{"id":"6672467b-845c-4e1e-b125-f120966c38e6","arxiv_id":"2501.01648","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GL-DMNet combines RGB and depth features through position and channel mutual fusion modules and a transformer-infused decoder, reporting state-of-the-art averages on six RGB-D salient object detection benchmarks.","lead":"A new neural network, GL-DMNet, fuses color and depth information with spatial and channel attention plus a transformer decoder to find salient objects in images. It reports top average scores on six RGB-D benchmarks, though by small margins and with some metrics still led by prior methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not yet controlled: the ~3% margin is mostly small E/S/F differences with no error bars, and S3Net/FCFNet use different training protocols; a matched multi-seed rerun is needed.","rationale":"The architecture is described in enough detail to be reproducible, the linked code and ablations support the method's internal consistency, and the datasets and metrics are standard. The load-bearing issue is the quantitative comparative claim itself: the headline margin over the second-best method is tiny on most cells and is not accompanied by error bars or multiple runs. The paper's own discussion concedes that S3Net uses semi-supervised training with supplementary priors and FCFNet uses a depth-quality filtering strategy, so the baseline numbers in Table II are not generated under a shared protocol. The '~3% average improvement' is also sensitive to how the average is computed: taking signed relative differences over the 20 available S3Net cells gives about 2.6%, while using absolute values gives about 3.0%, and the largest contributor is MAE, not the structural metrics. Because the strongest claim is precisely that GL-DMNet outperforms 24 methods, this uncontrolled comparison and the absence of significance testing leave the SOTA assertion conditional. The reader's CONDITIONAL verdict is therefore appropriate, and no change to that verdict is needed.","tokens_in":21433,"tokens_out":11246,"duration_ms":112444,"concrete_test":"Retrain GL-DMNet, S3Net, and FCFNet from their released code under a matched protocol (same NLPR/DUT-RGBD/NJUD fully-supervised train split, same input size 256×256, same augmentation, no extra unlabeled or weak supervision) with at least five seeds, and report mean±std for all 24 metric-dataset cells. If the 95% confidence intervals overlap on STEREO Sα, NLPR Fβ/MAE, or DUT-RGBD Eξ, or if the signed average relative improvement falls below ~2%, the abstract's 'better than 24 methods / ~3% improvement' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table II shows that on the 20 metric-dataset cells where S3Net has numbers, GL-DMNet's Eξ/Sα/Fβ advantages are typically ≤0.003, while S3Net actually wins STEREO Sα (0.913 vs 0.908) and NLPR MAE/Fβ (0.021 vs 0.022; 0.923 vs 0.926), and matches NLPR Eξ/Sα exactly. The abstract's '~3% average improvement' therefore depends heavily on relative MAE gains (e.g., 19.6% on SIP) and on averaging absolute signed differences; it is not a robust head-to-head margin. No error bars or multi-seed results are reported, and Sec. IV-C acknowledges S3Net is semi-supervised with supplementary priors while FCFNet discards low-quality depth, so the comparison is not controlled. If S3Net and FCFNet were retrained under the same fully-supervised split, the direction and size of the margin could change, undermining the central SOTA claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GL-DMNet, an RGB-D salient object detection network combining ResNet-50 backbone branches for RGB and depth, a dual mutual fusion module with position mutual fusion (PMF) and channel mutual fusion (CMF), and a cascade transformer-infused reconstruction (CTR) decoder built on PVTv2-B2. The model is trained with deep supervision using a weighted BCE plus IoU loss. Experiments on six RGB-D SOD benchmark datasets compare against 24 published methods under four metrics, with additional PR curves, ablations of the fusion module, decoder, and loss, plus qualitative visualizations. The central claim is that GL-DMNet outperforms all 24 methods and achieves an average improvement of about 3% over the second-best method, S3Net.","tokens_in":21619,"tokens_out":6115,"duration_ms":59260,"significance":"If the central claim is supported, GL-DMNet would be a new state of the art on six commonly used RGB-D SOD benchmarks, with the PMF/CMF modules and CTR decoder as the enabling components. The paper has clear strengths: the architecture is described in detail with equations, the ablation study isolates the contribution of each major module, the training setup is reproducible on a single GPU, and the authors promise to release code and results. However, the headline SOTA claim is not yet adequately controlled: baseline numbers are taken from published saliency maps under heterogeneous training protocols, and the reported margins over the closest competitor are often at the level of 0.001-0.003 with no error bars or multi-seed statistics. The central significance therefore depends on verification under matched experimental conditions.","major_comments":[{"comment":"The central claim that GL-DMNet 'performs better than 24 RGB-D SOD methods' and improves on the second-best model by ~3% is not based on a controlled comparison. Section IV-C states that baseline metrics are computed 'based