Help a Frozen Backbone? Global Attention at Resolution Predicts Scalable Adaptation for Camouflaged and Marine Animal Segmentation

Tyler Rust , Chandra Kambhamettu This is my paper
classification cs.CV
keywords frozen backbone adaptationLoRAisotropic Vision Transformerhierarchical transformerglobal attentionresolution scalingcamouflaged object detectionmarine animal segmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

When adapting frozen vision foundation models for fine-grained binary segmentation (camouflaged objects, marine animals, salient objects), the practical choice of backbone is not settled by size or pretraining alone. This paper shows that one structural property predicts whether a low-rank adapter turns higher input resolution into higher accuracy: whether the backbone applies global attention over a high-resolution token grid. Isotropic Vision Transformers keep the full token grid and attend globally at every depth, so they improve monotonically with resolution. Hierarchical backbones confine early attention to local windows and pool the grid before global stages, so they plateau near 512 pixels. A controlled six-backbone study under one fixed lightweight pipeline establishes the pattern; editing isotropic backbones isolates the cause, because pooling the grid leaves scaling intact while removing global attention collapses it. The effect is specific to low-rank (attention-level) adaptation. The same pipeline on a strong isotropic backbone then sets new state-of-the-art numbers on marine-animal benchmarks and leads the camouflaged and salient suites in a single RGB pass.

Core claim

Under one fixed low-rank adaptation pipeline, whether a frozen backbone applies global attention to a high-resolution token set predicts whether raising input resolution improves fine-grained segmentation accuracy. Isotropic ViTs keep improving with resolution; hierarchical backbones saturate by roughly 512 px. Pooling an isotropic grid preserves the scaling benefit; removing global attention removes most of it. The dichotomy is specific to attention-level (LoRA) adaptation rather than every adapter.

What carries the argument

Global attention at resolution: the property that an isotropic ViT attends globally over the full high-resolution token grid at every depth, whereas hierarchical backbones reach global stages only after pooling (or never apply global attention). This property, not capacity, pretraining, or token-grid preservation alone, carries the resolution-scaling dichotomy under LoRA.

Load-bearing premise

The observed scaling rule is treated as a general backbone-selection guide even though the paper itself shows it holds under low-rank attention adaptation and can reverse under a different feature-space adapter.

What would settle it

Repeat the identical resolution sweep under the same LoRA pipeline on a new isotropic backbone whose global attention has been fully replaced by local windows (or on a new hierarchical backbone that applies true global attention before pooling); if the isotropic model still scales or the hierarchical one no longer saturates, the causal claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper argues that, for low-rank adaptation of frozen vision foundation models on fine-grained binary segmentation (COD, MAS, SOD), a single structural property predicts whether resolution converts into accuracy: whether the backbone applies global attention to a high-resolution token set. Under a fixed lightweight pipeline (SALT: rank-8 LoRA + convolutional neck + attention-gated U-Net + CNN side-stem), three isotropic ViTs (SAM3 PE, DINOv2, DINOv3) improve monotonically with resolution while three hierarchical backbones (Hiera-L, Hiera-B+, Swin-L) saturate near 512 px (Table 5, Fig. 1). Confound controls (capacity, objective, pretraining) and direct edits on DINOv2/DINOv3 show that pooling the grid preserves scaling whereas removing global attention collapses it (Table 6); the dichotomy is specific to attention-level (LoRA) adaptation and disappears under a feature-space bottleneck (Table 7). With SALT-DINOv3 the method sets new SOTA on both MAS benchmarks (MAS3K mIoU 0.878) and obtains the best or tied S-measure on the four COD and five SOD sets in a single RGB pass at 672 px.

