{"id":"0809da30-edfc-4c43-9241-f45a6f8a38f7","arxiv_id":"2506.04054","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A three-stage video deblurring network (preprocessing, aligned deconvolution, reliability-map aggregation) reports improved PSNR over 2019-era baselines on Adobe240 and GOPRO.","lead":"This paper presents a three-stage neural network for removing motion blur from videos, combining frame preprocessing, alignment-based deblurring, and sharpness-based aggregation. The method reports higher PSNR than older video deblurring baselines on two public benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The state-of-the-art claim is not supported by the comparison set: all learned baselines are pre-2020 and the strongest 2019 baseline is beaten by less than 0.45 dB, so Table 1 cannot establish 'superior to existing SOTA.'","rationale":"The reader's weakest-assumption analysis focused on optical-flow accuracy and occlusion handling. That is a legitimate concern, but it is secondary because almost all alignment-based video deblurring methods share the same dependency, and the paper's ablation (PPN before ABDN outperforms ABDN+ABDN) partially addresses the misalignment risk. The more load-bearing issue is the empirical comparison itself: the central claim is explicitly a comparative state-of-the-art claim, and the evidence set stops at 2019 baselines while the paper is posted in 2025. This is not an internal inconsistency in the method, but it is a correctness risk for the conclusion as worded. The small PSNR margins over STFAN and the absence of statistical testing further weaken the claim. A concrete test using modern checkpoints would settle the matter. I agree with the reader's overall CONDITIONAL judgment, so the verdict should remain unchanged.","tokens_in":11031,"tokens_out":5810,"duration_ms":54455,"concrete_test":"Run official checkpoints of representative post-2019 video deblurring/restoration models (e.g., VRT, RVRT, ESTRNN) on the same Adobe240 and GOPRO test sets with the same PSNR computation, plus a paired per-sequence statistical comparison against STFAN and the reported DAN numbers. If any 2020+ model exceeds 31.618 dB or 33.003 dB, or if the DAN-vs-STFAN margin fails a paired p<0.05 test, the state-of-the-art claim as stated in the abstract is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DAN was demonstrated to be superior to existing state-of-the-art methods. The only quantitative evidence is Table 1, whose learned baselines are OVD (2017), DBN (2017), EDVR (2019, trained on REDS) and STFAN (2019). The EDVR row is admittedly an apples-to-oranges comparison because it was not trained on the same data. The margins over STFAN are 0.367 dB on Adobe240 and 0.443 dB on GOPRO, with no error bars, per-sequence statistics, or significance testing. A 2025 submission cannot use a comparison set that ends in 2019 to establish state-of-the-art status; methods from 2020 onward (recurrent and transformer video-restoration models) are absent. Thus the abstract's literal claim is under-supported rather than demonstrated. The internal ablations in Tables 3 and 4 are coherent and support the module-ordering design, but they do not bear on the comparative SOTA claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a video deblurring network, DAN, composed of three sub-networks: a preprocessing network (PPN) with non-local operations, an alignment-based deconvolution network (ABDN) using optical flow and occlusion maps, and a frame aggregation network (FAN) that merges deblurred frames according to learned reliability maps. The method is trained on the Adobe240 dataset and evaluated on Adobe240 and GOPRO. Table 1 reports average PSNRs of 31.618 dB and 33.003 dB, which the authors claim outperform existing state-of-the-art methods, with the strongest compared baseline (STFAN) at 31.251 dB and 32.560 dB. The paper also includes ablations showing that each module contributes positively to performance.","tokens_in":11210,"tokens_out":3096,"duration_ms":32103,"significance":"If the central claim were fully supported, the architecture would be a reasonable contribution: the idea of preprocessing blurry frames before optical-flow-based alignment, and then aggregating deblurred frames with reliability maps, is coherent and the ablations in Tables 3 and 4 provide internally consistent evidence that each module helps. However, the paper's headline claim of state-of-the-art performance rests on a comparison set whose newest learned baseline is from 2019, with one baseline (EDVR) trained on a different dataset, and with no error bars or statistical significance testing. The reported margins over STFAN are 0.367 dB and 0.443 dB, which are too small to establish superiority without uncertainty quantification. The paper does not provide code or pretrained models, so the empirical results are not independently reproducible from the manuscript. The strengths are the clear module-level analysis and the explicit statement of the training data and the EDVR training-data mismatch.","major_comments":[{"comment":"The central claim of state-of-the-art performance is not supported by the comparison set. All learned baselines are from 2017-2019, and video deblurring has advanced substantially since then with recurrent and transformer-based methods. To substantiate the abstract's claim, the authors should compare against recent strong methods, ideally retrained or fine-tuned on the same training data, and report per-sequence results and run-to-run variance.","section":"§5.1, Table 1"},{"comment":"The EDVR row is an apples-to-oranges comparison because the authors state that EDVR was trained on REDS, not on the Adobe240 