on the saliency maps derived from their original papers,' but the baselines include methods trained under substantially different protocols: S3Net is semi-supervised with supplementary priors, FCFNet discards low-quality depth maps, DWD is trained without depth, DENet and DHFR are weakly supervised, and DLMNet is unsupervised. These methods are not retrained under the fully supervised protocol of Section IV-B, so the reported ranking may reflect protocol differences rather than architectural superiority. The magnitude of the claimed gain is also fragile: in Table II many GL-DMNet advantages over S3Net are 0.001-0.003 in E_xi, S_alpha, and F_beta, while S3Net actually wins STEREO S_alpha (0.913 vs 0.908) and NLPR MAE (0.021 vs 0.022), and ties on NLPR E_xi and S_alpha. The abstract's ~3% figure is driven largely by large relative MAE differences on SIP and DUT-RGBD and by averaging across datasets and metrics. The authors should either retrain all baselines under a matched protocol or explicitly restrict the claim to 'outperforms published saliency maps' and remove the 'For fair comparisons' wording in Section IV-C.","section":"Sec. IV-C / Abstract"},{"comment":"The PMF module as written does not implement position mutual fusion. After reshaping f_RGB_i, f_D_i, and f_SP_i to R^{C x N}, Equation (6) computes M(f_SP_i \\otimes (f_RGB_i)^T), which is a C x C matrix, not an N x N spatial attention map over positions. Equation (9) then multiplies the C x N feature f_RGB_i by this C x C matrix, so the operation mixes channels rather than attending to spatial positions. To obtain a genuine spatial attention map, the multiplication order should be reversed, e.g., (f_RGB_i)^T \\otimes f_SP_i yielding an N x N map, with corresponding adjustments to Eqs. (9)-(11). As it stands, the PMF is formally analogous to the channel interaction in Eqs. (15)-(17), contradicting the claimed spatial/channel decomposition and the qualitative interpretation in Section IV-E.","section":"Sec. III-B, Eqs. (6)-(11)"},{"comment":"All quantitative results are single-run numbers with no error bars, no standard deviations, and no significance testing. Given that the closest competitor comparisons are frequently within 0.003 in E_xi, S_alpha, and F_beta, and that the paper's own training procedure involves a multi-phase freeze/unfreeze schedule with random augmentation, the reported differences are not shown to be statistically distinguishable from run-to-run variation. The authors should report mean and standard deviation over at least three independent training runs for their method, and ideally also for the closest baselines under a matched training protocol.","section":"Sec. IV-C / Table II"}],"minor_comments":[{"comment":"The abstract and the conclusion say four evaluation metrics, while Section IV-A lists five metrics and also presents PR curves; please clarify the count.","section":"Sec. IV-A"},{"comment":"In the CDINet row for NJUD, F_beta is reported as 0.827, which is inconsistent with the surrounding values and is likely a typo; the intended value is probably 0.927.","section":"Table II"},{"comment":"The mapping from the loss weights {0.8, 0.6, 0.4, 0.2} to decoder stages i=1..4 is not stated; please specify which weight corresponds to which level of deep supervision.","section":"Eq. (33)"},{"comment":"The moment normalization M(x) = sign(x) * x^{-1/2} is not standard and is not defined for matrix-valued inputs with negative or zero entries; please define the intended element-wise operation and explain its purpose.","section":"Eqs. (6)-(7)"},{"comment":"The notation Conv3(Conv1(...)) in Eqs. (1)-(2) and elsewhere is not defined in Table I; adding a convention for sequential composition would improve readability.","section":"Sec. III-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for an IEEE Transactions on Multimedia submission, and the proposed architecture is clearly presented. The main risk is the uncontrolled baseline comparison, which is the basis of the headline SOTA claim; a matched-protocol evaluation or a substantially softened claim is needed. The shape inconsistency in Eqs. (6)-(11) also needs to be resolved before the PMF contribution can be evaluated. I did not verify the GitHub repository, so the code availability statement should be treated as unverified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, workmanlike RGB-D SOD paper with a coherent architecture, honest ablations, and released code. The novelty is incremental—PMF/CMF are mutual-attention variants of DANet, and the decoder is a PVTv2 plus reconstruction scheme inspired by TriTransNet—but the paper says so. The problem is the abstract overclaims. It says GL-DMNet 'performs better than 24 methods' with ~3% average improvement, but Table II shows S3Net wins STEREO S-alpha and NLPR F-beta/MAE, and FCFNet wins NJUD and STEREO E-xi. The margin on many cells is under 0.003, which without error bars or multi-seed results is not a meaningful SOTA edge. The ~3% figure is carried by relative MAE gains like 19.6% on SIP, which is a different kind of comparison.\n\nThe other real issue is protocol mismatch. S3Net is semi-supervised with extra priors, FCFNet discards low-quality depth, and the paper doesn't retrain baselines under a matched setting. That's standard practice in SOD—reusing official maps—but it means the head-to-head SOTA claim is not controlled. The authors'd need a matched multi-seed rerun of the top baselines to justify the abstract. They also claim efficiency advantages over TriTransNet/GroupTransNet but give no parameter/FLOPs table; that's an easy fix.