Significance. If the result holds, the work supplies a concrete, falsifiable backbone-selection rule for practitioners performing LoRA-style PEFT on resolution-sensitive dense tasks, grounded in a controlled six-backbone sweep rather than a new module. Credit is due for the isolation of geometry (pipeline held fixed), the three confound controls, the causal attention-editing experiments on two isotropic models, the FLOPs/throughput reporting that turns the dichotomy into a cost-accuracy trade-off, the isolation study showing that pretrained features alone underperform prior work, and the strong external-benchmark results (new SOTA on MAS3K/RMAS, best S_m on all four COD sets). The authors carefully scope the mechanism to low-rank adaptation and acknowledge residual correlational aspects of the isotropic-versus-hierarchical comparison. This is a useful empirical contribution to understanding when frozen VFMs remain adaptable under lightweight fine-grained adaptation.

major comments (2)
  1. [Abstract, Sec. 1 contributions, Sec. 5.2, Table 7, Limitations, Sec. 6] The practical backbone-selection rule ('choose an isotropic ViT that applies global attention at high resolution') is stated in the abstract, contributions list, and conclusion, yet Table 7 and Sec. 5.2 show that a matched-rank feature-space bottleneck adapter lets the hierarchical Swin-L scale (+0.034) while the LoRA dichotomy vanishes. Although the Limitations section scopes the mechanism to LoRA, the framing of the rule itself remains broader. The manuscript should qualify the guidance as LoRA-/attention-level-specific in the abstract, introduction, and conclusion, and briefly discuss when a practitioner might prefer hierarchical geometry plus a feature-space adapter.
  2. [Table 6, Sec. 5.2, Limitations] Several load-bearing attention edits in Table 6 (fully windowed DINOv2, pool-after-block-2, reconstructed Hiera pattern) remain single-seed, while the unedited baselines and the DINOv3 no-global rows are two-seed means. The causal claim that 'removing global attention collapses scaling' rests on these slope reversals (e.g., +0.001 vs +0.030). Multi-seed means (or at least seed-band reporting) for every edited configuration, plus a short clarification of why the reconstructed hierarchical pattern saturates on DINOv2 but not DINOv3 (already noted in Limitations), would make the causal account more robust without changing the experimental design.
minor comments (5)
  1. [Table 5, Sec. 4.3] Resolution tiers are matched by design (336/504/672 vs 352/512/704) but are not pixel-identical. A one-sentence note of the resulting token-grid sizes (e.g., 24/36/48 vs the corresponding hierarchical stages) next to Table 5 would make the FLOPs/throughput comparison easier to interpret.
  2. [Table 1, Sec. 4.4] The component ablation (Table 1) is performed only at 336 px on the SAM3 PE. A single additional row confirming that LoRA remains the dominant term at the 672 px operating point used for the SOTA claims would strengthen the adaptation-isolation narrative.
  3. [Sec. 4.3, Sec. 6] Code, training configs, and the edited-backbone implementations are not mentioned as released. For a controlled empirical study whose primary value is the protocol and the attention edits, releasing the SALT pipeline and the Table-6 configurations would materially improve reproducibility.
  4. [Sec. 5.5, Fig. 3] The qualitative scan against BiRefNet maps (53 % / 48 % win rates on NC4K/CAMO) is helpful; a short quantitative summary of the IoU-delta distribution or one representative failure case would balance the high-margin successes shown in Fig. 3.
  5. [Sec. 3.4, Eq. (2)] Minor notation consistency: the structure loss is written L = L_w^BCE + L_w^IoU in Eq. (2) while the text sometimes refers to 'boundary-weighted' without restating the weighting scheme; a brief pointer to the source papers would help.

Circularity Check

0 steps flagged

No significant circularity: empirical backbone-resolution study on public benchmarks with fixed public checkpoints; no derivation that reduces to its inputs by construction.

