{"id":"6fc15960-0d01-4300-bd1f-db214e4c449e","arxiv_id":"1909.01738","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PAD-Net, a Siamese encoder-decoder with a ResNet-18 quality regressor, predicts no-reference stereoscopic image quality by modeling binocular rivalry through likelihood and prior maps, and reports state-of-the-art results on LIVE Phase I/II and Waterloo IVC Phase I.","lead":"PAD-Net is a deep network that predicts the perceptual quality of 3D stereo images without a reference by reconstructing each eye's view and fusing the reconstruction errors with learned feature maps. It reports top correlation with human scores on three standard stereo image quality databases, especially for asymmetrically distorted pairs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random 80/20 split by stereopair places the same reference scenes in both training and test sets, so the reported SROCC/PLCC gains may reflect content memorization; the cross-database tests are mixed and do not settle the issue.","rationale":"The paper is clearly written, the architecture is plausible, the source code is promised, and the ablations show that the likelihood and prior maps contribute under the chosen protocol. The central claim, however, is about better quality estimation than existing metrics, especially for unseen asymmetric distortions. That requires a test protocol that prevents reference-content leakage. The random 80/20 split does not satisfy this because the databases have few references and many distorted versions per reference. The cross-database table is the natural safeguard, but it is only partially supportive and too coarse to validate the asymmetric-distortion advantage. I therefore agree with the reader's conditional verdict and would not change it.","tokens_in":20580,"tokens_out":4527,"duration_ms":48311,"concrete_test":"Run leave-one-reference-out evaluation on LIVE Phase II and Waterloo IVC Phase I: for each reference scene, train PAD-Net on all distorted versions of the other references and test on all distorted versions of the held-out reference; average SROCC/PLCC over held-out references and report them separately for symmetric and asymmetric pairs. Apply the identical protocol to StereoQA-Net. If the asymmetric SROCC gap (0.954 vs 0.927 on LIVE Phase II) collapses or reverses, the manuscript's central claim is unsupported and should be revised; if the gap persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B states that distorted stereo pairs are randomly split 80/20 'according to [34]'. In all three databases, each reference scene has many distorted versions (LIVE Phase II: 8 references, 360 images; Waterloo IVC Phase I: 6 references, 330 images), so with high probability every reference appearing in the 20% test set also appears in the 80% training set. Since the network is trained on 256x256 crops and tested with heavily overlapping crops, test sub-images can be near-duplicates of training crops from the same reference, allowing the regressor to memorize reference-specific content and distortion signatures rather than learn a general quality mapping. This directly threatens the headline SROCC values (0.973/0.967/0.974) and the asymmetric-distortion advantage (Table IX, 0.954 vs 0.927 on LIVE Phase II). The 10-run t-test (Table VI) only measures variance of the same leaky protocol, not independence. The cross-database experiments (Table X) are the relevant counter-evidence, but they report only PLCC, no SROCC or asymmetric sub-scores, and PAD-Net is not the best method in the LIVE II to LIVE I direction (0.915 vs 0.932 for StereoQA-Net). Thus the central claim of SOTA generalization, especially for asymmetric distortion, is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PAD-Net, an end-to-end no-reference stereoscopic image quality metric. It uses a Siamese encoder-decoder to reconstruct the left and right views, interprets the reconstruction error as a likelihood map and high-level encoder features as a prior map, and fuses these normalized maps with the original views before a ResNet-18 regressor predicts the MOS/DMOS. The network is pretrained on 2D databases and jointly fine-tuned on 3D databases. Experiments on LIVE Phase I, LIVE Phase II, and Waterloo IVC Phase I report state-of-the-art SROCC/PLCC, including improved performance on asymmetrically distorted pairs, together with ablation, cross-database, and complexity analyses.","tokens_in":20878,"tokens_out":4568,"duration_ms":47355,"significance":"If the reported results hold under a contamination-free protocol, the paper would be a solid contribution to blind SIQM: it builds a network around a biologically motivated rivalry model, releases source code, and reports strong correlation with subjective scores, including an asymmetric-distortion breakdown and a Krasula-based significance analysis. The architecture is described in detail, and the ablation partially isolates the effect of the likelihood/prior maps. However, the empirical claims currently rest on a train/test split that does not guarantee content independence, and the cross-database evidence is mixed and under-reported. The contribution is significant but needs validation under a content-disjoint