{"id":"500e1c9d-cdba-4ac4-a9a0-f7bdb9401a06","arxiv_id":"2501.12246","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SPEINet detects sharp frames with a lightweight logistic-regression classifier and uses their features, together with edge-enhanced information, to deblur video, reporting top average PSNR on four benchmarks.","lead":"Researchers built SPEINet, a video deblurring model that finds naturally sharp frames in a blurred clip and uses them, plus edge details, to restore blurry frames. They also created a dataset with adjustable sharp-to-blur ratios to help models adapt to real-world video conditions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fair-comparison and table-integrity issues undermine the reported +3.2% average: baselines are trained on other distributions, and Table 1's average row cannot be reproduced from its own PSNR entries.","rationale":"The reader's weakest assumption is valid: cross-dataset training mismatch is the main reason the SOTA claim is not established. I add a more immediate red flag: the Average row in Table 1 cannot be reproduced from the printed per-dataset PSNRs for several baselines, and VDTR's GoProRS entry is implausible, so even the arithmetic basis of the headline is in doubt. These are fixable issues, so I keep the reader's conditional assignment rather than moving to reject or unverified. The decisive check is retraining a leading baseline on GoProRS and recomputing all averages from per-sequence metrics; if the corrected numbers still favor SPEINet by the claimed margin, the central claim would be credible.","tokens_in":19825,"tokens_out":16318,"duration_ms":154860,"concrete_test":"Retrain VRT (or TSPNL) on the GoProRS training split using SPEINet's exact protocol (patches, loss, epochs, evaluation code) and recompute Table 1's Average row from per-sequence PSNR/SSIM values using a single stated formula. If the retrained baseline's corrected average is still below SPEINet's 30.557 with a >2.7% gap, the concern fails; otherwise the SOTA claim is an artifact of training distribution and arithmetic.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The flagship result, a +3.2% average PSNR improvement (abstract; conclusion says +2.74%), is computed from Table 1's Average row. That row conflates architecture with training distribution: SPEINet is trained on GoProRS, while VRT, VDTR, TSP, TSPNL, DSTNet, and PVDNet are trained on GoProO or GoProS. The per-dataset lines show SPEINet is not best on GoProO (31.527 vs VRT 34.79) or GoProS (33.861 vs D2Net(S) 35.554); the advantage appears only in the average dominated by GoProRS, the distribution SPEINet was trained on. The numerical basis is also unstable. As printed, simple means of the six PSNR entries are 29.44 for TSPNL and 29.03 for VRT, not 29.6033 and 28.5287, while SPEINet's 30.557 matches a simple mean; VDTR's GoProRS value (5.96 dB, SSIM 0.0066) is implausible and suggests an evaluation artifact. Until the average is recomputed by a stated formula and at least one strong baseline is retrained on GoProRS, the headline improvement is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GoProRS, a new video deblurring dataset in which the proportion of sharp frames in a sequence is user-specified, and proposes SPEINet, a network that integrates a lightweight logistic-regression sharp-frame detector, a Richardson-Lucy edge-emphasis stage, and an attention-based encoder-decoder with search-transfer modules. The authors report that SPEINet, trained on GoProRS, achieves an average PSNR of 30.557 on GoProO, GoProS, GoProRS, and BSD, corresponding to a +3.2% average improvement over recent techniques in the abstract (the conclusion states +2.74%). The paper also includes ablations on the detection module and sub-module contributions.","tokens_in":20096,"tokens_out":2978,"duration_ms":28099,"significance":"If the claims are supported, the GoProRS dataset is a useful generalization of existing fixed-ratio sharp-frame datasets, and the lightweight classical-metric-based sharp detector is a practical contribution that could make sharp-frame prior methods more deployable. The paper is an empirical benchmark paper rather than a theoretical derivation, so the circularity burden is low; however, the central empirical claim currently rests on a comparison that is not controlled for training distribution, and the reported average numbers are not reproducible from the paper's own tables. These issues are fixable in principle, so the contribution is potentially significant but requires re-verification.","major_comments":[{"comment":"The Average row in Table 1 cannot be reproduced from the per-dataset PSNR entries. For example, TSPNL's six PSNR values (32.9042, 