Pith. sign in

REVIEW 5 major objections 5 minor 75 references

This paper claims that salient object detection can be solved in a training-free way by organizing the knowledge already present in foundation models, and introduces FOCUS, a pipeline that beats supervised methods on 13 benchmarks.

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 · deepseek-v4-flash

2026-08-03 11:18 UTC pith:OTMHSWI5

load-bearing objection A serious empirical paper with a genuinely useful diagnostic benchmark and a strong training-free pipeline, but the 'zero-shot' claim is oversold because the per-dataset observer protocols leak annotation conventions from the very papers that define the test labels. the 5 major comments →

arxiv 2607.29222 v1 pith:OTMHSWI5 submitted 2026-07-31 cs.CV

Is It Time for the Renaissance of Salient Object Detection in the Era of MLLMs?

Wenzhuo Zhao , Xiuzhi Li , Zhongkuan Mao , Ronghao Xian , Yao Jiang , Zhao Gao , Keren Fu , Qijun Zhao
show 1 more author
Jian Cheng
This is my paper
classification cs.CV
keywords salient object detectionmultimodal large language modelszero-shot learningforeground organizationobserver protocolBayesian surpriseself-supervised featurespromptable 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.

The paper argues that salient object detection no longer needs task-specific supervision: a training-free pipeline that combines a multimodal large language model, a self-supervised feature graph, and a promptable segmenter can organize the knowledge already in foundation models into masks that match or beat fully supervised methods. To support this, the authors build SaliLLM, a diagnostic benchmark that separates saliency localization from segmentation, and show that MLLMs localize salient entities better than SOD state of the art but segment them worse. The paper identifies foreground organization—deciding the right granularity, extent, and inclusion of the mask—as the bottleneck, and introduces FOCUS, which uses Bayesian-surprise calibration of dataset observation protocols and bottom-up perceptual grouping to fix it. Across 13 RGB, RGB-D, and RGB-T benchmarks, FOCUS reduces mean absolute error by 11% versus fully supervised, 34% versus weakly supervised, and 48% versus self-supervised methods. The broader claim is that SOD should shift from learning dataset-specific masks to organizing general visual knowledge under observer protocols.

Core claim

On the paper's own terms, the discovery is that the remaining gap between MLLMs and supervised salient object detection is not perception but organization. SaliLLM shows that MLLMs already know where salient entities are—they outperform supervised methods on box-level localization—but their masks have high coverage and low purity, covering about 2.86 times the correct area. By decomposing the failure into cardinality, granularity, and extent mismatches, the paper shows that these are protocol mismatches: different SOD datasets encode different rules for what counts as foreground. FOCUS addresses this directly by deriving each dataset's observer protocol from its publication text, turning tha

What carries the argument

The central mechanism is protocol-conditioned foreground organization, implemented as a two-pathway attention system. The top-down pathway reads the dataset's published annotation protocol (granularity, extent, inclusion) and converts it into a categorical prior via fixed verbalizers; Bayesian surprise between that prior and the MLLM's default prior produces weights that strengthen protocol dimensions where the model is most mismatched. The bottom-up pathway constructs an entity-centric perceptual manifold from self-supervised features (a k-NN affinity graph and spectral soft charts) and propagates sparse MLLM evidence over it by reweighting visual-token attention with a log(1+s) saliency sc

Load-bearing premise

The load-bearing assumption is that each dataset's observer protocol can be faithfully recovered from its publication text via a small set of fixed verbalizers, and that this protocol aligns with how the ground-truth masks were actually annotated.

What would settle it

Run FOCUS on a held-out dataset while deliberately swapping its protocol description with another dataset's description. If performance does not drop (or if random protocol permutations produce the same MAE distribution), then protocol calibration is not actually driving the gains and the method is relying on the segmenter's default behavior. Conversely, if omitting the protocol text entirely causes MAE to jump back to MLLM-baseline levels, the zero-shot claim is qualified as dataset-aware.