\n\nWhat the paper does well: equations are clean, training details are complete, ablations cover the fusion module, transformer choice, reconstruction, and loss. The dual mutual fusion design shows consistent improvement in Table III. The qualitative feature visualizations are useful. This is a reproducible, honest piece of engineering.\n\nBottom line: the method is sound and the direction of results is believable, but the central claim needs tempering and the comparisons need tightening. I'd send it to review, asking for a revised abstract, an efficiency table, and ideally some significance testing or at least multi-seed numbers for the closest baselines. It's not a breakthrough, but it's a useful contribution to a narrow subfield.","headline":"Solid, reproducible RGB-D SOD paper that overclaims its SOTA margin; the architecture and ablations hold up, but the abstract needs tempering and the baseline comparisons need control.","tokens_in":22226,"tokens_out":2185,"would_cite":false,"duration_ms":21611,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Bidirectional attention fusion tops 24 RGB-D saliency models.","keywords":["RGB-D salient object detection","dual attention mechanism","cross-modal fusion","position mutual fusion","channel mutual fusion","cascade transformer-infused reconstruction","PVTv2","saliency benchmarks"],"falsifier":"Retrain the leading baselines, including S3Net and FCFNet, with the same training split, the same loss, and no extra priors or depth filtering, then recompute E-measure, S-measure, F-measure, and MAE with the same official evaluation tool; if GL-DMNet's average margin over S3Net falls well below 3%, the reported state-of-the-art claim would not survive.","tokens_in":1588,"feed_emoji":"🎯","tokens_out":2072,"duration_ms":73047,"temperature":0.7,"pith_summary":"This paper tries to establish that RGB and depth should refine each other mutually during fusion rather than being combined by a fixed, one-way rule. The proposed GL-DMNet fuses the two modalities through a position mutual fusion module and a channel mutual fusion module, then decodes the fused features with a cascade transformer-infused reconstruction network. On six RGB-D salient object detection benchmarks, the paper reports that GL-DMNet outperforms 24 prior methods, with an average gain of about 3% across four metrics over the second-best model, S3Net. If that comparison holds, the contribution is a concrete architecture for cross-modal fusion that uses bidirectionally exchanged attention and stage-wise global context.","feed_headline":"Mutual attention fusion tops 24 RGB-D saliency models","feed_subtitle":"On six benchmarks the model averages roughly 3 percent better than the runner-up, S3Net.","key_machinery":"The load-bearing mechanism is the dual mutual fusion module, consisting of a position mutual fusion (PMF) module and a channel mutual fusion (CMF) module. PMF computes per-position similarity between the fused feature map and each modality's own map by matrix multiplication, so spatial attention weights are exchanged in both directions; CMF does the same for channel attention. The second mechanism is the cascade transformer-infused reconstruction (CTR) decoder, which feeds each stage's fused features through a separate stage of a PVTv2 transformer and reconstructs the saliency map by letting high-level features guide low-level ones through upsampling, channel attention, and residual connections. Together these modules let local features gather global context while keeping each stage's features distinct, since the transformer stages do not share parameters.","core_discovery":"The central claim is that mutual attention exchange between RGB, depth, and their fusion, in both spatial and channel dimensions, yields better saliency maps than unilateral or manual fusion. The paper's architecture encodes RGB and depth with ResNet-50 branches, runs the features through parallel position mutual fusion and channel mutual fusion modules, and reconstructs the saliency map with a cascade transformer-infused decoder based on PVTv2. The quantitative claim is that this design reaches best or second-best performance on nearly every metric of the six benchmarks, improving on S3Net by about 3% on average. Ablations trace the gain to the dual mutual fusion module, the transformer embedding, the reconstruction decoder, and the combination of BCE and IoU losses. The paper also claims robustness when depth is missing or low-quality, though it does not formalize a separate test for that case.","pith_inferences":["Because baseline scores are computed from published saliency maps rather than from retrained models, the roughly 3% average margin could shrink or disappear under a fully controlled re-training with identical supervision and data splits; this is a testable risk, not a claim the paper makes.","The mutual-attention fusion recipe of PMF plus CMF is modality-agnostic in form, so it could plausibly transfer to other paired-input tasks such as RGB-thermal detection or RGB-depth semantic segmentation, which face the same discrepancy problem.","A natural follow-up experiment would train GL-DMNet on RGB-only inputs, either by removing the depth branch or feeding it a constant, to quantify how much of the gain comes