32.611, 32.608, 26.573, 25.401, 26.522) have a simple mean of about 29.44, but the table reports 29.6033; VRT's values have a simple mean of about 29.03, but the table reports 28.5287. The paper does not state the averaging formula (e.g., frame-count-weighted or per-sequence averaging), and the discrepancy undermines the headline average improvement. Please provide the exact aggregation procedure and recompute the table, or correct the entries.","section":"§5.3, Table 1"},{"comment":"The central performance claim conflates architecture with training distribution. SPEINet is trained on GoProRS, while VRT, VDTR, TSP, TSPNL, DSTNet, and PVDNet are trained on GoProO or GoProS. On GoProO, SPEINet (31.527 dB) is well below VRT (34.79 dB), and on GoProS, the GoProS-trained D2Net(S) (35.554 dB) outperforms SPEINet (33.861 dB). The average advantage is therefore driven largely by the GoProRS and BSD columns, which favor the training distribution of SPEINet. To support the claim that the architecture is responsible for the improvement, at least one strong baseline (e.g., VRT or TSPNL) should be retrained on GoProRS under identical settings, or the comparison should be explicitly framed as a cross-dataset generalization study with the appropriate caveats.","section":"§5.3, Table 1"},{"comment":"The VDTR results are implausible and indicate an evaluation artifact: across all tested sharp-frame ratios, VDTR reports PSNR between 5.58 and 6.64 dB with SSIM between 0.0060 and 0.0105, which is worse than a constant-gray prediction and physically meaningless for a video deblurring transformer. This also contaminates the GoProRS row of Table 1, where VDTR is listed at 5.96 dB. Please rerun the VDTR evaluation with the correct model weights, data preprocessing, and metric computation, and update all affected tables and averages.","section":"§5.4, Table 2"},{"comment":"The abstract claims an average '+3.2% PSNR improvement', while the conclusions state 'an improvement of over 2.74%'. Both refer to the same set of experiments and the same average PSNR of 30.557, so the two numbers are mutually inconsistent. Moreover, the percentage is not defined: it is unclear whether it is relative to the best baseline, the average of all baselines, or a specific method such as D2Net(RS). Please state the reference baseline and recompute the percentage from the corrected Table 1.","section":"Abstract and §8 Conclusions"},{"comment":"The comparison between the proposed logistic-regression detector (LD) and the Bi-LSTM detector is not matched on training data: the Bi-LSTM was trained on GoProS only, while LD is trained on GoProRS (Table 3) or GoProS (Table 4). The paper acknowledges this confound, but the sentence 'LD outperforms Bi-LSTM, improving detection accuracy by 15.56%' is still misleading, since the comparison in Table 3 is cross-dataset. Additionally, the 15.56% is a difference in percentage points (75.95 vs 60.39), not a 15.56% relative improvement. Please present a same-training-set comparison or clearly label the numbers as percentage-point differences.","section":"§6.1, Tables 3 and 4"},{"comment":"The ablation results in Table 5 report relative gains of 4.262% and 4.361% for the sharp-frame and edge modules, respectively. These percentages are not defined against a clear baseline: the rows compare different combinations of modules, but the baseline row (✗ ✗ ✗) has PSNR 29.592, and the +4.361% row (✗ ✓ ✓) has PSNR 33.953, which is a 14.7% relative increase, not 4.361%. Please clarify whether the percentages are relative improvements, absolute percentage-point increases, or something else, and make the computation consistent.","section":"§5.3, Table 1 and §6.2, Table 5"}],"minor_comments":[{"comment":"The abstract contains typos: 'essential task' should be 'an essential task', and 'the the frequency' has a duplicated article.","section":"Abstract"},{"comment":"The indexing in Eq. (1) is confusing: the definition of the starting index σ(i)j is not self-contained, and the superscript on v is unclear. Please rewrite with explicit ranges and clarify how windows are non-overlapping.","section":"§3, Eq. (1)"},{"comment":"The caption says 'the other method were trained on GoProO'; the grammar should be 'the other methods were trained on GoProO'.","section":"§5.4, Table 2 caption"},{"comment":"The phrase 'an logistic-regression-based detector' should be 'a logistic-regression-based detector'.","section":"§8 Conclusions"},{"comment":"Several references are cited without complete context (e.g., [12], [13], [41]). Please ensure all citations have full bibliographic entries and that the described contributions match the cited works.","section":"§2 Related works"},{"comment":"Figure 