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

If this is right

  • Fully supervised, weakly supervised, and self-supervised SOD methods can be outperformed without any task-specific training, challenging the need for expensive dense mask annotation.
  • The bottleneck for MLLMs in dense prediction is foreground organization, not saliency reasoning, so fine-tuning MLLMs for segmentation may be less effective than improving the organization stage.
  • SOD evaluation should separate localization from segmentation; a model can be right about where but wrong about what to include, and a single mask metric hides that distinction.
  • Observer protocols derived from dataset papers are transferable across RGB, RGB-D, and RGB-T, enabling one frozen framework to handle all three modalities without architectural changes.
  • The same protocol-conditioned Bayesian-surprise mechanism could be used to calibrate any dataset-dependent prior in other zero-shot perception tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The framework is 'zero-shot' only with respect to images and masks: it consumes each dataset's paper text as a protocol description, so in practice it is dataset-aware; a fair comparison with fully unsupervised methods would require withholding the protocol text and measuring the drop.
  • If the protocol verbalizers were replaced by a learned mapping from dataset text to annotation rules, the same architecture could adapt to unseen datasets without per-dataset hand-crafting, at the cost of introducing a training signal.
  • The reported 11%/34%/48% MAE reductions are strong but should be tested for statistical significance per dataset; the largest gains appear in RGB-D and RGB-T, where auxiliary modalities give extra contrast, so the protocol-calibration contribution may vary by modality.
  • The dependence on fixed self-supervised features and a promptable segmenter means FOCUS's ceiling is tied to those frozen components; as better foundation models appear, the framework should inherit their gains without retraining.

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

5 major / 5 minor

Summary. The paper argues that multimodal large language models (MLLMs) already possess strong zero-shot salient-entity localization but are much weaker at segmentation, and that this gap is due to mismatches in foreground cardinality, granularity, and extent. To support this, the authors construct SaliLLM, a diagnostic benchmark that augments existing RGB/RGB-D/RGB-T SOD datasets with phrases, boxes, categories, and attributes, and they present analyses of MLLM internal representations. They then propose FOCUS, a training-free framework that couples a top-down per-dataset 'observer-protocol' prior with bottom-up propagation of MLLM evidence over DINOv3-derived perceptual manifolds, using SAM3 to segment the resulting prompt and a Protocol-Conditioned Bayesian Surprise (PCBS) module to select among mask candidates. Across 13 RGB, RGB-D, and RGB-T benchmarks, FOCUS is reported to generally surpass fully, weakly, and self-supervised SOD methods, with mean absolute error reductions of 11%, 34%, and 48%, respectively.

Significance. If the reported results hold, the paper would make a strong case that task-specific SOD supervision can be replaced by a combination of frozen foundation models and hand-derived dataset priors. The diagnostic SaliLLM benchmark and the layer-wise analysis of MLLM saliency are useful contributions in themselves, and the paper is commendably explicit about excluding benchmark images and label statistics from protocol construction. The strengths are the breadth of evaluation (13 datasets across three modalities), the use of fixed frozen backbones, and the attempt to decompose SOD into localization and segmentation. However, the central 'zero-shot without training' claim rests on per-dataset observer protocols whose derivation is not independently validated, and the localization comparison in the diagnostic benchmark is not apples-to-apples. These issues are load-bearing and require additional experiments or careful reframing.