from depth information versus from the mutual fusion and decoder alone.","The paper's qualitative robustness to missing depth could be turned into a quantitative test by evaluating GL-DMNet on RGB-D benchmarks with depth maps artificially corrupted or set to zero, measuring the drop in each metric."],"forward_implications":["If GL-DMNet's reported numbers hold, effective RGB-D fusion can be built by exchanging spatial and channel attention in both directions among RGB, depth, and their fusion, rather than by choosing a single fusion rule.","The stage-wise, non-shared transformer decoder provides a template for adding global context to multi-level fusion without repeatedly running a large transformer, which the paper connects to its reported 40 FPS inference speed.","On the six standard benchmarks, downstream applications that consume saliency maps would receive foreground masks that are on average about 3% closer to the ground truth than those of the runner-up among 24 compared methods.","Deep supervision with per-stage BCE and IoU losses is shown to speed convergence and improve the final single saliency output, suggesting that auxiliary multi-level outputs are worth retaining in similar decoding designs.","The paper's own tables show that FCFNet and S3Net still beat GL-DMNet on a few metric-dataset cells, so the claimed advantage is an average margin concentrated on some datasets rather than a universal win on every measurement."],"supporting_citations":[{"why":"Provides the dual attention (position and channel) mechanism on which the PMF and CMF modules are built.","marker":"[7]"},{"why":"The triplet transformer embedding network that motivates the joint transformer-based encoder and serves as a key comparison.","marker":"[10]"},{"why":"Supplies the spatial and channel attention blocks used inside the mutual fusion modules.","marker":"[40]"},{"why":"ResNet-50 backbone that extracts the multi-level RGB and depth features.","marker":"[50]"},{"why":"PVTv2 transformer backbone used as the stage-wise encoder in the cascade transformer-infused reconstruction decoder.","marker":"[55]"},{"why":"Defines the six benchmark datasets and the official evaluation tools used for all reported metrics.","marker":"[1]"},{"why":"S3Net is the second-best model against which the roughly 3% average improvement is measured.","marker":"[77]"},{"why":"FCFNet is the strongest competitor on NJUD and STEREO E-measure, using a strategy of discarding low-quality depth maps.","marker":"[24]"},{"why":"Squeeze-and-excitation channel attention used in the multi-level feature reconstruction.","marker":"[56]"}],"fun_headline_variants":["Mutual attention fusion beats 24 RGB-D saliency models","Mutual learning lifts RGB-D saliency ~3% on six benchmarks","Global-local mutual attention fuses RGB and depth better","Spatial and channel mutual fusion improve RGB-D saliency"],"cache_read_input_tokens":24320,"weakest_assumption_plain":"The roughly 3% claim assumes the 24 comparison methods were evaluated under matched training and evaluation protocols, even though several of them were trained with extra supervision, discarded low-quality depth maps, or used different data splits.","fun_headline_variants_meta":{"raw":{"variants":["Mutual attention fusion beats 24 RGB-D saliency models","Mutual learning lifts RGB-D saliency ~3% on six benchmarks","Global-local mutual attention fuses RGB and depth better","Spatial and channel mutual fusion improve RGB-D saliency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00091,"raw_usage":{"total_tokens":3916,"prompt_tokens":957,"completion_tokens":2959,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2888}},"tokens_in":573,"tokens_out":2959,"duration_ms":23054,"temperature":1.0,"reasoning_tokens":2888,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:22:44.693697+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the leading baselines, including S3Net and FCFNet, with the same training split, the same loss, and no extra priors or depth filtering, then recompute E-measure, S-measure, F-measure, and MAE with the same official evaluation tool; if GL-DMNet's average margin over S3Net falls well below 3%, the reported state-of-the-art claim would not survive.","supporting_citations":[{"cited_title":"Dual attention network for scene segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the dual attention (position and channel) mechanism on which the PMF and CMF modules are built."},{"cited_title":"Tritransnet: Rgb-d salient object detection with a triplet transformer embedding network,","cited_arxiv_id":null,"evidence_quote":"The triplet transformer embedding network that motivates the joint transformer-based encoder and serves as a key comparison."},{"cited_title":"Cbam: Convolutional block attention module,","cited_arxiv_id":null,"evidence_quote":"Supplies the spatial and channel attention blocks used inside the mutual fusion modules."},{"cited_title":"S 3 net: Self-supervised self-ensembling network for semi-supervised rgb-d salient object detection,","cited_arxiv_id":null,"evidence_quote":"S3Net is the second-best model against which the roughly 3% average improvement is measured."},{"cited_title":"Feature calibrating and fusing network for rgb-d salient object detection,","cited_arxiv_id":null,"evidence_quote":"FCFNet is the strongest competitor on NJUD and STEREO E-measure, using a strategy of discarding low-quality depth maps."}],"review_version":1}