protocol.","major_comments":[{"comment":"The random 80/20 split by distorted stereopairs does not guarantee content independence because each reference scene has many distorted versions (e.g., LIVE Phase II has 8 references and 360 images; Waterloo IVC Phase I has 6 references and 330 images). The same reference content therefore almost certainly appears in both training and testing, and the testing procedure uses heavily overlapping 256×256 crops with stride 192×104. This allows the regressor to memorize reference-specific content and distortion signatures, inflating the headline SROCC values (0.973, 0.967, 0.974) and the asymmetric-distortion advantage in Table IX. The 10-run t-test in Table VI only measures the variance of the same protocol. Please re-run the evaluation with a content-disjoint split (e.g., split by reference scene) and report SROCC, PLCC, and RMSE, including the symmetric/asymmetric breakdown; the central state-of-the-art claim is not supported until this is done.","section":"IV-B (Tables IV, V, IX)"},{"comment":"The ablation study does not cleanly establish that the likelihood/prior maps are the cause of the improvement. In the joint optimization stage, the encoder-decoder weights w1, the regression weights w2, and the newly introduced w3 are all updated with loss l3 on the target MOS labels (Eqs. 14-16). Adding the map-generation and fusion branches therefore adds trainable parameters that are directly supervised by the target scores, so the gain over the baseline could come from additional capacity rather than from the binocular-rivalry prior. To support the mechanistic claim, the maps should be generated by a frozen pretrained encoder-decoder or by an unsupervised objective, and the comparison repeated.","section":"III-D, Fig. 7"},{"comment":"The cross-database experiments, which are the main evidence that does not suffer from the same-content training and testing issue, are only reported in terms of PLCC and are mixed: trained on LIVE Phase II and tested on LIVE Phase I, PAD-Net (0.915) is below StereoQA-Net (0.932). Moreover, no SROCC or symmetric/asymmetric sub-scores are given, so the claim of superior generalization for asymmetric distortions is not established by this table. Please report complete correlation metrics and a per-condition breakdown for both cross-database directions.","section":"IV-E, Table X"}],"minor_comments":[{"comment":"The word 'commutation' appears to be a typo; 'communication systems' is the likely intended phrase.","section":"Abstract"},{"comment":"The sentence 'which will be explained in Section III D' should refer to Section IV-D, where the symmetric/asymmetric evaluation is presented.","section":"IV-B"},{"comment":"'Normlization' is misspelled, and the labels 'Square7a'/'Square7b' are inconsistently formatted.","section":"Table II"},{"comment":"The testing-stage cropping rule is described only as a stride of 192 for width and 104 for height; please clarify how boundary crops are handled when the image dimensions are not exact multiples of the stride.","section":"III-D"}],"recommendation":"major_revision","confidential_remarks":"The content-disjoint split issue is likely to be raised by other reviewers and should be addressed head-on in the revision. The authors should also be encouraged to provide the cross-database results with the same detail as the within-database tables, since those results are the cleanest available test of generalization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, PAD-Net is a reasonable, clearly described neural network for blind stereoscopic image quality prediction: a Siamese encoder-decoder built from Ballé-style components produces likelihood/prior maps from reconstruction error and high-level features, and those maps plus the original views feed a ResNet-18 regressor. Code is public. Second, the main empirical claim—state of the art on LIVE Phase I/II and Waterloo IVC—is weakened by the evaluation protocol. Section IV-B says distorted stereopairs are randomly split 80/20 \"according to [34]\". Because each reference scene appears in many distorted versions, the same content lands in both train and test. With 256x256 crops and heavy overlap at test time, the regressor can memorize reference-specific content rather than learn a general quality mapping. The 10-run t-test reports variance under the same leaky protocol. The cross-database results, which are the right sanity check, are mixed: PAD-Net trails StereoQA-Net on LIVE II→I (0.915 vs 0.932) and beats it on LIVE I→II by 0.002 (0.854 vs 0.852). So the SOTA claim, especially the asymmetric-distortion advantage in Table IX, is not yet established.\n\nWhat is genuinely new is the specific architecture combination and the attempt to model binocular rivalry through normalized likelihood/prior maps. The ablation and runtime tables are useful, and the paper is honest enough to include per-distortion results where it does not always win (e.g. JPEG on LIVE Phase I). I also credit the authors for releasing source code.