major comments (5)
  1. [Top-Down Observer-Protocol Calibration (Eq. (2), Table 5)] The per-dataset protocol p_D is derived from dataset papers and encodes granularity, extent, and inclusion—exactly the properties that define the ground-truth annotation rules of each benchmark. Table 5 shows that the observer prompt/protocol components provide the largest gains (V2→V3), so the reported superiority over supervised methods may be substantially due to dataset-specific prior knowledge rather than generic foreground organization. The paper provides no ablation with a single generic protocol across all datasets, and no evidence that protocols can be reliably recovered from text alone by an independent reader. Please add (i) a 'generic protocol' version of FOCUS, (ii) a test of protocol-extraction reliability, or (iii) a clear statement that FOCUS is not dataset-agnostic but uses per-dataset protocol priors. Without this, the 'without training'/'zero-shot' claims are overstate
  2. [Figure 1, SaliLLM localization comparison] The claim that MLLMs outperform SOTA SOD methods in localization is based on comparing boxes generated directly by MLLMs with boxes obtained by fitting tight boxes to the predicted masks of fully supervised SOD models. These are not equivalent tasks: SOD models are trained for mask quality, not box regression, and mask-derived boxes inherit all mask errors (both over- and under-segmentation). This mismatch undermines the diagnostic conclusion that MLLMs are already better at localization. Please evaluate supervised models with a box head or an equivalent localization mechanism, or compare MLLM box outputs through the same mask-to-box conversion pipeline, and report the results.
  3. [Experiments, Tables 2–4] The central quantitative claims ('generally surpasses SOTA', 11%/34%/48% MAE reductions) are made without any error bars, confidence intervals, or significance tests. Although deterministic decoding reduces run-to-run variance, the comparisons are across different papers and implementations, and the aggregate percentages are not defined precisely (which datasets/methods are averaged, and are they paired across the same images?). Please report per-dataset deviations, paired tests across datasets, or at least a clear aggregation formula. Without this, the strength of the headline claim cannot be assessed.
  4. [Implementation Details; Eqs. (3)–(5)] The method depends on a large number of hyperparameters and undefined components: manifold scales τ_f, σ_p, σ_a, chart count K_D, PCBS coefficients β, gate limits λ_min/λ_max, SAM3 candidate budget J, and the definitions of 'support recall', 'chart coherence', 'leakage', and 'fragmentation' in r_D(m). The text says these are 'provided in the supplementary material', but no supplement is included with the manuscript. Since the claim that all settings are 'fixed globally and shared across datasets' is central to the training-free argument, please provide the full configuration and exact definitions in the main text or an available supplement.
  5. [Eq. (1), Method overview] The method is framed as the solution to a KL-projection optimization, but the paper states that 'one visual-token update suffices' and then implements Eq. (4) as a single per-row softmax tilt. No justification or sensitivity study is given for this approximation. If this step is a heuristic, the formal framing in Eq. (1) is misleading; if it is meant as a exact projection, it needs proof or empirical validation. Please clarify the role of Eq. (1) and report sensitivity to this approximation.
minor comments (5)
  1. [SaliLLM Data Annotation] The human verification process is described but no inter-annotator agreement or quality metrics are reported. Since SaliLLM is a new benchmark, please add agreement statistics or at least a sample-level consistency measure.
  2. [Abstract and Conclusion] The phrase 'generally surpasses' and 'strongest overall or average performance' are vague. Specify whether the claim is average rank, mean per dataset, or the majority of datasets.
  3. [Table 1] The error-decomposition columns (S+/B−, S−/B+) are not defined clearly; please spell out the notation and the matching procedure used for one-to-one entity matching.
  4. [References] Reference [60] appears to have a truncated title ('for Salient Object Detection') and reference [18] is cited as both 'HyPSAM' and 'DFNet'; please verify these entries.
  5. [Figure 3] Panel labels (C1, C2, C3, D, E) are referenced in text but are not all visually labeled in the figure; please align text callouts with the figure panels.

Circularity Check

1 steps flagged

Observer protocols are read from the same annotation papers that define the test masks, so the zero-shot gains partly encode label semantics.

specific steps
  1. self definitional [Sec. 1 (Introduction); Sec. 3.2 Top-Down Observer-Protocol Calibration, Eq. (2); mask selection in Eq. (5)]
    "Different datasets encode distinct preferences regarding foreground entities, granularity, and spatial extent. We recover this semantic prior only from original dataset papers and public acquisition or annotation descriptions; benchmark images and derived label statistics are excluded from protocol construction."

    The protocol prior p_D = T(R_D, ρ_D) is built deterministically from R_D, the dataset papers and annotation descriptions that specify how the test ground-truth masks were produced. Its categorical dimensions (single/group foreground, tight/broad extent, inclusion/exclusion) are exactly the annotation rules that generated the target masks. Eq. (5) then reranks SAM candidates using protocol likelihood, so the supposedly training-free prediction is conditioned on a compressed encoding of the test-label generation rule. This is not a numerical fit to masks, so it is a mild definitional leakage rather than a tautology, but the reported 11%/34%/48% MAE reductions are not a pure measure of zero-shot foreground organization.

full rationale

FOCUS itself performs no SOD-specific training and no numerical fitting to benchmark masks; the SaliLLM diagnostic benchmark is independently constructed; the bottom-up manifold injection and PCBS selection use fixed, shared settings. No load-bearing self-citation or imported uniqueness theorem is present. The one substantive circularity burden is the top-down observer protocol: p_D is extracted from the very annotation descriptions that define the test masks, using dimensions (granularity, extent, inclusion) along which the masks vary. Because p_D enters mask selection through Eq. (5), the method embeds label semantics in its 'zero-shot' prior. This is a real but bounded form of circularity: the paper is transparent about the exclusion of images and label statistics, but the protocol source is not independent of the targets.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central empirical claim rests on per-dataset protocol priors, several hand-set hyperparameters, and an assumption that frozen DINOv3 and SAM3 provide the right bottom-up grouping and candidate masks. None of these are fitted to ground-truth masks, but several are not fully specified in the main text.

