{"id":"8e49b63c-b23f-44fb-bb5b-87f1821dbc15","arxiv_id":"2505.24002","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A depth-guided cross-attention network with a transformer-CNN bridge improves no-reference image quality prediction and cross-dataset generalization over several published baselines.","lead":"DGIQA is a no-reference image quality model that adds an automatically estimated depth map as a second input, using cross-attention to focus on salient objects and a transformer-CNN bridge to fuse global and local features. It reports top results on several IQA benchmarks and larger quality-score separation on unseen low-light, hazy, and lens-flare images than two strong baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-CAR ablation is confounded by parameter count: removing depth also removes the second Swin-T backbone (65M vs 103M params), so the cross-dataset gains in Table 5 may be capacity effects rather than depth guidance.","rationale":"Good-faith reading: the paper's headline empirical claims may well be true as measured. Table 2 shows DGIQA is competitive with MANIQA and others, and the cross-dataset numbers in Table 3 are plausible. The concern is not about intentional misreporting; it is about what the ablation can prove. Because the Depth-CAR module is coupled with the second pre-trained backbone, the controlled experiment required to attribute improvement to depth is missing. The reader's weakest_assumption (depth maps on distorted inputs preserve structure) is also relevant, but it is a secondary reliability question; the primary causal question is whether depth information, not extra parameters, drives the reported gains. Both could be settled by the same capacity-matched control using ground-truth or predicted depth and a non-depth auxiliary stream. This does not change the overall CONDITIONAL verdict: the empirical claims need additional controlled evidence and variance reporting, but there is no basis for rejection. Therefore verdict stays unchanged.","tokens_in":18450,"tokens_out":6772,"duration_ms":70479,"concrete_test":"Implement a capacity-matched control: replace the depth input of the second Swin-T backbone with the same RGB image (or a fixed random projection) while keeping Depth-CAR, TCB, dilation, loss, and training schedule identical. Retrain on Koniq10k and evaluate on LIVE-C and LIVE-FB, matching Table 5. If the control reproduces the +3.2%/+8.2% SROCC gains over config #1, the depth-specific benefit is not established. If the control does not reproduce the gains, the depth guidance claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is the Depth-CAR mechanism, which is claimed to make NR-IQA generalize by distilling depth into quality features. The only direct evidence for this is the ablation in Tables 4-5. In Table 5, config #1 ('w/o D-CAR') is compared against config #4 (full model) to attribute gains of +3.2% SROCC on LIVE-C and +8.2% on LIVE-FB. But config #1 has 65M parameters while config #4 has 103M: removing Depth-CAR also removes the entire second Swin-T backbone that consumes the depth map. The comparison therefore conflates 'depth semantics' with 'added trainable capacity'. A model with a second RGB stream (or any auxiliary stream) might show the same improvement without any depth information. This is a load-bearing gap because the paper's novelty claim in Sec. 3.3 and the abstract rests on depth being the cause of the gains. It is distinct from the depth-reliability limitation the authors acknowledge in Sec. 5.2: even if DepthAnything preserves structure, the current experiments cannot establish that the preserved structure is what is being used.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DGIQA, a no-reference image quality assessment model that combines two Swin Transformer backbones (one for RGB, one for depth), Transformer-CNN Bridge (TCB) blocks, a depth-guided cross-attention and refinement (Depth-CAR) module, and a dilated convolution stack. The model is trained with MSE plus a consistency loss, and is evaluated on seven benchmark datasets, cross-dataset transfer, and five natural distortion datasets (low light, haze, flare, motion blur). The authors claim state-of-the-art performance on synthetic and authentic benchmarks, superior cross-dataset generalization, and introduce a 'density separation' criterion to quantify generalization on unseen distortions. The paper includes ablations, Grad-CAM visualizations, t-SNE plots, and an FR-IQA adaptation.","tokens_in":18730,"tokens_out":6005,"duration_ms":49798,"significance":"If the central claims were properly supported, depth-guided