7 seems to show only six frames, but the text says 'six blur frames' and the figure caption says 'six frames'—please clarify whether these are frames from one sequence or separate sequences, and label the rows/columns accordingly.","section":"§7 Visualization"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially within scope for Pattern Recognition, but the experimental validation needs substantial rework before the central claim can be accepted. The most serious issues are the irreproducible average row in Table 1, the training-distribution confound in the headline comparison, the implausible VDTR numbers, and the abstract/conclusion discrepancy. These are not mere presentation issues; they directly affect the validity of the claimed +3.2% improvement. That said, the proposed dataset and the lightweight detector are concrete contributions that could be made solid with a corrected and properly controlled experimental section. I would recommend major revision with the expectation that the authors retrain baselines on GoProRS, recompute all tables, and align the reported percentages."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Name],\n\nQuick read on arXiv:2501.12246. The useful contribution here is GoProRS, a variable-ratio version of GoProS that lets you set the sharp-frame frequency. That's a genuine gap: prior sharp-frame deblurring datasets fix the ratio at 0.5, which is unrealistic. The paper also shows that a logistic-regression detector on six classical autofocus metrics is lightweight and, within 2.5% accuracy of a Bi-LSTM on GoProS, much faster. That part is solid and reproducible.\n\nThe model SPEINet is an assembly of known pieces — SwinIR, TTSR-style search-transfer, Richardson-Lucy edge emphasis, and an attention encoder-decoder. Nothing conceptually new, but the integration is reasonable and the ablation shows each component helps.\n\nWhere it gets shaky is the experimental comparison. The abstract claims +3.2% average PSNR over recent methods; the conclusion says over 2.74%. More importantly, the average row in Table 1 cannot be reproduced from the per-dataset PSNR entries. For TSPNL the six entries average to 29.44, but the table lists 29.6033; for VRT the simple mean is 29.03, not 28.5287. That's a red flag — either the averages are computed by a different formula or the numbers come from inconsistent runs. Also, VDTR's GoProRS numbers (PSNR 5.95, SSIM 0.0066) are implausible and suggest an evaluation bug. On top of that, most baselines were trained on GoProO or GoProS while SPEINet is trained on GoProRS, so the headline average mainly reflects training-distribution advantage. SPEINet is not best on GoProO or GoProS; it wins on GoProRS, the distribution it trained on.\n\nTo the paper's credit, the D2Net(RS) comparison is fair — both models trained on GoProRS — and SPEINet beats it by 0.8 dB. That gives some support for the architecture. The detector analysis is also careful.\n\nOverall: the dataset and the lightweight detector are worth publishing, and the model is a reasonable engineering effort. But as submitted, the central SOTA claim is not supported. A serious referee should ask for retraining at least VRT and D2Net on GoProRS, for a reproducible average table, for an explanation of the VDTR numbers, and for the actual code/data release.\n\nI'd send this to peer review, but with the expectation of major revision. It's not a desk reject.","headline":"Useful variable-ratio sharp-frame dataset and a fast logistic-regression detector, but the headline SOTA claim is undercut by unfair baselines and an unreproducible Table 1.","tokens_in":20657,"tokens_out":3258,"would_cite":false,"duration_ms":29068,"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":"This paper claims that making the sharp-frame frequency in a video deblurring dataset a tunable parameter, combined with a network that detects and transfers sharp-frame features and edge emphasis, improves average deblurring quality by…","keywords":["video deblurring","sharp frame detection","GoProRS dataset","attention-based encoder-decoder","Richardson-Lucy edge enhancement","logistic regression focus metrics","motion blur","domain adaptation"],"falsifier":"Retrain SPEINet on the same data as each baseline, such as GoProO only, and evaluate on the same four datasets; if the average PSNR advantage over VRT or D2Net disappears or reverses, the central claim that the architecture plus GoProRS generalizes better is falsified. A second check is to measure detection accuracy of the logistic-regression detector on videos whose sharp-frame ratio falls outside the training distribution, such as $r=0.02$; if sharp detection fails there, the claimed robustness to variable