free parameters (5)
  • Per-dataset observer-protocol priors (z_g, z_r, z_e) = Categorical choices (single/group, tight/broad, include/exclude) per dataset; not numerically specified in main text
    Hand-mapped by fixed verbalizers from dataset papers; act as per-dataset inference-time priors on foreground granularity, extent, and inclusion.
  • PCBS coefficient vector β and gate limits λ_min, λ_max = Not reported in main text ('fixed before evaluation')
    Eq. (5) uses β and per-image gate λ_X = λ_min + (λ_max - λ_min) σ(η_X); values are shared across datasets but deferred to the supplement.
  • Manifold scales τ_f, σ_p, σ_a and chart count K_D = Not reported; in supplementary
    Eq. (3) uses feature/chart scales and Top-K chart selection; fixed globally before evaluation, but hand-chosen hyperparameters.
  • SAM3 candidate budget J and other candidate-generation settings = Not reported in main text
    PCBS selects among J mask candidates; candidate budget affects the ceiling of achievable performance.
  • Auxiliary modality fusion settings (PHFuse for RGB-T) = Not reported
    An external frozen fusion network is used; its settings affect RGB-T results.
axioms (5)
  • domain assumption Dataset papers encode faithful observer-protocol choices.
    Sec. 4.1 assumes fixed verbalizers can convert natural-language dataset descriptions into discrete granularity/extent/inclusion choices; if papers are ambiguous, uniform prior is used.
  • domain assumption DINOv3 dense features induce entity-centric perceptual manifolds whose top-K charts approximate salient object extents.
    Sec. 4.2; the entire bottom-up pathway relies on this empirical property of self-supervised features and on k-NN graph/spectral chart settings.
  • domain assumption MLLM-generated phrase+box evidence is a reliable saliency prior and can be steered by manifold support without retraining.
    Sec. 4 and Fig. 1; FOCUS assumes Qwen3-VL's phrase and box outputs are trustworthy starting points for foreground organization.
  • domain assumption SAM3's candidate mask set Q(h*) contains a protocol-aligned mask for each salient entity.
    Sec. 4.3; the final mask is selected from SAM3 candidates; if the correct region is not in Q(h*), FOCUS cannot recover it.
  • ad hoc to paper One visual-token update suffices to approximate the KL projection in Eq. (1).
    The paper states 'We approximate Eq. (1) with p_D and one visual-token update' without convergence or optimality analysis.
invented entities (1)
  • Observer protocol variable p_D no independent evidence
    purpose: Abstracts dataset-specific annotation policies (granularity, extent, inclusion) to condition inference and mask selection.
    A modeling construct introduced by the paper; it is not independently measurable and its mapping from papers is hand-authored, so it cannot be falsifiably verified outside the paper.

pith-pipeline@v1.3.0-daily-deepseek · 14248 in / 15203 out tokens · 153859 ms · 2026-08-03T11:18:23.763926+00:00 · methodology

0 comments
read the original abstract

The zero-shot capabilities of multimodal large language models (MLLMs) are pushing salient object detection (SOD) beyond task-specific supervision. To disentangle MLLMs beyond conventional mask-based evaluation, we decompose SOD into localization and segmentation, and re-engineer datasets with phrases, boxes, and attributes, establishing a diagnostic benchmark for MLLM saliency perception (SaliLLM). SaliLLM uncovers a striking capability mismatch: MLLMs outperform state-of-the-art (SOTA) methods in localization, yet remain substantially weaker in segmentation. Further analyses attribute this gap primarily to mismatches between MLLMs and annotations over foreground cardinality, granularity, and extent. Motivated by this diagnosis, we recast zero-shot SOD as protocol-aligned Foreground Organization and introduce the first training-free framework that leverages Gestalt-inspired Collaborative attention for Unified SOD (FOCUS). FOCUS couples top-down Bayesian-surprise calibration of protocol-conditioned foreground granularity with bottom-up propagation of MLLMs evidence over entity-centric perceptual manifolds induced by self-supervised features, yielding coherent object extents as prompts for a general segmenter. Across 13 RGB, RGB-D, and RGB-T SOD benchmarks, FOCUS generally surpasses SOTA methods without training, reducing mean absolute error by 11\%, 34\%, and 48\% compared with fully, weakly, and self-supervised methods, respectively. Our findings signal the renaissance of SOD: from task-specific supervision to zero-shot foreground organization. Code is available in the supplementary material.

