Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Graph-Based Uncertainty Modeling and Multimodal Fusion for Salient Object Detection

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper claims that DUP-MCRNet, built on dynamic uncertainty graph convolution and learnable gating of RGB, depth, and edge attention, outperforms prior salient object detection methods on most standard benchmarks, especially on edge clar

desk verdict Promising graph-based uncertainty module, but the multimodal fusion story is untested: the experiments use only RGB data while Eqs. 8-9 require depth and edge streams, and the tables contain errors that contradict the 'outperforms most' claim. read the letter →

arxiv 2508.20415 v1 pith:NK6JKXJR submitted 2025-08-28 cs.CV

classification cs.CV
keywords SalientObjectDetectionDynamicUncertaintyGraphConvolutionMultimodalFusionCollaborativeReasoningEdgePreservationBenchmarkEvaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that salient object detection can be improved by treating uncertainty as a message passed between feature layers and by letting the network reweight RGB, depth, and edge evidence per image. Its first module builds a sparse graph over spatial positions using a blend of coordinate distance and feature similarity, runs three residual graph convolutions to propagate uncertainty, and uses the result as a pixel-wise gate that emphasizes uncertain boundaries and small structures. Its second module applies self-attention separately to RGB, depth, and edge features and fuses the resulting maps with softmax-normalized learnable weights, so the model can suppress unreliable modalities. The model is trained with multi-scale BCE and IoU losses plus a cross-scale consistency term. The paper reports that on five standard RGB benchmarks the full model is best or second-best on most metrics, winning all four metrics on ECSSD and HKU-IS.

What carries the argument

The load-bearing components are two modules. DUGC (Dynamic Uncertainty Graph Convolution) constructs a sparse graph on the feature map, with adjacency Aij = 1 only for the K nearest neighbors under Dcomb(i,j) = α‖Pi−Pj‖2 + (1−α)(1 − cosine similarity between features). Three graph-convolution layers X(t+1) = X(t) + ReLU(D̂^{−1/2}ÂD̂^{−1/2}X(t)W_g^{(t)}) propagate uncertainty across nodes; a light MLP turns the propagated features into a pixel-level uncertainty weight map that multiplies the input features, focusing computation on edges and small structures. MCF (Multimodal Collaborative Fusion) maps RGB, depth, and edge features to a shared dimension, runs self-attention on each, and combine

What would settle it

Train the released model with the MCF gate weights frozen to zero for depth and edge (effectively removing those inputs) and evaluate on DUTS and ECSSD; if performance does not drop, the multimodal fusion contribution is not carrying the reported gains.

Watch

Extended reading notes

Core claim

DUP-MCRNet makes two substantive claims. First, uncertainty can be propagated dynamically through a sparse graph whose nodes are spatial positions and whose edges are drawn from the top-K smallest combinations of pixel coordinate distance and feature dissimilarity. Three symmetric-normalized residual graph-convolution layers let uncertain regions exchange information across feature scales, and an MLP readout produces a per-pixel uncertainty weight that amplifies features where boundaries or small objects are at risk. Second, multimodal evidence should be fused by attention and gating: RGB, depth, and edge features each pass through self-attention, and the three attention maps are combined wi

Load-bearing premise

The multimodal fusion claim depends on depth and edge streams actually being available during the experiments, but the paper evaluates only RGB images and never specifies how those maps are produced.

Editorial extensions

If this is right

  • If DUGC works as claimed, boundary details and small-object detection can be improved without external edge supervision, because the graph propagates uncertainty learned from data.
  • If MCF's gating works as claimed, applications with RGB-D or RGB-D-edge sensors can adaptively emphasize the reliable modality in occluded or low-texture scenes.
  • The multi-scale BCE+IoU plus cross-scale consistency loss is a transferable training recipe for multi-scale dense prediction decoders.
  • The ablation pattern—each module helps alone and together they help more—means the two mechanisms are complementary, not redundant.