\n\nThe softer spots: the predictive-coding framing is looser than the prose implies, because the prior/likelihood maps are fine-tuned on MOS, so the ablation gains show the maps help regression, not that they implement binocular rivalry. And the authors never compare to their own earlier predictive-coding SIQA [55], which would be the natural baseline. These are fixable. The content-leak issue is less fixable by tweaks; it needs a different split protocol.\n\nWho this is for: anyone working in no-reference image/video quality assessment, especially stereo. It deserves a serious referee and could become a solid methods paper after the authors re-run the core experiments with content-disjoint splits and report SROCC/PLCC with variance across splits. I would not cite the headline numbers until that re-run happens, but I would engage with the architecture.","headline":"A useful NR-SIQM architecture with open code, but the headline SOTA numbers rest on a content-leaky 80/20 split and should be re-run on content-disjoint splits before being trusted.","tokens_in":21397,"tokens_out":3159,"would_cite":true,"duration_ms":31872,"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":"Modeling binocular rivalry as predictive coding gives a blind stereoscopic quality metric its edge, particularly on unequally distorted view pairs.","keywords":["stereoscopic image quality assessment","no-reference quality metric","binocular rivalry","predictive coding","encoder-decoder network","asymmetric distortion","Siamese network","quality regression"],"falsifier":"Re-run the LIVE Phase II and Waterloo IVC Phase I evaluations with training and test sets separated by reference scene, so no distorted version of any test scene appears in training, and compare asymmetric-pair SROCC to the reported 0.954 and 0.978; a large drop would show the rivalry maps are memorizing content rather than measuring quality.","tokens_in":20373,"feed_emoji":"👁","tokens_out":8322,"duration_ms":71614,"temperature":0.7,"pith_summary":"The paper tries to show that a no-reference stereoscopic image quality metric can be improved by building in a specific model of how the brain handles mismatched views: binocular rivalry as predictive coding. Instead of treating the left and right views as independent images to be averaged, the proposed PAD-Net reconstructs each view with an auto-encoder, converts reconstruction error into a likelihood map and high-level features into a prior map, then lets those maps compete through normalization. These rivalry maps, together with the distorted views, are fed to a quality regression network. On three benchmark databases the reported correlations with human opinion scores exceed those of the compared full-, reduced-, and no-reference metrics, with the clearest gains on asymmetrically distorted pairs.","feed_headline":"A rivalry-aware network tops blind 3D image quality benchmarks","feed_subtitle":"Using error and prior maps to model eye dominance lifts asymmetric-pair SROCC to 0.95 on LIVE Phase II.","key_machinery":"The load-bearing object is the pair of normalized rivalry maps derived from the Siamese encoder-decoder. The likelihood map is the squared difference between the input view and its reconstruction, inverted so small error means high likelihood; the prior map is the softplus-activated high-level encoder feature. Equations (7) and (8) normalize the left and right versions by their sum, and the left likelihood uses the right error (and vice versa), so each map expresses one eye's dominance relative to the other. A convolution-plus-GDN fusion layer combines the two views with the two likelihood and two prior maps into a three-channel input for a ResNet-18 regressor with max pooling and a fully connected layer. This fusion is what lets the network read which view is winning the rivalry, and the paper's ablation shows the maps and the fusion both contribute on the asymmetric-heavy LIVE Phase II database.","core_discovery":"The central claim is that predictive coding offers a usable computational recipe for binocular rivalry in quality assessment. For each view, the encoder-decoder's reconstruction error defines a likelihood map (small error means high likelihood), while the compressed high-level representation defines a prior map; normalizing each map by the sum of the two views' maps turns them into explicit dominance probabilities. The paper reports that this design, trained with two 2D pre-training stages followed by joint optimization on stereo databases, outperforms prior metrics: SROCC of 0.973 on LIVE Phase I, 0.967 on LIVE Phase II, and 0.974 on Waterloo IVC Phase I, with the largest margin over rivals on asymmetric distortion subsets. It also reports better statistical significance in t-tests and better discrimination under the Krasula methodology on Waterloo IVC Phase I.","pith_inferences":["If the rivalry maps are doing genuine perceptual work, they should predict the outcome of psychophysical binocular-rivalry experiments: the view with the larger normalized likelihood or prior sum should be the one subjects report seeing, which could be tested directly.","The same likelihood-prior construction could be lifted into a no-reference 2D quality model by treating the reconstruction-error map as a distortion mask, an extension the authors do not explore.","The content-split concern suggests a stronger evaluation: train on distortions from some reference scenes and test on