ratios does not hold.","tokens_in":19604,"feed_emoji":"🎬","tokens_out":6074,"duration_ms":54717,"temperature":0.7,"pith_summary":"This paper argues that fixed sharp-to-blur ratios in video deblurring datasets hinder generalization, and that letting the ratio vary makes deblurring models more adaptable. To test this, the authors build GoProRS from GoPro footage by sampling window lengths so a user-chosen fraction $r$ of frames come out sharp, and train SPEINet, a network that detects nearby sharp frames with logistic regression on six classical autofocus metrics, emphasizes edges via Richardson-Lucy, and transfers sharp-frame features into an attention-based encoder-decoder. The paper reports that SPEINet trained on GoProRS reaches an average PSNR of 30.557 across GoProO, GoProS, GoProRS, and BSD, roughly 3.2% higher than recent baselines, and that its advantage grows as sharp-frame frequency increases. If correct, the work would mean variable sharp-frame availability is a key axis for video deblurring performance, not a nuisance to be fixed at a single ratio.","feed_headline":"Tunable sharp-frame ratios lift video deblurring by 3.2%","feed_subtitle":"SPEINet detects sharp frames with classical autofocus metrics, then beats existing methods on four datasets.","key_machinery":"The load-bearing machinery is the conditional transfer of sharp-frame features through a search-and-transfer module patterned on texture transformers. Sharp frames are located by a logistic-regression detector over six classical focus measures (MIS3, GRA7, LAP1, STA3, DCT3, WAV1); each detected sharp frame is encoded at three scales, while the blurry triplet and its Richardson-Lucy edge-emphasized version are encoded separately. A SwinIR block fuses the blur and edge encodings, a similarity search finds corresponding sharp-feature patches, and soft-attention refinement with cross-scale feature integration lets the decoder reconstruct the central frame. When no sharp frame is found, a self-search branch substitutes transfer with within-blur similarity.","core_discovery":"The central claim is that a video deblurring model can be made both more accurate and more general by treating sharp frames as a tunable prior rather than a fixed assumption. Concretely, SPEINet detects the closest sharp frame among the previous seven frames using a logistic-regression classifier over six autofocus metrics, sharpens edges with Richardson-Lucy deconvolution, encodes blurry and sharp frames with local-attention residual blocks, and then transfers the sharp frame's features into the decoder through search-and-transfer and self-search branches. Trained on GoProRS with sharp ratios 0.1, 0.3, and 0.5, the model outperforms the compared state of the art on average and by a wider margin as the sharp ratio rises, which the authors interpret as evidence that the fixed 50% sharp ratio in GoProS biases models and hurts domain adaptation.","pith_inferences":["The paper does not test what happens when all methods are trained on GoProRS; a re-run with matched training data would separate the dataset's contribution from the architecture's contribution.","A natural extension the paper leaves open is whether the detector's seven-frame search window is optimal for clustered bursts of sharp frames rather than randomly scattered ones.","The reported gains are averaged across datasets with different blur severity; per-dataset error bars would show whether the claimed advantage is consistent or driven by one test split.","The detector's accuracy advantage is partly inherited from training on the same distribution, so evaluating on videos whose sharp ratio falls outside the training set, such as $r=0.02$, would test the claimed robustness more strictly."],"forward_implications":["If SPEINet's claim is right, video deblurring models trained on variable sharp ratios generalize to both heavily blurred and real-world footage better than models trained on fixed ratios.","The logistic-regression sharp detector is fast enough to replace a learned Bi-LSTM detector: about 5.71 seconds versus 46 seconds per GoProRS sequence, with comparable final PSNR.","Edge emphasis from Richardson-Lucy contributes materially, with an ablation gain of roughly 4.36% PSNR over the baseline without sharpness or edge modules.","Performance scales with sharp-frame frequency: SPEINet's PSNR on GoProRS rises from 32.032 at $r=0.02$ to 35.452 at $r=0.5$, while most compared baselines stay flat or degrade.","A variable-ratio dataset such as GoProRS can be used to retrain existing sharp-frame-based models, improving their generalization across datasets outside their original training distribution."],"supporting_citations":[{"why":"Introduces GoProS and