full rationale

The paper's load-bearing claim is an empirical dichotomy under a fixed adapter (SALT/LoRA): isotropic ViTs that apply global attention over a high-resolution token grid keep converting input resolution into accuracy, while hierarchical backbones that confine global attention to post-pooling stages saturate near 512 px. That claim is established by a controlled six-backbone resolution sweep (Table 5, Fig. 1), confound controls (capacity, objective, pretraining), and causal edits that pool the grid or remove global attention on DINOv2/DINOv3 (Table 6), then evaluated on external public MAS/COD/SOD benchmarks with public foundation-model checkpoints. No parameter is fitted to a subset and re-reported as a prediction; global attention is not defined in terms of the scaling outcome; and the authors do not import a uniqueness theorem or ansatz from their own prior work. The adapter-specificity finding (Table 7: a feature-space bottleneck lets Swin-L scale) is an explicit scope limitation, not a circular reduction. Isolation of pretrained features alone (0.784 mIoU, below SAM2-UNet) further shows the result is not tautological with the backbone. The derivation chain is therefore self-contained measurement, not circular.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 1 invented entities

Empirical CV paper; load-bearing content is experimental design rather than free parameters or new physical entities. Free parameters are the usual PEFT and training choices held fixed across backbones. Axioms are standard architectural facts about ViT vs. hierarchical transformers plus the assumption that the fixed SALT pipeline isolates backbone geometry.

free parameters (3)
  • LoRA rank r and scale α = r=8, α=16
    Fixed at r=8, α=16 for all backbones; the central claim is scoped to this low-rank attention adaptation regime.
  • Resolution tiers (336/504/672 vs 352/512/704) = three matched tiers
    Chosen to match patch sizes and produce comparable token grids; the scaling slopes are measured across these discrete points.
  • SALT decoder capacity (2.7-5.6 M trainable params) = 2.7-5.6 M
    Neck, attention-gated U-Net, side-stem and heads are fixed; only backbone varies, so decoder size is a controlled constant of the study.
axioms (3)
  • domain assumption Isotropic ViTs maintain a uniform token grid and apply global self-attention in every block; hierarchical backbones (Hiera, Swin) use early local windows and pool before global stages.
    Standard architectural fact used throughout Sec. 1, 3.2 and 5.2 to define the two geometries.
  • domain assumption Public pretrained checkpoints (SAM3 PE, DINOv2/v3, SAM2 Hiera, Swin-L) are representative of their geometry classes and can be frozen without further pretraining.
    All experiments start from these released weights (Sec. 4.3).
  • ad hoc to paper The fixed SALT pipeline (LoRA + convolutional neck + attention-gated U-Net + CNN side-stem) does not itself favor one geometry; any residual neck asymmetry favors hierarchical stages.
    Stated in Sec. 3.2 and used to attribute the dichotomy to the backbone rather than the adapter (Sec. 5.2).
invented entities (1)
  • SALT pipeline no independent evidence
    purpose: Single fixed lightweight adapter (LoRA + side-stem attention-gated U-Net) used to hold everything constant while only the frozen backbone varies.
    Combination of known components (LoRA, U-Net, ViTDet pyramid, side-stem); not a new theoretical object, but the experimental constant of the study.

pith-pipeline@v1.1.0-grok45 · 19321 in / 2910 out tokens · 27899 ms · 2026-07-12T07:34:30.972235+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of When Does Resolution Help a Frozen Backbone? Global Attention at Resolution Predicts Scalable Adaptation for Camouflaged and Marine Animal Segmentation." pith.science (2026). https://pith.science/paper/LPPJXKLT

@misc{pith2026260702708,
  author       = {Pith},
  title        = {Pith review of: When Does Resolution Help a Frozen Backbone? Global Attention at Resolution Predicts Scalable Adaptation for Camouflaged and Marine Animal Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LPPJXKLT}},
  note         = {Machine review of arXiv:2607.02708}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Adapting frozen vision foundation models to fine-grained segmentation now largely depends on backbone selection. Whether the backbone applies global attention to a high-resolution token set predicts whether a low-rank adapter turns resolution into accuracy. Isotropic ViTs attend globally over the full grid and keep improving with resolution; hierarchical backbones confine early attention to local windows and pool the grid before their global stages, plateauing at lower resolutions. A controlled six-backbone study establishes the pattern, and editing the backbone points to the cause: pooling keeps the benefit, removing global attention does not. The effect is specific to low-rank adaptation. Under one fixed pipeline, SALT (Side-stem, Attention-gated U-Net, Low-rank Tuning), one RGB-only pass on a strong isotropic backbone wins the best S-measure on the four data-matched camouflaged sets, and leads every marine and salient set. It reaches a new state of the art on both marine-animal benchmarks (MAS3K mIoU 0.878).