training set used for the other methods. A model trained on a different data distribution cannot be directly compared in a benchmark table. Either retrain EDVR on the same data or remove this row; in its current form it should not be used as evidence for superiority.","section":"§5.1, Table 1"},{"comment":"The reported margins over STFAN are only 0.367 dB on Adobe240 and 0.443 dB on GOPRO, yet no error bars, multiple-run statistics, or significance tests are provided. Given that training stochasticity alone can produce differences of this magnitude, the claim that DAN is 'superior' needs at least standard deviation across multiple seeds or per-sequence breakdowns to be convincing.","section":"§5.1, Table 1"},{"comment":"The statement that the light version of the proposed method 'also outperforms the other video deblurring methods' inherits the same comparison limitations as Table 1: the baselines are all pre-2020, and the margins are small. In addition, the processing time comparison is incomplete without information about the hardware and software versions used for all methods; the 1.10-1.21 s/frame figures for the proposed method are considerably slower than OVD and STFAN, which should be discussed in relation to the 'advantage' claim.","section":"§5.3, Table 2"}],"minor_comments":[{"comment":"The occlusion-detection thresholds α1 and α2 are fixed to 0.01 and 0.5 without a sensitivity analysis; a short experiment varying these values would strengthen the claim that the chosen values are robust.","section":"§3.3, Eq. (5)"},{"comment":"The notation for time-step superscripts and subscripts is confusing, e.g., 'Pt+1_t' and 'Pt+1_t+1'. Please clarify the indexing so that the reader can map the equations to Figure 2 without ambiguity.","section":"§3.2, Eqs. (1)-(4)"},{"comment":"The qualitative comparisons would be more informative with zoomed crops or a quantitative sharpness/error map in addition to the small red/yellow boxes; as presented, it is hard to judge the claimed visual differences.","section":"§5.2, Figures 5 and 6"},{"comment":"The related-work section does not cite any video deblurring method from 2020 or later; this makes the comparison in Section 5.1 appear dated and also gives the reader no context for how the method relates to recent recurrent or transformer-based approaches.","section":"§2.3"},{"comment":"The paper has no explicit limitations section. Given the reliance on optical flow and the relatively high inference time, a brief discussion of failure cases (e.g., large occlusions or textureless regions) and of the real-time feasibility would improve the manuscript.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"The paper appears technically sound in its internal design and ablations, but the empirical claim of state-of-the-art performance is the main selling point and it is currently under-supported. If the authors can add recent baselines, retrain EDVR on the same data, and provide uncertainty estimates, the paper could become a solid contribution. Without those additions, the manuscript would be better framed as 'competitive with 2019-era methods' rather than 'superior to existing state-of-the-art.'"},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, incremental video-deblurring paper whose ablations are more convincing than its headline comparison. The proposed DAN chains three known ideas—non-local preprocessing, optical-flow alignment with occlusion correction, and reliability-map aggregation—and the module-level experiments actually show each stage pulling its weight. That is real work and the design story is coherent.\n\nWhat's new: the specific combination, plus the choice to align on preprocessed frames rather than raw blurry ones, which is a sensible answer to the misalignment problem. The PPN+ABDN vs ABDN+ABDN ablation in Table 3 is a nice piece of evidence. I also appreciate the light/full versions and the reporting of runtimes.\n\nSoft spots: the comparison set is the problem. Every learned baseline is from 2017 or 2019, and the EDVR row is admittedly trained on REDS, so it is not a same-protocol competitor. The margin over STFAN is 0.37–0.44 dB, with no error bars, per-sequence stats, or significance testing. A 2025 submission cannot call that 'superior to existing state-of-the-art' without including 2020+ methods. The claim in the abstract is under-supported, not false; within the tested comparison set, the method does win. The other weaknesses are also real but minor-to-moderate: no code or weights, no qualitative failure cases, and the occlusion threshold constants are reported but not sensitivity-analyzed.\n\nOn the math/data side: nothing is circular. The central claim is empirical; PSNR is measured. The flow-consistency occlusion test is standard, and the authors cite it properly. No invented entities, no parameter-fitting dressed as prediction.\n\nBottom line: this deserves a serious referee, but as a revision, not as-is. An editor should send it out; a referee should ask for modern baselines, error bars or per-sequence results, and code/weights. If those land, the empirical claim becomes credible. As it stands, the right verdict is 'revise before accept,' and the abstract should be toned down to 'outperforms prior published deep video deblurring methods evaluated under the same protocol' until the comparison set is updated.