attention would be a useful and reasonably novel contribution to NR-IQA, and the TCB block is a sensible parameter-reduction idea. The paper reports a 10-split evaluation protocol and makes model and code publicly available, which are strengths. However, the SOTA claim is weakened by Table 2, where DGIQA is not the best method on several datasets, and the key ablation in Tables 4-5 is confounded because removing Depth-CAR also removes the second Swin-T backbone (65M vs 103M parameters). The proposed density-separation criterion is not formally defined and is not validated against human opinion or established generalization metrics. The contribution is therefore plausible but not yet established.","major_comments":[{"comment":"The SOTA claim in the Abstract ('achieves state-of-the-art (SOTA) performance on both synthetic and authentic benchmark datasets') is not supported by Table 2. On TID2013, MANIQA reports SROCC/PLCC of 0.937/0.943 while DGIQA reports 0.934/0.940; on LIVE-FB, TOPIQ reports 0.652/0.745 and Re-IQA 0.645/0.733 versus DGIQA 0.591/0.685; and on LIVE-C, LIQE reports 0.904/0.910 versus DGIQA 0.891/0.910. The claim should be revised to 'competitive' or 'top-two', and the comparisons need variance information.","section":"Abstract; Table 2"},{"comment":"The ablation that attributes the cross-dataset gains to Depth-CAR is confounded by model capacity. In Table 5, configuration #1 (DGIQA w/o D-CAR) has 65M parameters and configuration #4 (full DGIQA) has 103M parameters; removing Depth-CAR also removes the entire second Swin-T backbone that consumes the depth map. The reported improvements of +3.2% SROCC on LIVE-C and +8.2% on LIVE-FB could therefore be due to increased trainable capacity rather than to depth-guided attention. A controlled comparison, for example a second RGB stream with the same capacity or a frozen depth backbone, is necessary to support the central novelty claim.","section":"Sec. 4.4, Tables 4-5"},{"comment":"The proposed 'density separation' criterion is not defined formally, and it is used as the main evidence for generalization on natural distortions. The paper does not specify how the Gaussian overlap is computed, how the 41-50% / 20-21% / 5-13% / 65-95% / 3-60% figures are derived, or whether these differences are statistically significant across the 10 training splits. Since this criterion is introduced by the authors and is not validated against human opinion scores or an established metric, it should be presented as an auxiliary diagnostic rather than as standalone evidence of SOTA generalization.","section":"Sec. 5.2"},{"comment":"The depth maps are generated by DepthAnything on distorted inputs, and Appendix B shows that severe white noise degrades the structural information in these maps. The paper acknowledges this limitation in Sec. 5.2, but there is no quantitative analysis of how often DepthAnything fails or how much such failures affect DGIQA's predictions. To support the claim that depth guidance is robust, the authors should report performance under corrupted or noisy depth inputs, or compare against a model trained on clean depth maps.","section":"Sec. 4.2; Appendix B"},{"comment":"The cross-dataset claim is stated as outperforming SOTA 'in most dataset pairs', but Table 3(a) shows that on LIVE-C, LoDa achieves SROCC 0.811 versus DGIQA 0.808, and DGIQA does not consistently beat all baselines across pairs. Furthermore, no standard deviations or significance tests are reported for any of the 10-split means, so small differences such as 0.808 vs 0.811 are not interpretable. Please provide per-split statistics or error bars, and qualify the statement accordingly.","section":"Sec. 4.3, Table 3"}],"minor_comments":[{"comment":"There are typos: 'seimesi architecture' should be 'Siamese architecture' in Sec. 2, and 'albumentation library' should be 'Albumentations library' in Sec. 4.2.","section":"Sec. 2; Sec. 4.2"},{"comment":"Equation (7) writes 'LSME' instead of 'LMSE' for the mean-squared-error loss term.","section":"Appendix A.2"},{"comment":"The checkmarks in Table 4 are not legible in the typeset version; the four configurations should be explicitly named in the caption or in the table.","section":"Sec. 4.4, Table 4"},{"comment":"The notation for dilated convolutions is inconsistent: the text says 'dilation rates of 2 and 4', but Eq. (4) writes 'dltn=2,4' without a clear definition of the symbol 'dltn'.","section":"Sec. 3.1; Eq. (4)"},{"comment":"The claim that TCB blocks 'reduce the model parameters by 23.3%' is not directly verifiable from Tables 4-5; the parameter count drops from 127M (w/o TCB) to 103M (full), which is about 18.9%, so please clarify the reference configuration for this percentage.","section":"Sec. 4.4"},{"comment":"The phrase 'a new evaluation criteria' should be 'a new evaluation criterion'.