Figures

Figures reproduced from arXiv: 2607.29222 by Jian Cheng, Keren Fu, Qijun Zhao, Ronghao Xian, Wenzhuo Zhao, Xiuzhi Li, Yao Jiang, Zhao Gao, Zhongkuan Mao.

Figure 1
Figure 1. Figure 1: Rethinking SOD with MLLMs: In contrast to annotation- and training-intensive SOD methods, our diagnostic [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the SaliLLM benchmark: construction pipeline, semantic taxonomy, and dataset statistics. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of MLLMs on SaliLLM: progressive salient object perception and incomplete spatial closure in segmentation. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FOCUS: protocol-conditioned foreground organization through collaborative top-down and bottom-up attention. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] 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

75 extracted references · 6 linked inside Pith

  1. [1]

    Achanta, R.; Hemami, S.; Estrada, F.; and Susstrunk, S. 2009. Frequency-tuned salient region detection. In CVPR, 1597–1604. IEEE

    2009

  2. [2]

    Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng, L.; Ding, W.; Gao, C.; Ge, C.; et al

  3. [3]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; et al

  4. [4]

    V.; Khedr, H.; Huang, A.; et al

    Carion, N.; Gustafson, L.; Hu, Y.-T.; Debnath, S.; Hu, R.; Suris, D.; Ryali, C.; Alwala, K. V.; Khedr, H.; Huang, A.; et al. 2025. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719

    Pith/arXiv arXiv 2025

  5. [5]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923

    Pith/arXiv arXiv

  6. [6]

    Cong, R.; Chen, Z.; Fang, H.; Kwong, S.; and Zhang, W. 2025. Breaking Barriers, Localizing Saliency: A Large-scale Benchmark and Baseline for Condition- Constrained Salient Object Detection.IEEE TPAMI

    2025

  7. [7]

    Chen, J.; Li, G.; Zhang, Z.; Chang, L.; and Zeng, D

  8. [8]

    Fan, D.; Gong, C.; Cao, Y.; Ren, B.; Cheng, M.; and Borji, A. 2018. Enhanced-alignment Measure for Bi- nary Foreground Map Evaluation. InIJCAI

    2018

  9. [9]

    Fan, D.; Lin, Z.; Zhang, Z.; Zhu, M.; and Cheng, M. 2020. Rethinking RGB-D salient object detection: Models,datasets,andlarge-scalebenchmarks.TNNLS, 32(5): 2075–2089

    2020

  10. [10]

    Ding, Y.; Chen, W.; Zhang, G.; Feng, Z.; and Li, X

  11. [11]

    Cross-ModalWeaklySupervisedRGB-DSalient Object Detection with a Focus on Filamentary Struc- tures.Sensors, 25(10): 2990

  12. [12]

    Highly efficient RGB-D salient object detection with adaptivefusionandattentionregulation.IEEETCSVT

    Gao,H.;Wang,F.;Wang,M.;Sun,F.;andLi,H.2024. Highly efficient RGB-D salient object detection with adaptivefusionandattentionregulation.IEEETCSVT

    2024

  13. [13]

    Gao,L.;Liu,B.;Fu,P.;Xu,M.;Zhang,Y.;andHuang, Y. 2025. Self-Supervised Pre-training with Multi- modality Representation Enhancement for Salient Ob- ject Detection in RGB-D Images.IEEE TIM

    2025

  14. [14]

    Fan, D.-P.; Cheng, M.-M.; Liu, Y.; Li, T.; and Borji, A. 2017. Structure-measure: A new way to evaluate foreground maps. InICCV, 4548–4557

    2017

  15. [15]

    Shifting more attention to video salient object detec- tion

    Fan,D.-P.;Wang,W.;Cheng,M.-M.;andShen,J.2019. Shifting more attention to video salient object detec- tion. InCVPR, 8554–8564

    2019

  16. [16]

    He, J.; Fu, K.; Liu, X.; and Zhao, Q. 2025. Samba: A unified mamba-based framework for general salient object detection. InCVPR, 25314–25324

    2025

  17. [17]

    Hou, J.; Wang, S.; Zhao, Q.; Ma, J.; Sheng, X.; Sun, Y.; Wen, H.; and Yan, C. 2026. Empirical study on fusionstrategyinRGB-Tsalientobjectdetection.IEEE TCSVT

    2026

  18. [18]

    Geva, M.; Caciularu, A.; Wang, K.; and Goldberg, Y

  19. [19]