Figures

Figures reproduced from arXiv: 2607.02708 by Chandra Kambhamettu, Tyler Rust.

Figure 1
Figure 1. Figure 1: Global attention at resolution, not backbone iden￾tity, predicts scaling. On MAS3K, the three isotropic backbones (blue: SAM3 PE, DINOv3, DINOv2) keep turning resolution into accuracy and stay above the strongest prior (dotted, SAM2-UNet), while the three hierarchical ones (red: Hiera-L, Hiera-B+, Swin￾L) saturate near 512px, regardless of size, pretraining, or design. tical question is which one to adapt.… view at source ↗
Figure 2
Figure 2. Figure 2: The shared SALT decoder. A frozen backbone is adapted with rank-8 LoRA; multi-scale neck features are compressed by a convolutional neck, decoded by an attention-gated U-Net, and fused with a high-frequency CNN side-stem. This pipeline is applied to every backbone; only the neck differs (isotropic ViTs get a ViTDet-style feature pyramid, the hierarchical Hiera uses native stages). pooling or by merging red… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative COD results for SALT-DINOv3 at 672 (RGB-only, single pass) against SAM2-UNet and BiRefNet (1024 px). Columns: input, ground truth, SAM2-UNet, BiRefNet, ours. Top: a texture-matched CAMO subject both baselines frag￾ment; bottom: an NC4K dragonfly whose thin wings and tail both baselines drop. Ours recovers both (Sec. 5.5). broad advantage, not isolated wins. 6. Conclusion We asked which property… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 2 linked inside Pith

  1. [1]

    Token merging: Your ViT but faster

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. InICLR, 2023. 3

    2023

  2. [2]

    SAM 3: Segment anything with concepts

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, et al. SAM 3: Segment anything with concepts. InICLR, 2026. 1, 2, 3, 4

    2026

  3. [3]

    SAM-Adapter: Adapting segment anything in under- performed scenes

    Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. SAM-Adapter: Adapting segment anything in under- performed scenes. InICCV Workshops, 2023. 6

    2023

  4. [4]

    SAM2- Adapter: Evaluating and adapting segment anything 2 in downstream tasks.arXiv preprint arXiv:2408.04579, 2024

    Tianrun Chen, Lanyun Zhu, Chaotao Ding, et al. SAM2- Adapter: Evaluating and adapting segment anything 2 in downstream tasks.arXiv preprint arXiv:2408.04579, 2024. 1, 3

    Pith/arXiv arXiv 2024

  5. [5]

    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. InICLR, 2023. 2

    2023

  6. [6]

    Structure- measure: A new way to evaluate foreground maps

    Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, et al. Structure- measure: A new way to evaluate foreground maps. InICCV,

  7. [7]

    Enhanced- alignment measure for binary foreground map evaluation

    Deng-Ping Fan, Cheng Gong, Yang Cao, et al. Enhanced- alignment measure for binary foreground map evaluation. In IJCAI, 2018. 4

    2018

  8. [8]

    Camouflaged object detection

    Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, et al. Camouflaged object detection. InCVPR, 2020. 1, 2, 4

    2020

  9. [9]

    PraNet: Par- allel reverse attention network for polyp segmentation

    Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, et al. PraNet: Par- allel reverse attention network for polyp segmentation. In MICCAI, 2020. 4

    2020

  10. [10]

    MASNet: A robust deep ma- rine animal segmentation network.IEEE Journal of Oceanic Engineering, 2023

    Zhenqi Fu, Ruizhe Chen, Yue Huang, En Cheng, Xinghao Ding, and Kai-Kuang Ma. MASNet: A robust deep ma- rine animal segmentation network.IEEE Journal of Oceanic Engineering, 2023. 4, 6

    2023

  11. [11]