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the ablation confound: the depth contribution is conflated with added parameter capacity, and this directly affects the paper's central claim. If the authors add a capacity-matched RGB-stream control and temper the SOTA claim, the paper could be a reasonable contribution. The density-separation criterion needs formalization and validation before it is used as evidence. I see no grounds for outright rejection if these fixes are made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of the DGIQA paper. The model is a competent NR-IQA pipeline: two Swin-T backbones (RGB and depth), SE-style transformer-CNN bridges, a depth-query cross-attention block, and dilated convs. It's evaluated on seven standard benchmarks plus low-light/haze/flare/blur datasets, and the authors release code. That is solid work, and the engineering is careful.\n\nThe core contribution is the Depth-CAR mechanism, claimed to improve generalization by distilling depth. The ablation evidence for this is weak. In Table 5, the 'w/o D-CAR' config removes not just the cross-attention but the entire depth branch, dropping parameters from 103M to 65M. So the +3.2%/+8.2% SROCC gains on LIVE-C and LIVE-FB could simply be extra trainable capacity. A control with a second RGB stream of similar size is missing. The Grad-CAM visualizations are suggestive but qualitative, and the paper does not compare against earlier depth-guided IQA methods, so the novelty over that line is unclear.\n\nSecond, the abstract overstates the benchmark results. In Table 2, MANIQA ties or beats DGIQA on LIVE and TID2013, LIQE beats it on LIVE-C, and TOPIQ beats it on LIVE-FB. 'State-of-the-art on both synthetic and authentic benchmarks' is not what the table shows. Third, no standard deviations are reported despite 10-split evaluation, and hyperparameters (lambda, crop count) are tuned on validation sets. The new 'density separation' criterion is clever, but since it is defined by the authors, the strong overlap-reduction numbers need external validation or standard metrics to back them up.\n\nOn the positive side, the depth maps on severely distorted images are shown to degrade (Appendix B), and the authors honestly list limitations where depth guidance fails. The t-SNE and visualizations add some evidence of structure.\n\nBottom line: this is a serious, reproducible piece of work that deserves peer review, but the main generalization claim needs a properly controlled ablation. If the authors add an equal-capacity RGB-only baseline, report variances, and tone down the abstract, it could be a useful contribution. Recommend: send to peer review with major revision.","headline":"Depth-guided IQA with a promising architecture, yet the key ablation is confounded by parameter count and the abstract oversells the results.","tokens_in":19246,"tokens_out":3554,"would_cite":false,"duration_ms":33833,"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":"Adding depth-guided cross-attention to a no-reference image quality model makes its quality scores generalize to unseen distortions such as low light, haze, and lens flares.","keywords":["image quality assessment","no-reference IQA","depth-guided cross-attention","Transformer-CNN Bridge","generalization","density separation","low-light distortion","lens flare"],"falsifier":"Retrain DGIQA with the estimated depth maps replaced by random noise or by depth from a deliberately corrupted estimator; if cross-dataset SROCC and PLCC on the unseen natural distortion sets stay the same or improve, then depth content is not what drives the reported gains, and the depth-guidance claim collapses.","tokens_in":18261,"feed_emoji":"🖼️","tokens_out":6519,"duration_ms":57153,"temperature":0.7,"pith_summary":"No-reference image quality assessment (NR-IQA) aims to predict how humans judge an image without access to an original, but models trained on known distortion types often fail on unseen natural distortions. This paper claims that adding scene depth as a guiding signal fixes much of that gap: a depth-guided cross-attention mechanism, Depth-CAR, makes the model attend to salient foreground objects and structural boundaries, which are the regions humans weight when judging quality. A companion Transformer-CNN Bridge (TCB) fuses global and local features while cutting parameters by about 23%. Tested across