    Hu, J.; Lin, J.; Yan, J.; and Gong, S. 2024. Leveraging hallucinations to reduce manual prompt dependency in promptable segmentation.NeurIPS, 37: 107171– 107197

    2024

  20. [20]

    Guan, H.; Lin, J.; and Lau, R. W. 2025. A contrastive- learningframeworkforunsupervisedsalientobjectde- tection.IEEE TIP

    2025

  21. [21]

    I.; Zhu, X.; and Ullah, I

    Khan, R.; Alzaben, N.; Daradkeh, Y. I.; Zhu, X.; and Ullah, I. 2025. Pyramidal attention with progres- sive multi-stage iterative feature refinement for salient object segmentation.Image and Vision Computing, 105670

    2025

  22. [22]

    Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.;Gustafson,L.;Xiao,T.;Whitehead,S.;Berg,A.C.; Lo, W.-Y.; et al. 2023. Segment anything. InICCV, 4015–4026

    2023

  23. [23]

    Hou, R.; Li, X.; Ren, T.; Zhou, D.; Wu, G.; and Cao, J. 2025. HyPSAM: Hybrid Prompt-driven Segment Anything Model for RGB-Thermal Salient Object De- tection.IEEE TCSVT

    2025

  24. [24]

    Li, G.; and Yu, Y. 2015. Visual saliency based on multiscale deep features. InCVPR, 5455–5463

    2015

  25. [25]

    Ju, R.; Ge, L.; Geng, W.; Ren, T.; and Wu, G. 2014. Depth saliency based on anisotropic center-surround difference. InICIP, 1115–1119. IEEE

    2014

  26. [26]

    Li,Y.;Hou,X.;Koch,C.;Rehg,J.M.;andYuille,A.L

  27. [27]

    Liu, N.; Luo, Z.; Zhang, N.; and Han, J. 2024. Vst++: Efficientandstrongervisualsaliencytransformer.IEEE TPAMI, 46(11): 7300–7316

    2024

  28. [28]

    Li, A.; Mao, Y.; Zhang, J.; and Dai, Y. 2023. Mu- tual information regularization for weakly-supervised RGB-D salient object detection.IEEE TCSVT, 34(1): 397–410

    2023

  29. [29]

    SAMSOD: Rethinking SAM Optimization for RGB-T Salient Object Detection.IEEE TMM

    Liu,Z.;Wang,X.;Fang,X.;Tu,Z.;andWang,L.2026. SAMSOD: Rethinking SAM Optimization for RGB-T Salient Object Detection.IEEE TMM

    2026

  30. [30]

    Li, L.; Ji, S.; Luo, Z.; Li, Z.; Zhang, D.; Han, J.; and Liu, N. 2025. Saliency-R1: Incentivizing Uni- fied Saliency Reasoning Capability in MLLM with Confidence-Guided Reinforcement Learning.arXiv preprint arXiv:2511.00396

    arXiv 2025

  31. [31]

    Luo, Z.; Liu, N.; Zhao, W.; Yang, X.; Zhang, D.; Fan, D.-P.; Khan, F.; and Han, J. 2024. VSCode: General VisualSalientandCamouflagedObjectDetectionwith 2D Prompt Learning. InCVPR, 17169–17180. IEEE

    2024

  32. [32]

    Leveraging stereopsis for saliency analysis

    Niu,Y.;Geng,Y.;Li,X.;andLiu,F.2012. Leveraging stereopsis for saliency analysis. InCVPR, 454–461

    2012

  33. [33]

    Peng, H.; Li, B.; Xiong, W.; Hu, W.; and Ji, R. 2014. RGBD salient object detection: A benchmark and al- gorithms. InECCV, 92–109. Springer

    2014

  34. [34]

    Liu, Z.; Deng, S.; Wang, X.; Wang, L.; Fang, X.; and Tang, B. 2025. SSFam: Scribble supervised salient object detection family.IEEE TMM

    2025

  35. [35]

    Piao, Y.; Wu, W.; Zhang, M.; Jiang, Y.; and Lu, H

  36. [36]

    VSCode-V2:DynamicPrompt Learning for General Visual Salient and Camouflaged Object Detection With Two-Stage Optimization.IEEE TPAMI

    Luo,Z.;Liu,N.;Yang,X.;Zhang,D.;Fan,D.-P.;Khan, F.S.;andHan,J.2025. VSCode-V2:DynamicPrompt Learning for General Visual Salient and Camouflaged Object Detection With Two-Stage Optimization.IEEE TPAMI