Reading between the lines

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

  • Beyond the paper: since the experiments are RGB-only and never state how depth or edge maps are produced, the MCF contribution would need a dedicated RGB-D or RGB-D-edge benchmark (or a synthetic depth/edge pipeline) to verify that the gating weights actually matter.
  • Beyond the paper: the same sparse-graph uncertainty propagation could be applied to other dense prediction tasks with boundary ambiguity, such as semantic segmentation or monocular depth estimation, where the per-pixel uncertainty map could act as a learned attention signal.
  • Beyond the paper: the learned gate weights θ could be inspected per image as a reliability signal, potentially indicating when the model trusts depth over color; that would turn the fusion module into an interpretable confidence estimator.
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

3 major / 4 minor

Summary. The paper proposes DUP-MCRNet for salient object detection, with two main contributions: a Dynamic Uncertainty Graph Convolution (DUGC) module that propagates features along a sparse graph built from spatial and semantic distances, and a Multimodal Collaborative Fusion (MCF) strategy that fuses self-attentive features from RGB, depth, and edge inputs using learnable modality-gating weights. Training uses multi-scale BCE and IoU losses, a cross-scale consistency loss, and uncertainty-aware mask supervision. The method is evaluated on five RGB-only benchmarks (DUT-O, DUTS, ECSSD, HKU-IS, PASCAL-S) plus ablation results on SOD, and the paper claims improvements over several recent SOD methods, especially in edge clarity and robustness to complex backgrounds.

Significance. If validated, the DUGC and MCF modules could make a useful contribution to salient object detection, particularly for boundary refinement and cross-modal fusion. The graph-based uncertainty propagation idea is interesting, the loss design is comprehensive, and the code is publicly available. However, the experiments as presented do not exercise the multimodal part of the system: all datasets are RGB-only, and no depth or edge input generation is described. Consequently, the central multimodal claim is currently untested, and the reported quantitative results cannot be attributed to the MCF module as specified. The DUGC component is plausible, but its benefit is entangled with the unvalidated multimodal module in the ablations.

major comments (3)
  1. [Section 3.2, Eqs. (8)–(9)] MCF is defined for three modality inputs F_out^(m), m=1,2,3 (RGB, Depth, Edge) and fuses them via Eqs. (8)-(9). Section 4.1 lists only RGB-only datasets, and Section 4.2 gives no depth-estimation or edge-extraction pipeline. It is therefore impossible to know whether the experiments actually use the described three-stream input. If depth and edge maps are generated from RGB, the generator must be specified; if they are not used, Eq. (9) degenerates to a weighted sum of RGB-derived attention maps, and the central multimodal contribution is unvalidated. The Limitation section (Sec. 5) even refers to 'multimodal saliency detection benchmarks,' while all evaluation datasets are RGB-only, an internal inconsistency.
  2. [Table 1 and abstract/conclusion] The abstract and conclusion claim that DUP-MCRNet 'outperforms various SOD methods on most common benchmark datasets.' In Table 1, Ours has the worst MAE on DUT-O (0.066 vs. LDF's 0.052) and is not best on DUTS (MAE 0.036 vs. 0.034 for LDF/UGRAN) or PASCAL-S (MAE 0.061 vs. 0.051 for LDF). The wins on ECSSD and HKU-IS are real, but 'most' overstates the evidence. The claims should be qualified with per-dataset and per-metric caveats, or supported by significance testing.
  3. [Section 4.4, Table 2] The text states that combining DUGC and MCF achieves the best results on almost all indicators, but the table shows several contradictions. On ECSSD, the combined model's MAE (0.030) is worse than DUGC-only (0.028); on HKU-IS, the combined model's MAE (0.027) is worse than MCF-only (0.026); and on ECSSD, DUGC-only decreases Sm and Fw relative to Base. This does not clearly support the claimed complementarity. The authors should either correct the numbers, add statistical significance, or provide a more nuanced interpretation of the ablation.
minor comments (4)
  1. [Section 4.1] The text says the model is evaluated on six benchmark datasets, but Table 1 reports only five, with SOD appearing only in the ablation study. Please clarify the train/test split and which datasets are used for the main comparison.
  2. [Table 1] Header contains a typo: 'P ACSAL-S' should be 'PASCAL-S'.
  3. [Eq. (13)] The detach operation is described as preventing gradient propagation to higher-level predictions. In practice, detach(·) stops gradients flowing into the detached target; please clarify the intended direction of gradient blocking.
  4. [Eq. (5)] The 'pixel-level uncertainty weight map' is the output of a learned MLP applied to graph-convolved features. This is an operational definition, but 'uncertainty' is potentially misleading. A short explicit definition would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DUGC and MCF are learnable modules fitted to supervised losses, and the central comparison is against external benchmarks.