seven synthetic and authentic benchmarks plus unseen low-light, hazy, lens-flare, and motion-blur datasets, the resulting model, DGIQA, is claimed to reach state-of-the-art correlation with human scores and to separate high-quality from distorted images far more cleanly than existing models.","feed_headline":"Depth cues sharpen blind image quality scores on unseen distortions","feed_subtitle":"DGIQA's depth-guided attention beats top models on cross-dataset tests, including low light, haze, and flares.","key_machinery":"The load-bearing component is the Depth-CAR block, a depth-guided cross-attention and refinement mechanism: features from the depth stream provide the queries, while RGB features supply keys and values, so the attention weights emphasize close, salient objects and structural boundaries that human viewers tend to prioritize. The TCB block is the second mechanism; it applies squeeze-and-excitation channel recalibration to transformer patch embeddings followed by 3x3 convolutions, distilling global context into local hierarchical features and shrinking parameter count by 23.3%. A final dilated-convolution stack with rates 2 and 4 widens the receptive field before global pooling and a fully connected layer predict the quality score. The training objective combines MSE with a consistency loss that penalizes disagreement between an image and its horizontal flip.","core_discovery":"The paper's central claim is that depth-guided attention improves NR-IQA generalization, and DGIQA is the empirical demonstration. Two parallel Swin-Transformer backbones extract features from the RGB image and its estimated depth map; four TCB blocks recalibrate channels and add local convolutions; then the Depth-CAR block uses depth-derived features as queries in scaled dot-product attention over RGB keys and values, followed by self-attention refinement and a dilated-convolution stack before score prediction. On the benchmark evaluations the model ranks in the top two across all seven datasets and takes first place on LIVE, CSIQ, Kadid10k, LIVE-C, and Koniq10k. In cross-dataset transfers it reports gains of roughly 0.8–5.0% in SROCC and up to 6.4% in PLCC over state-of-the-art baselines, and on unseen natural distortions its predicted-score distributions for low-light, haze, lens-flare, and motion-blur images overlap 3–95% less with high-quality distributions than the baselines do, with the strongest separation on hazy scenes (1.73% overlap). The authors also introduce 'density separation' as a new criterion for measuring NR-IQA generalization.","pith_inferences":["Beyond the paper: if depth maps remain stable under corruptions that destroy RGB texture (haze, blur, low light), depth-guided attention could extend naturally to video quality assessment, where temporal depth coherence would give a stronger prior than per-frame RGB statistics.","Beyond the paper: the density-separation measure could be adopted as a standard generalization test for NR-IQA, but its overlap percentages depend on the choice of high/low quality thresholds and should be calibrated against multiple baselines before being used as a headline metric.","Beyond the paper: because the depth estimator is trained mostly on clean scenes, heavily noisy or synthetic-distortion images may yield unreliable depth; testing DGIQA with depth maps from a distortion-robust estimator would clarify whether the improvement comes from depth quality or from the attention mechanism itself."],"forward_implications":["If the depth-guidance claim holds, NR-IQA models can be made substantially more robust to unseen real-world distortions by adding a depth branch, without requiring new subjective datasets.","The TCB design shows that fusing transformer global features with CNN local features can improve quality prediction while cutting parameters, pointing toward cheaper IQA models.","The reported cross-dataset gains imply that models trained on one large authentic dataset such as Koniq10k transfer better to other authentic and synthetic domains when depth-guided.","Density separation on external datasets such as LOL, IHAZE, Flare7k, and GoPro offers a practical protocol for testing generalization to natural distortions beyond standard benchmarks."],"supporting_citations":[{"why":"Generates the depth maps used as the second input modality when ground-truth depth is unavailable, making the depth branch possible.","marker":"[58]"},{"why":"Provides the Swin Transformer backbone whose hierarchical features feed the TCB and Depth-CAR blocks.","marker":"[31]"},{"why":"State-of-the-art transformer NR-IQA baseline that