    H2Former: An efficient hierarchical hybrid transformer for medical image segmentation.IEEE Transac- tions on Medical Imaging, 2023

    Along He, Kai Wang, Tao Li, Chengkun Du, Shuang Xia, and Huazhu Fu. H2Former: An efficient hierarchical hybrid transformer for medical image segmentation.IEEE Transac- tions on Medical Imaging, 2023. 6

    2023

  12. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR, 2016. 2

    2016

  13. [13]

    LoRA: Low- rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, et al. LoRA: Low- rank adaptation of large language models. InICLR, 2022. 1, 3

    2022

  14. [14]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, et al. Segment anything. InICCV, 2023. 1, 2, 6

    2023

  15. [15]

    Anabranch network for camouflaged object segmentation

    Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, et al. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, 2019. 2, 4

    2019

  16. [16]

    Visual saliency based on multi- scale deep features

    Guanbin Li and Yizhou Yu. Visual saliency based on multi- scale deep features. InCVPR, 2015. 4

    2015

  17. [17]

    MAS3K: An open dataset for marine animal segmentation

    Lin Li, Eric Rigall, Junyu Dong, and Geng Chen. MAS3K: An open dataset for marine animal segmentation. InInter- national Symposium on Benchmarking, Measuring and Op- timizing (Bench), 2020. 4

    2020

  18. [18]

    The secrets of salient object segmentation

    Yin Li, Xiaodi Hou, Christof Koch, et al. The secrets of salient object segmentation. InCVPR, 2014. 4

    2014

  19. [19]

    Exploring plain vision transformer backbones for object de- tection

    Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object de- tection. InECCV, 2022. 2, 3

    2022

  20. [20]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 3, 4

    2021

  21. [21]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InICLR, 2019. 4

    2019

  22. [22]

    VSCode: Gen- eral visual salient and camouflaged object detection with 2d prompt learning

    Ziyang Luo, Nian Liu, Wangbo Zhao, et al. VSCode: Gen- eral visual salient and camouflaged object detection with 2d prompt learning. InCVPR, 2024. 2

    2024

  23. [23]

    Simultaneously lo- calize, segment and rank the camouflaged objects

    Yunqiu Lv, Jing Zhang, Yuchao Dai, et al. Simultaneously lo- calize, segment and rank the camouflaged objects. InCVPR,

  24. [24]

    How to evaluate foreground maps? InCVPR, 2014

    Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? InCVPR, 2014. 4

    2014

  25. [25]

    Vo, et al

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V. Vo, et al. DINOv2: Learning robust visual features without supervision.Transactions on Machine Learning Research (TMLR), 2024. 1, 2, 3, 4

    2024

  26. [26]

    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. InCVPR, 2022. 6

    2022

  27. [27]

    Zoom- NeXt: A unified collaborative pyramid network for camou- flaged object detection.IEEE TPAMI, 2024

    Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, et al. Zoom- NeXt: A unified collaborative pyramid network for camou- flaged object detection.IEEE TPAMI, 2024. 2, 6

    2024

  28. [28]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. InICML, 2021. 1

    2021

  29. [29]

    SAM 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, et al. SAM 2: Segment anything in images and videos. InICLR, 2025. 1, 2, 3, 4

    2025

  30. [30]

    U-Net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. InMICCAI, 2015. 2, 4

    2015

  31. [31]

    Hi- era: A hierarchical vision transformer without the bells-and- whistles

    Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, et al. Hi- era: A hierarchical vision transformer without the bells-and- whistles. InICML, 2023. 2

    2023

  32. [32]

    Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, et al

    Oriane Sim ´eoni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, et al. DINOv3.arXiv preprint arXiv:2508.10104, 2025. 1, 2, 3, 4

    Pith/arXiv arXiv 2025

  33. [33]

    Very deep convo- lutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition. InICLR,

  34. [34]

    Animal camouflage analysis: CHAMELEON database

    Przemys law Skurowski, Hassan Abdulameer, et al. Animal camouflage analysis: CHAMELEON database. InUnpub- lished, 2018. 2, 4