full rationale

The paper's derivation chain is self-contained. The DUGC module constructs a sparse graph from spatial and feature distances (Eqs. 1-3), propagates along it with graph convolution (Eq. 4), and produces a pixel-level weight via an MLP (Eq. 5); all of these are learned functions of the input features, supervised through the saliency losses, so no target result is encoded by construction. The MCF module (Eqs. 8-9) forms an attention-weighted sum of three modal streams with softmax gating weights w_m from a learnable parameter θ; these weights are fitted to the training loss, which is standard supervised learning rather than fitting a quantity that is then relabeled as a prediction. The loss functions (Eqs. 10-14) are standard BCE, IoU, and cross-scale consistency losses against ground-truth saliency maps; they do not smuggle in the benchmark metrics used for evaluation. The one self-citation, [29] (Xiong & Wen, same first/corresponding authors), is used only to attribute the cross-attention operation in Eq. 6, which is fully defined by the equation itself, so the citation is not load-bearing. The Limitation section's phrase 'multimodal saliency detection benchmarks' is inaccurate because all five evaluated datasets are RGB-only and no depth/edge generation pipeline is specified; this is a reproducibility/validation gap for the claimed multimodal contribution, but it does not make the derivation circular. Since the central improvements are measured against external SOD methods on standard benchmarks, there is no fitted-parameter-renamed-as-prediction or self-citation chain forcing the result.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The ledger tracks the unstated or under-specified inputs the central claim depends on. Most network weights are ordinary trained parameters and are not listed; the listed items are hand-set hyperparameters (α, K, λ_c), learned modality weights whose values are not disclosed, the assumed availability of depth and edge modalities, and the ad hoc 'uncertainty' map.

free parameters (4)
  • DUGC distance mixing weight α = not reported
    Eq. 1 balances spatial distance vs feature similarity in graph construction; chosen by hand, not listed or ablated.
  • Graph neighbor count K = not reported
    Eq. 3 selects top-K neighbors per node; K is a hyperparameter with no stated value or sensitivity study.
  • Consistency loss weight λ_c = not reported
    Eq. 13 weights the cross-scale consistency term in L_total; no value or ablation is reported.
  • Modality gating weights θ_m = learned during training; values not reported
    Eq. 9; these are trained on DUTS, not hand-set, but the paper does not report their final values, which are needed for full reproduction and interpretation.
assumptions (3)
  • domain assumption Depth and edge modalities are available or derivable for all training and test images (Sec. 3.2, Eqs. 8-9)
    The MCF module takes three input streams, but the experimental section lists only RGB-only datasets and gives no depth or edge extraction method; if this assumption is false, the multimodal fusion module is not operating as claimed.
  • ad hoc to paper The learned 'uncertainty weight map' from MLP(X^(3)) (Eq. 5) improves saliency accuracy
    This map is called uncertainty but is just a data-dependent attention scaling; its benefit is asserted via ablation, not derived from a first-principles uncertainty model.
  • standard math Cross-attention as implemented in Eq. 6 is a valid feature fusion mechanism for SOD
    Attributed to self-citation [29]; it is standard attention, so the underlying math is accepted, but the citation is not the canonical source.
invented entities (1)
  • Pixel-level uncertainty weight map (Eq. 5)
    purpose: Multiply feature maps to emphasize uncertain regions
    An internal learned map named 'uncertainty'; it is not tied to any measured or estimated predictive uncertainty, and no external handle is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph-Based Uncertainty Modeling and Multimodal Fusion for Salient Object Detection." pith.science (2026). https://pith.science/paper/NK6JKXJR

@misc{pith2026250820415,
  author       = {Pith},
  title        = {Pith review of: Graph-Based Uncertainty Modeling and Multimodal Fusion for Salient Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NK6JKXJR}},
  note         = {Machine review of arXiv:2508.20415}
}
read the original abstract