    2025

  37. [37]

    V.; Seitzer, M.; Baldassarre, F.; Oquab, M.; Jose, C.; Khalidov, V.; Szafraniec, M.; Yi, S.; Ramamonjisoa, M.; et al

    Siméoni, O.; Vo, H. V.; Seitzer, M.; Baldassarre, F.; Oquab, M.; Jose, C.; Khalidov, V.; Szafraniec, M.; Yi, S.; Ramamonjisoa, M.; et al. 2025. Dinov3.arXiv preprint arXiv:2508.10104

    Pith/arXiv arXiv 2025

  38. [38]

    Tan, H.; Zhang, Y.; Zhang, L.; Li, J.; Hu, T.; and Wu, F. 2026. Multi-Modal Hierarchical Fusion with Cross- AgentforRGB-DSalientObjectDetection. InICASSP, 12952–12956. IEEE

    2026

  39. [39]

    Tang, H.; Li, Z.; Zhang, D.; He, S.; and Tang, J. 2024. Divide-and-conquer:Confluenttriple-flownetworkfor RGB-T salient object detection.IEEE TPAMI, 47(3): 1958–1974

    2024

  40. [40]

    Piao, Y.; Ji, W.; Li, J.; Zhang, M.; and Lu, H. 2019. Depth-induced multi-scale recurrent attention network for saliency detection. InICCV, 7254–7263

    2019

  41. [41]

    Tu, Z.; Ma, Y.; Li, Z.; Li, C.; Xu, J.; and Liu, Y. 2022. RGBT salient object detection: A large-scale dataset and benchmark.IEEE TMM, 25: 4163–4176

    2022

  42. [42]

    Noise-sensitive adversarial learning for weakly supervised salient object detection.IEEE TMM, 25: 2888–2897

  43. [43]

    Qi, Y.; Guo, R.; Li, Z.; Niu, D.; and Qu, L. 2024. Masked visual pre-training for RGB-d and RGB-t salient object detection. InChinese conference on pat- tern recognition and computer vision (PRCV), 49–66. Springer

    2024

  44. [44]

    Wang, G.; Li, C.; Ma, Y.; Zheng, A.; Tang, J.; and Luo, B. 2018. RGB-T saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In IGTA, 359–369. Springer

    2018

  45. [45]

    Wang, L.; Lu, H.; Wang, Y.; Feng, M.; Wang, D.; Yin, B.; and Ruan, X. 2017. Learning to detect salient ob- jectswithimage-levelsupervision. InCVPR,136–145

    2017

  46. [46]

    Wang, W.; Gao, Z.; Gu, L.; Pu, H.; Cui, L.; Wei, X.; Liu, Z.; Jing, L.; Ye, S.; Shao, J.; et al. 2025. In- ternvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265

    Pith/arXiv arXiv 2025

  47. [47]

    Tang, H.; Xie, C.-W.; Wang, H.; Bao, X.; Weng, T.; Li, P.; Zheng, Y.; and Wang, L. 2026. Ufo: A unified approach to fine-grained visual perception via open- ended language interface.NeurIPS, 38: 83761–83791

    2026

  48. [48]

    X.; and Misra, I

    Wang, X.; Girdhar, R.; Yu, S. X.; and Misra, I. 2023. Cut and learn for unsupervised object detection and instance segmentation. InCVPR, 3124–3134

    2023

  49. [49]

    Tu, Z.; Xia, T.; Li, C.; Wang, X.; Ma, Y.; and Tang, J

  50. [50]

    Xia,Z.;Han,D.;Han,Y.;Pan,X.;Song,S.;andHuang, G. 2024. Gsva: Generalized segmentation via multi- modal large language models. InCVPR, 3858–3869

    2024

  51. [51]

    Wan, B.; Cong, R.; Zhou, X.; Fang, H.; Lv, C.; and Kwong, S. 2026. Rsonet: Region-guided selective op- timization network for rgb-t salient object detection. IEEE TCSVT

    2026

  52. [52]

    Yang, C.; Zhang, L.; Lu, H.; Ruan, X.; and Yang, M.- H. 2013. Saliency detection via graph-based manifold ranking. InCVPR, 3166–3173

    2013

  53. [53]

    Yang, Y.; Tian, J.; Zhang, S.; and Cao, L. 2026. Dis- cover, Segment, and Select: A Progressive Mechanism for Zero-shot Camouflaged Object Segmentation. In CVPR, 34745–34754