    2018

  35. [35]

    Learning to detect salient objects with image-level supervision

    Lijun Wang, Huchuan Lu, Yifan Wang, et al. Learning to detect salient objects with image-level supervision. InCVPR,

  36. [36]

    Pixels, regions, and objects: Multiple enhancement for salient object detection

    Yi Wang, Ruili Wang, Xin Fan, et al. Pixels, regions, and objects: Multiple enhancement for salient object detection. InCVPR, 2023. 6

    2023

  37. [37]

    F3Net: Fu- sion, feedback and focus for salient object detection

    Jun Wei, Shuhui Wang, and Qingming Huang. F3Net: Fu- sion, feedback and focus for salient object detection. InAAAI,

  38. [38]

    Pyramid grafting network for one-stage high-resolution saliency de- tection

    Chenxi Xie, Changqun Xia, Mingcan Ma, et al. Pyramid grafting network for one-stage high-resolution saliency de- tection. InCVPR, 2022. 4

    2022

  39. [39]

    SAM3- UNet: Simplified adaptation of segment anything model 3

    Xinyu Xiong, Zihuang Wu, Lei Lu, and Yufa Xia. SAM3- UNet: Simplified adaptation of segment anything model 3. arXiv preprint arXiv:2512.01789, 2025. 3, 6, 7

    arXiv 2025

  40. [40]

    SAM2-UNet: Segment anything 2 makes strong encoder for natural and medical image segmentation.Visual Intelligence, 4(1):2, 2026

    Xinyu Xiong, Zihuang Wu, Shuangyi Tan, Wenxue Li, Fei- long Tang, Ying Chen, Siying Li, Jie Ma, and Guanbin Li. SAM2-UNet: Segment anything 2 makes strong encoder for natural and medical image segmentation.Visual Intelligence, 4(1):2, 2026. 1, 2, 3, 4, 5, 6, 7

    2026

  41. [41]

    Hierarchical saliency detection

    Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. InCVPR, 2013. 4

    2013

  42. [42]

    MAS-SAM: Segment any marine animal with aggregated features

    Tianyu Yan, Zifu Wan, Xinhao Deng, Pingping Zhang, Yang Liu, and Huchuan Lu. MAS-SAM: Segment any marine animal with aggregated features. InIJCAI, 2024. 2, 5, 6

    2024

  43. [43]

    Saliency de- tection via graph-based manifold ranking

    Chuan Yang, Lihe Zhang, Huchuan Lu, et al. Saliency de- tection via graph-based manifold ranking. InCVPR, 2013. 4

    2013

  44. [44]

    FOCUS: Towards universal foreground segmentation

    Zuyao You, Lingyu Kong, Lingchen Meng, and Zuxuan Wu. FOCUS: Towards universal foreground segmentation. In AAAI, 2025. 2

    2025

  45. [45]

    Towards high-resolution salient object detection

    Yi Zeng, Pingping Zhang, Jianming Zhang, et al. Towards high-resolution salient object detection. InICCV, 2019. 4

    2019

  46. [46]

    Fantastic animals and where to find them: Segment any ma- rine animal with dual SAM

    Pingping Zhang, Tianyu Yan, Yang Liu, and Huchuan Lu. Fantastic animals and where to find them: Segment any ma- rine animal with dual SAM. InCVPR, 2024. 6

    2024

  47. [47]

    Bilateral reference for high-resolution dichotomous image segmenta- tion.CAAI Artificial Intelligence Research, 2024

    Peng Zheng, Dehong Gao, Deng-Ping Fan, et al. Bilateral reference for high-resolution dichotomous image segmenta- tion.CAAI Artificial Intelligence Research, 2024. 1, 2, 4, 6, 7

    2024

  48. [48]

    Salient object detection via integrity learning.IEEE TPAMI, 2023

    Mingchen Zhuge, Deng-Ping Fan, Nian Liu, et al. Salient object detection via integrity learning.IEEE TPAMI, 2023. 6

    2023