In view of the problems that existing salient object detection (SOD) methods are prone to losing details, blurring edges, and insufficient fusion of single-modal information in complex scenes, this paper proposes a dynamic uncertainty propagation and multimodal collaborative reasoning network (DUP-MCRNet). Firstly, a dynamic uncertainty graph convolution module (DUGC) is designed to propagate uncertainty between layers through a sparse graph constructed based on spatial semantic distance, and combined with channel adaptive interaction, it effectively improves the detection accuracy of small structures and edge regions. Secondly, a multimodal collaborative fusion strategy (MCF) is proposed, which uses learnable modality gating weights to weightedly fuse the attention maps of RGB, depth, and edge features. It can dynamically adjust the importance of each modality according to different scenes, effectively suppress redundant or interfering information, and strengthen the semantic complementarity and consistency between cross-modalities, thereby improving the ability to identify salient regions under occlusion, weak texture or background interference. Finally, the detection performance at the pixel level and region level is optimized through multi-scale BCE and IoU loss, cross-scale consistency constraints, and uncertainty-guided supervision mechanisms. Extensive experiments show that DUP-MCRNet outperforms various SOD methods on most common benchmark datasets, especially in terms of edge clarity and robustness to complex backgrounds. Our code is publicly available at https://github.com/YukiBear426/DUP-MCRNet.

Figures

Figures reproduced from arXiv: 2508.20415 by the authors.

Figure 1
Figure 1. shows the overview architecture of our model. First, the input image is passed through the CNN or Transformer based backbone to extract features of different scales. Then, the features of different levels are passed in turn to the dy￾namic uncertainty graph convolution and channel adaptive interaction module we proposed to interact with information of different scales. Then, the multi￾modal collaborative fusion part… view at source ↗
Figure 2
Figure 2. Precision-recall curves of different methods on the salient object detection task. The blue dashed line represents our model, and the other lines represent the models we compare against [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Visualization of comparative results for various saliency detection models. Our method better preserves salient structures and suppresses noise from shadows and low￾saturation regions [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [1]

    Cen, J., Wu, Y., Wang, K., Li, X., Yang, J., Pei, Y., Kong, L., Liu, Z., Chen, Q.: Sad: Segment any rgbd (2023),https://arxiv.org/abs/2305.14207

  2. [2]

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

    Chen, H., Li, Y.: Progressively complementarity-aware fusion network for rgb- d salient object detection. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3051–3060 (2018).https://doi.org/10.1109/CVPR. 2018.00322

  3. [3]

    Chen, S., Tan, X., Wang, B., Hu, X.: Reverse attention for salient object detection (2019), https://arxiv.org/abs/1807.09940

  4. [4]

    IEEE Transactions on Pattern Analysis and Machine In- telligence 37(3),569–582(2015)

    Cheng,M.M.,Mitra,N.J.,Huang,X.,Torr,P.H.S.,Hu,S.M.:Globalcontrastbased salient region detection. IEEE Transactions on Pattern Analysis and Machine In- telligence 37(3),569–582(2015). https://doi.org/10.1109/TPAMI.2014.2345401

  5. [5]

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

    Feng, M., Lu, H., Ding, E.: Attentive feedback network for boundary-aware salient object detection. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1623–1632 (2019). https://doi.org/10.1109/CVPR. 2019.00172

  6. [6]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 41(4), 815–828 (Apr 2019)

    Hou, Q., Cheng, M.M., Hu, X., Borji, A., Tu, Z., Torr, P.H.S.: Deeply su- pervised salient object detection with short connections. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(4), 815–828 (Apr 2019). https: //doi.org/10.1109/tpami.2018.2815688, http://dx.doi.org/10.1109/TPAMI. 2018.2815688

  7. [7]

    In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M

    Ji, W., Li, J., Zhang, M., Piao, Y., Lu, H.: Accurate rgb-d salient object detection via collaborative learning. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision – ECCV 2020. pp. 52–69. Springer International Publishing, Cham (2020)

  8. [8]

    Xiong et al

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2017),https: //arxiv.org/abs/1412.6980 14 Y. Xiong et al

Show all 38 references
  1. [9]

    In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Li, G., Yu, Y.: Visual saliency based on multiscale deep features. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5455–5463 (2015). https://doi.org/10.1109/CVPR.2015.7299184

  2. [10]

    Li, Y., Hou, X., Koch, C., Rehg, J.M., Yuille, A.L.: The secrets of salient object segmentation (2014), https://arxiv.org/abs/1406.2807