    2026

  54. [54]

    M.; and Sadeghian,A.2024

    Yasarla, R.; Weng, R.; Choi, W.; Patel, V. M.; and Sadeghian,A.2024. 3SD:self-supervisedsaliencyde- tection with no labels. InWACV, 313–322

    2024

  55. [55]

    Wang,W.;Lai,Q.;Fu,H.;Shen,J.;Ling,H.;andYang, R. 2021. Salient object detection in the deep learning era: An in-depth survey.IEEE TPAMI, 44(6): 3239– 3259

    2021

  56. [56]

    Yuan, Y.; Liu, W.; Gao, P.; Dai, Q.; and Qin, J

  57. [57]

    Wang,Y.;Wang,R.;He,X.;Lin,C.;Wang,T.;Jia,Q.; and Fan, X. 2024. WBNet: Weakly-supervised salient object detection via scribble and pseudo-background priors.Pattern Recognition, 154: 110579

    2024

  58. [58]

    Zhan, Y.; Zeng, Z.; Liu, H.; Tan, X.; and Tian, Y

  59. [59]

    Yan, Q.; Xu, L.; Shi, J.; and Jia, J. 2013. Hierarchical saliency detection. InCVPR, 1155–1162

    2013

  60. [60]

    Zhang, S.; and Wang, C. 2026. for Salient Object De- tection. InPattern Recognition and Computer Vision: 8thChineseConference,PRCV2025,Shanghai,China, October15-18,2025,Proceedings,PartII,31.Springer Nature

    2026

  61. [61]

    Q.; and Artzi, Y

    Zhang, T.; Kishore, V.; Wu, F.; Weinberger, K. Q.; and Artzi, Y. 2019. Bertscore: Evaluating text generation with bert.arXiv preprint arXiv:1904.09675

    Pith/arXiv arXiv 2019

  62. [62]

    Zhou, Q.; Wang, J.; Li, J.; Zhou, C.; Hu, H.; and Hu, K. 2025. RMFDNet: Redundant and missing feature decoupling network for salient object detection.En- gineering Applications of Artificial Intelligence, 139: 109459

    2025

  63. [63]

    Yin, C.; Li, H.; Yang, K.; Li, J.; Zhu, P.; and Li, X

  64. [64]

    InACM MM, 3741–3750

    Stepwise Decomposition and Dual-stream Fo- cus: A Novel Approach for Training-free Camouflaged Object Segmentation. InACM MM, 3741–3750

  65. [67]

    Zhai, S.; Liu, C.; Tu, Z.; Li, C.; and Gao, L. 2025. Weakly Supervised RGBT salient object detection via SAM-Guided Label Optimization and Progressive Cross-modal Cross-scale Fusion.Information Fusion, 120: 103048

    2025

  66. [69]

    Mambasod: Dual mamba-driven cross-modal fusion network for rgb-d salient object detection.Neu- rocomputing, 631: 129718

  67. [70]

    Zhang, S.; Huang, J.; Tang, W.; Wu, Y.; Hu, T.; Xu, X.; and Liu, J. 2025. DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection. InAAAI, 10103–10111

    2025

  68. [74]

    Zhou,T.;Fan,D.-P.;Cheng,M.-M.;Shen,J.;andShao, L. 2021. RGB-D salient object detection: A survey. Computational Visual Media, 7(1): 37–69

    2021

  69. [75]

    Zhuang,C.;Kuang,H.;Wang,H.;Wen,C.;Liu,C.;and Yuan,G.2025.PHFuse:Unsupervisedcolorvisibleand infrared image fusion with preserved hue.Scientific Reports, 15(1): 31458

    2025

  70. [2014]

    In CVPR, 280–287

    The secrets of salient object segmentation. In CVPR, 280–287

  71. [2019]

    RGB-T image saliency detection via collabora- tive graph learning.IEEE TMM, 22(1): 160–173

  72. [2022]

    InProceedings of the 2022 conference on empirical methods in natural language processing, 30–45

    Transformer feed-forward layers build predic- tions by promoting concepts in the vocabulary space. InProceedings of the 2022 conference on empirical methods in natural language processing, 30–45

    2022

  73. [2024]

    InIJCAI, 1616–1624

    Unified unsupervised salient object detection via knowledge transfer. InIJCAI, 1616–1624

  74. [2025]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631

    Pith/arXiv arXiv

  75. [2026]

    STENet: Superpixel token enhancing network for RGB-D salient object detection.IEEE TMM