D2Net, the sharp-frame-based baseline with a fixed 50% sharp ratio that GoProRS generalizes and against which SPEINet is compared.","marker":"[38]"},{"why":"Provides the original GoPro high-frame-rate videos from which GoProRS is synthesized.","marker":"[23]"},{"why":"Supplies the six focus-measure operators used by the logistic-regression sharp detector.","marker":"[29]"},{"why":"Richardson-Lucy method on which the regularized edge emphasis module is based.","marker":"[34]"},{"why":"Provides the search-transfer and soft-attention decoder machinery adapted for sharp-to-blur feature transfer.","marker":"[51]"},{"why":"SwinIR fuses blur and edge features before the search-and-transfer stage.","marker":"[17]"},{"why":"Temporal sharpness prior and non-local similarity baseline (TSPNL) used in the comparison tables.","marker":"[26]"},{"why":"VRT video restoration transformer baseline that performs best on GoProO and serves as a strong comparison.","marker":"[16]"}],"fun_headline_variants":["SPEINet boosts video deblurring by 3.2% with tunable sharp frames","Adaptive sharp-frame detection sharpens video by 3.2%","GoProRS dataset and SPEINet lift video deblurring accuracy","Sharpness prior and edge details add 3.2% to video deblurring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main load-bearing assumption is that averaging numbers across datasets with different training distributions is a fair comparison: most baselines were trained on GoProO or GoProS while SPEINet was trained on GoProRS, so the reported PSNR advantage is assumed to reflect the model rather than the training data.","fun_headline_variants_meta":{"raw":{"variants":["SPEINet boosts video deblurring by 3.2% with tunable sharp frames","Adaptive sharp-frame detection sharpens video by 3.2%","GoProRS dataset and SPEINet lift video deblurring accuracy","Sharpness prior and edge details add 3.2% to video deblurring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000378,"raw_usage":{"total_tokens":2011,"prompt_tokens":949,"completion_tokens":1062,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":971}},"tokens_in":565,"tokens_out":1062,"duration_ms":8988,"temperature":1.0,"reasoning_tokens":971,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:21:11.222384+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain SPEINet on the same data as each baseline, such as GoProO only, and evaluate on the same four datasets; if the average PSNR advantage over VRT or D2Net disappears or reverses, the central claim that the architecture plus GoProRS generalizes better is falsified. A second check is to measure detection accuracy of the logistic-regression detector on videos whose sharp-frame ratio falls outside the training distribution, such as $r=0.02$; if sharp detection fails there, the claimed robustness to variable ratios does not hold.","supporting_citations":[{"cited_title":"Bringing events into video deblurring with non-consecutively blurry frames, in: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"Introduces GoProS and D2Net, the sharp-frame-based baseline with a fixed 50% sharp ratio that GoProRS generalizes and against which SPEINet is compared."},{"cited_title":"Deep multi-scale convolutional neural network for dynamic scene deblurring, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Provides the original GoPro high-frame-rate videos from which GoProRS is synthesized."},{"cited_title":"Analysis of focus measure opera- tors for shape-from-focus","cited_arxiv_id":null,"evidence_quote":"Supplies the six focus-measure operators used by the logistic-regression sharp detector."},{"cited_title":"Bayesian-based iterative method of image restora- tion","cited_arxiv_id":null,"evidence_quote":"Richardson-Lucy method on which the regularized edge emphasis module is based."},{"cited_title":"Learning texture transformer network for image super-resolution, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Provides the search-transfer and soft-attention decoder machinery adapted for sharp-to-blur feature transfer."},{"cited_title":"Swinir: Image restoration using swin transformer, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"SwinIR fuses blur and edge features before the search-and-transfer stage."},{"cited_title":"Cascaded deep video deblurring using temporal sharpness prior and non-local spatial-temporal similarity","cited_arxiv_id":null,"evidence_quote":"Temporal sharpness prior and non-local similarity baseline (TSPNL) used in the comparison tables."},{"cited_title":"Vrt: A video restoration transformer","cited_arxiv_id":null,"evidence_quote":"VRT video restoration transformer baseline that performs best on GoProO and serves as a strong comparison."}],"review_version":1}