  3. [11]

    https://doi.org/ 10.1109/CVPR.2019.00404

    Liu, J.J., Hou, Q., Cheng, M.M., Feng, J., Jiang, J.: A simple pooling-based design forreal-timesalientobjectdetection.In: 2019 IEEE/CVFConferenceonComputer VisionandPatternRecognition(CVPR).pp.3912–3921(2019). https://doi.org/ 10.1109/CVPR.2019.00404

  4. [12]

    IEEE Transactions on Image Processing29, 6438–6451 (2020)

    Liu, N., Han, J., Yang, M.H.: Picanet: Pixel-wise contextual attention learning for accurate saliency detection. IEEE Transactions on Image Processing29, 6438–6451 (2020). https://doi.org/10.1109/TIP.2020.2988568

  5. [13]

    In: 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR)

    Liu, N., Zhang, N., Han, J.: Learning selective self-mutual attention for rgb-d saliency detection. In: 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR). pp. 13753–13762 (2020). https://doi.org/10.1109/ CVPR42600.2020.01377

  6. [14]

    Liu, N., Zhang, N., Wan, K., Shao, L., Han, J.: Visual saliency transformer (2021), https://arxiv.org/abs/2104.12099

  7. [15]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV)

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin trans- former: Hierarchical vision transformer using shifted windows. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9992–10002 (2021). https://doi.org/10.1109/ICCV48922.2021.00986

  8. [16]

    IEEE Transac- tions on Circuits and Systems for Video Technology 35(2), 1041–1054 (2025)

    Mao, Y., Zhang, J., Wan, Z., Tian, X., Li, A., Lv, Y., Dai, Y.: Generative transformer for accurate and reliable salient object detection. IEEE Transac- tions on Circuits and Systems for Video Technology 35(2), 1041–1054 (2025). https://doi.org/10.1109/TCSVT.2024.3469286

  9. [17]

    In: 2010 IEEE Computer Society Confer- ence on Computer Vision and Pattern Recognition - Workshops

    Movahedi, V., Elder, J.H.: Design and perceptual validation of performance mea- sures for salient object segmentation. In: 2010 IEEE Computer Society Confer- ence on Computer Vision and Pattern Recognition - Workshops. pp. 49–56 (2010). https://doi.org/10.1109/CVPRW.2010.5543739

  10. [18]

    Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection (2020),https://arxiv.org/abs/2007.09062

  11. [19]

    In: 2019 IEEE/CVF International Con- ference on Computer Vision (ICCV)

    Piao, Y., Ji, W., Li, J., Zhang, M., Lu, H.: Depth-induced multi-scale recurrent attention network for saliency detection. In: 2019 IEEE/CVF International Con- ference on Computer Vision (ICCV). pp. 7253–7262 (2019).https://doi.org/10. 1109/ICCV.2019.00735

  12. [20]

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

    Qin, X., Zhang, Z., Huang, C., Gao, C., Dehghan, M., Jagersand, M.: Bas- net: Boundary-aware salient object detection. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7471–7481 (2019). https://doi.org/10.1109/CVPR.2019.00766

  13. [21]

    IEEE Transactions on Image Processing26(5), 2274–2285 (May 2017).https://doi.org/10.1109/tip.2017.2682981, http://dx.doi.org/ 10.1109/TIP.2017.2682981

    Qu, L., He, S., Zhang, J., Tian, J., Tang, Y., Yang, Q.: Rgbd salient object de- tection via deep fusion. IEEE Transactions on Image Processing26(5), 2274–2285 (May 2017).https://doi.org/10.1109/tip.2017.2682981, http://dx.doi.org/ 10.1109/TIP.2017.2682981

  14. [22]

    Sun, G., Wang, W., Dai, J., Gool, L.V.: Mining cross-image semantics for weakly supervised semantic segmentation (2020),https://arxiv.org/abs/2007.01947

  15. [23]

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

    Wang, L., Lu, H., Wang, Y., Feng, M., Wang, D., Yin, B., Ruan, X.: Learning to detect salient objects with image-level supervision. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3796–3805 (2017).https: //doi.org/10.1109/CVPR.2017.404 Uncertain...