DGIQA compares against and outperforms on benchmark and cross-dataset evaluations.","marker":"[59]"},{"why":"State-of-the-art CNN NR-IQA baseline used in benchmark comparison and in the density-separation generalization analysis.","marker":"[1]"},{"why":"Source of the consistency-loss idea and a transformer-based baseline; its self-consistency regularizer is incorporated into DGIQA's objective.","marker":"[15]"},{"why":"Squeeze-and-excitation mechanism that the TCB block adapts for channel recalibration.","marker":"[19]"},{"why":"Scaled dot-product attention and self-attention formulation used by Depth-CAR for cross-attention and refinement.","marker":"[46]"},{"why":"Koniq10k, the authentic dataset used to train models for the unseen-distortion generalization experiments.","marker":"[18]"}],"fun_headline_variants":["DGIQA: depth-guided attention beats SOTA on unseen distortions","Depth cues sharpen blind IQA scores on haze, low light, and flares","Depth-guided feature attention improves blind IQA generalization","DGIQA uses depth to outperform cross-dataset blind IQA baselines","Depth-aware blind IQA model tops unseen distortions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that depth maps computed by a pretrained monocular estimator on distorted images preserve enough structural information for depth-guided attention to help quality prediction; the paper itself notes that on flat or depth-uniform scenes, or when estimated depth is noisy, the depth guidance does not help and can even cause the model to diverge.","fun_headline_variants_meta":{"raw":{"variants":["DGIQA: depth-guided attention beats SOTA on unseen distortions","Depth cues sharpen blind IQA scores on haze, low light, and flares","Depth-guided feature attention improves blind IQA generalization","DGIQA uses depth to outperform cross-dataset blind IQA baselines","Depth-aware blind IQA model tops unseen distortions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00077,"raw_usage":{"total_tokens":3442,"prompt_tokens":1009,"completion_tokens":2433,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":2346}},"tokens_in":625,"tokens_out":2433,"duration_ms":18172,"temperature":1.0,"reasoning_tokens":2346,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:37:27.381935+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DGIQA with the estimated depth maps replaced by random noise or by depth from a deliberately corrupted estimator; if cross-dataset SROCC and PLCC on the unseen natural distortion sets stay the same or improve, then depth content is not what drives the reported gains, and the depth-guidance claim collapses.","supporting_citations":[{"cited_title":"Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data","cited_arxiv_id":null,"evidence_quote":"Generates the depth maps used as the second input modality when ground-truth depth is unavailable, making the depth branch possible."},{"cited_title":"Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows","cited_arxiv_id":null,"evidence_quote":"Provides the Swin Transformer backbone whose hierarchical features feed the TCB and Depth-CAR blocks."},{"cited_title":"Maniqa: Multi-Dimension Attention Network for No- Reference Image Quality Assessment","cited_arxiv_id":null,"evidence_quote":"State-of-the-art transformer NR-IQA baseline that DGIQA compares against and outperforms on benchmark and cross-dataset evaluations."},{"cited_title":"Arniqa: Learning Distortion Mani- fold for Image Quality Assessment","cited_arxiv_id":null,"evidence_quote":"State-of-the-art CNN NR-IQA baseline used in benchmark comparison and in the density-separation generalization analysis."},{"cited_title":"No-Reference Image Quality Assessment Via Transformers, Relative Ranking, and Self-Consistency","cited_arxiv_id":null,"evidence_quote":"Source of the consistency-loss idea and a transformer-based baseline; its self-consistency regularizer is incorporated into DGIQA's objective."},{"cited_title":"Squeeze-and-Excitation Net- works","cited_arxiv_id":null,"evidence_quote":"Squeeze-and-excitation mechanism that the TCB block adapts for channel recalibration."},{"cited_title":"Gomez, Łukasz Kaiser, and Il- lia Polosukhin","cited_arxiv_id":null,"evidence_quote":"Scaled dot-product attention and self-attention formulation used by Depth-CAR for cross-attention and refinement."},{"cited_title":"KonIQ-10k: An Ecologically Valid Database for Deep Learning of Blind Image Quality Assessment","cited_arxiv_id":null,"evidence_quote":"Koniq10k, the authentic dataset used to train models for the unseen-distortion generalization experiments."}],"review_version":1}