entirely unseen scenes, because the existing cross-database test only partially isolates memorization from quality learning."],"forward_implications":["On the three evaluated databases, the reported correlations with human scores exceed those of the compared metrics, so a rivalry-aware design appears to capture information that view-averaging or cyclopean-image approaches miss.","Asymmetric distortion handling improves: on the LIVE Phase II asymmetric pairs the reported SROCC is 0.954, clearly above the best compared method's 0.927.","The normalized likelihood and prior maps give a per-pixel account of which eye's signal dominates, making the model's predictions interpretable rather than a black-box score.","Transfer learning from large 2D quality databases to the small 3D databases is enough to train a competitive deep stereo metric, reducing the need for large 3D subjective datasets.","At 0.906 seconds per image, the method is fast enough for practical use, roughly nine times faster than the compared CNN baseline."],"supporting_citations":[{"why":"It supplies the binocular-rivalry-as-predictive-coding theory that motivates the likelihood and prior formulation.","marker":"[49]"},{"why":"It provides the hierarchical predictive-coding model that the encoder-decoder reconstruction is built to mimic.","marker":"[52]"},{"why":"It supplies the encoder-decoder structure with GDN/IGDN layers used to reconstruct each view.","marker":"[65]"},{"why":"It contributes the sub-image training and end-to-end regression strategy used to turn images into quality scores.","marker":"[59]"},{"why":"It is the leading compared deep no-reference baseline and the source of the 80/20 train-test protocol.","marker":"[34]"},{"why":"It provides the Waterloo IVC Phase I asymmetric-distortion database and the finding that blur quality is set by the good view while noise is set by the bad view.","marker":"[45]"},{"why":"It provides LIVE Phase I, one of the three benchmark databases with subjective scores used in evaluation.","marker":"[62]"},{"why":"It provides LIVE Phase II with its asymmetric pairs and the cyclopean full-reference baseline that accounts for rivalry.","marker":"[23]"}],"fun_headline_variants":["Rivalry-based autoencoder sets new bar for blind stereo quality","Predictive coding eye model tops blind 3D image quality tests","Error maps mimic eye rivalry to beat stereo QA benchmarks","Blind stereo QA improved by modeling binocular rivalry","PAD-Net: rivalry-aware blind stereo quality measurement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that randomly splitting distorted stereo pairs into 80% training and 20% testing makes the sets independent, even though distorted versions of the same reference scene can appear in both, so the reported correlations could come partly from content memorization rather than quality understanding.","fun_headline_variants_meta":{"raw":{"variants":["Rivalry-based autoencoder sets new bar for blind stereo quality","Predictive coding eye model tops blind 3D image quality tests","Error maps mimic eye rivalry to beat stereo QA benchmarks","Blind stereo QA improved by modeling binocular rivalry","PAD-Net: rivalry-aware blind stereo quality measurement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000802,"raw_usage":{"total_tokens":3530,"prompt_tokens":953,"completion_tokens":2577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2495}},"tokens_in":569,"tokens_out":2577,"duration_ms":18336,"temperature":1.0,"reasoning_tokens":2495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:08:23.574806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the LIVE Phase II and Waterloo IVC Phase I evaluations with training and test sets separated by reference scene, so no distorted version of any test scene appears in training, and compare asymmetric-pair SROCC to the reported 0.954 and 0.978; a large drop would show the rivalry maps are memorizing content rather than measuring quality.","supporting_citations":[{"cited_title":"Predictive coding explains binocular rivalry: An epistemological review,","cited_arxiv_id":null,"evidence_quote":"It supplies the binocular-rivalry-as-predictive-coding theory that motivates the likelihood and prior formulation."},{"cited_title":"Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects,","cited_arxiv_id":null,"evidence_quote":"It provides the hierarchical predictive-coding model that the encoder-decoder reconstruction is built to mimic."},{"cited_title":"Dual-stream interactive networks for no- reference stereoscopic image quality assessment,","cited_arxiv_id":null,"evidence_quote":"It is the leading compared deep no-reference baseline and the source of the 80/20 train-test protocol."},{"cited_title":"Quality prediction of asymmetrically distorted stereoscopic 3d images,","cited_arxiv_id":null,"evidence_quote":"It provides the Waterloo IVC Phase I asymmetric-distortion database and the finding that blur quality is set by the good view while noise is set by the bad view."},{"cited_title":"Subjective evaluation of stereoscopic image quality,","cited_arxiv_id":null,"evidence_quote":"It provides LIVE Phase I, one of the three benchmark databases with subjective scores used in evaluation."}],"review_version":1}