  16. [24]

    Wang, W., Lai, Q., Fu, H., Shen, J., Ling, H., Yang, R.: Salient object detection in the deep learning era: An in-depth survey (2021),https://arxiv.org/abs/1904. 09146

  17. [25]

    Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions (2021), https://arxiv.org/abs/2102.12122

  18. [26]

    Wei, J., Wang, S., Huang, Q.: F3net: Fusion, feedback and focus for salient object detection (2019), https://arxiv.org/abs/1911.11445

  19. [27]

    Wei, J., Wang, S., Wu, Z., Su, C., Huang, Q., Tian, Q.: Label decoupling framework for salient object detection (2020),https://arxiv.org/abs/2008.11048

  20. [28]

    Wu, Z., Su, L., Huang, Q.: Cascaded partial decoder for fast and accurate salient object detection (2019),https://arxiv.org/abs/1904.08739

  21. [29]

    Xiong, Y., Wen, Y.: Non-stationary time series forecasting based on fourier analysis and cross attention mechanism (2025),https://arxiv.org/abs/2505.06917

  22. [30]

    In: 2013 IEEE Conference on Computer Vision and Pattern Recognition

    Yan, Q., Xu, L., Shi, J., Jia, J.: Hierarchical saliency detection. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1155–1162 (2013). https://doi.org/10.1109/CVPR.2013.153

  23. [31]

    In: 2013 IEEE Conference on Computer Vision and Pat- tern Recognition

    Yang, C., Zhang, L., Lu, H., Ruan, X., Yang, M.H.: Saliency detection via graph- based manifold ranking. In: 2013 IEEE Conference on Computer Vision and Pat- tern Recognition. pp. 3166–3173 (2013). https://doi.org/10.1109/CVPR.2013. 407

  24. [32]

    In: Ide, I., Kompatsiaris, I., Xu, C., Yanai, K., Chu, W.T., Nitta, N., Riegler, M., Yamasaki, T

    Yu, J., Liu, Y., Wu, X., Xu, K., Li, J.: Pa2net: Pyramid attention aggregation network for saliency detection. In: Ide, I., Kompatsiaris, I., Xu, C., Yanai, K., Chu, W.T., Nitta, N., Riegler, M., Yamasaki, T. (eds.) MultiMedia Modeling. pp. 186–200. Springer Nature Singapore, ...

  25. [33]

    IEEE Transactions on Image Processing34, 2301–2314 (2025)

    Yuan, Y., Gao, P., Dai, Q., Qin, J., Xiang, W.: Uncertainty-guided refinement for fine-grained salient object detection. IEEE Transactions on Image Processing34, 2301–2314 (2025). https://doi.org/10.1109/TIP.2025.3557562

  26. [34]

    IEEE Transactions on Image Processing 30, 8727–8742 (2021).https://doi.org/10.1109/tip.2021.3116793, http://dx.doi.org/10.1109/TIP.2021.3116793

    Zhai, Y., Fan, D.P., Yang, J., Borji, A., Shao, L., Han, J., Wang, L.: Bifurcated backbone strategy for rgb-d salient object detection. IEEE Transactions on Image Processing 30, 8727–8742 (2021).https://doi.org/10.1109/tip.2021.3116793, http://dx.doi.org/10.1109/TIP.2021.3116793

  27. [35]

    In: 2017 IEEE International Conference on Computer Vision (ICCV)

    Zhang, P., Wang, D., Lu, H., Wang, H., Ruan, X.: Amulet: Aggregating multi-level convolutional features for salient object detection. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 202–211 (2017).https://doi.org/ 10.1109/ICCV.2017.31

  28. [36]

    In: 2019 IEEE/CVF International Con- ference on Computer Vision (ICCV)

    Zhao, J., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: Egnet: Edge guid- ance network for salient object detection. In: 2019 IEEE/CVF International Con- ference on Computer Vision (ICCV). pp. 8778–8787 (2019).https://doi.org/10. 1109/ICCV.2019.00887

  29. [37]

    Computational Visual Media 7(1), 37–69 (Mar 2021)

    Zhou, T., Fan, D.P., Cheng, M.M., Shen, J., Shao, L.: Rgb-d salient ob- ject detection: A survey. Computational Visual Media 7(1), 37–69 (Mar 2021). https://doi.org/10.1007/s41095-020-0199-z, http://dx.doi.org/10. 1007/s41095-020-0199-z

  30. [38]

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

    Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12734–12745 (2023).https://doi. org/10.1109/ICCV51070.2023.01174

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.