\n\nFor a reading group: worth a session on how to evaluate 'SOTA' when the field moves fast. I wouldn't cite it in my own work yet, but I'd revisit it after revision.","headline":"A competent three-stage video deblurring pipeline whose module ablations are convincing, but the state-of-the-art claim rests on a pre-2020 baseline set and needs recalibration.","tokens_in":11788,"tokens_out":4362,"would_cite":false,"duration_ms":36302,"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":"A three-stage network improves video deblurring beyond prior baselines on two benchmarks","keywords":["video deblurring","motion blur","non-local operations","optical flow alignment","frame aggregation","reliability map","occlusion detection","deep neural network"],"falsifier":"Run the full DAN on a video containing large textureless regions and heavy motion blur, then replace the estimated optical flow with ground-truth flow obtained from an ultra-high-speed capture; if the PSNR jumps by far more than the 0.37 to 0.44 dB margin over STFAN, the claimed result is dominated by flow accuracy rather than by the network design.","tokens_in":1824,"feed_emoji":"🎬","tokens_out":2830,"duration_ms":66132,"temperature":0.7,"pith_summary":"The paper argues that video deblurring fails when neighbor frames are used without first making them sharp enough to align reliably. It proposes a deconvolution and aggregation network (DAN) with three sub-networks: a preprocessing network applies non-local operations to reduce blur before alignment; an alignment-based deconvolution network warps neighbors to the central frame using optical flow and occlusion maps; and a frame aggregation network fuses the deblurred frames according to pixel-wise reliability maps. Trained with the Adobe240 protocol, DAN reports average PSNR of 31.618 dB on Adobe240 and 33.003 dB on GOPRO, above the strongest compared baseline STFAN at 31.251 dB and 32.560 dB. The claim is that a proper combination of deconvolution before alignment and aggregation after it lets neighbor information be used correctly.","feed_headline":"Video deblurring beats prior state of the art on two benchmarks","feed_subtitle":"It pre-clears frames with non-local operations, aligns them by optical flow, then fuses pixels by estimated sharpness.","key_machinery":"The load-bearing object is the three-stage pipeline whose linking step is the forward-backward optical flow consistency inequality, Eq. (5): $|w_f(x)+w_b(x+w_f(x))| < \\alpha_1(|w_f(x)|^2 + |w_b(x+w_f(x))|^2) + \\alpha_2$, with $\\alpha_1=0.01$ and $\\alpha_2=0.5$. Non-local blocks in the PPN operate on concatenated feature maps rather than raw frames to keep memory use low, and the FAN's final frame is $A_t = \\hat{A}_t^f \\odot RM_{t-1} + D_t \\odot RM_t + \\hat{D}_t^r \\odot RM_{t+1}$, a sharpness-weighted sum of a warped previous output, the current deblurred frame, and a warped future deblurred frame. The occlusion maps are also fed into the FAN so that the reliability maps learn to distrust occluded regions.","core_discovery":"The central discovery is that deblurring and alignment should be treated as a combined problem, with deblurring done twice: first a rough restoration by non-local feature matching across frames, then a restoration on flow-aligned preprocessed frames, and finally a fusion of multiple restored frames guided by reliability maps. Each stage contributes: the PPN removes enough blur that optical flow between frames becomes trustworthy; the ABDN uses forward-backward flow consistency to mark occlusions and replaces occluded warped pixels with the central frame's pixels; and the FAN weights each pixel by estimated sharpness rather than by a fixed temporal scheme. The result shown is that each module adds measurable PSNR and the full pipeline outperforms the compared methods on both benchmark test sets.","pith_inferences":["Editorial inference: the same two-pass recipe, rough restoration before alignment, could transfer to other flow-dependent video restoration tasks such as super-resolution or denoising, where blur or noise currently corrupts motion estimation.","Editorial inference: the reliability-map fusion in Eq. (8) is a learned per-pixel weighting over a three-frame window; extending it to longer recurrent memory could help when blur persists across many frames.","Editorial inference: the reported margin over STFAN is less than 0.5 dB on both datasets, so on harder real videos with large textureless regions or severe occlusions the margin may shrink if the optical-flow consistency check fails.","Editorial inference: a direct test of the paper's central premise is to replace the estimated optical flow with ground-truth flow obtained from a high-speed capture; if the gain over STFAN mostly disappears, then accurate flow is the real driver of the improvement."],"forward_implications":["On both Adobe240 and GOPRO, DAN's average PSNR exceeds OVD, DBN, EDVR, and STFAN: 31.618 dB versus 31.251 dB on Adobe240 and 33.003 dB versus 32.560 dB on GOPRO.","Using the PPN before the ABDN outperforms applying the ABDN twice (31.228 dB versus 30.472 dB), indicating that aligning preprocessed frames matters more than an extra deconvolution pass.","Occlusion maps improve both the ABDN (31.10 dB to 31.23 dB) and the FAN (31.57 dB to 31.62 dB), showing that explicit occlusion handling helps both deblurring and aggregation.","The light version with a 1.05M-parameter ABDN still reaches 31.34 to 31.41 dB on Adobe240, surpassing OVD, DBN, and EDVR and roughly matching STFAN with fewer parameters.","Each module contributes additively in the ablation, so the architecture gives a reusable recipe: rough deblur, align, deblur, then fuse by reliability."],"supporting_citations":[{"why":"Supplies the Adobe240 video deblurring dataset, the training/test split, and the DBN baseline compared in Table 1.","marker":"[8]"},{"why":"Supplies the non-local operation used in the PPN to enhance blurry frames before alignment.","marker":"[17]"},{"why":"Supplies the forward-backward optical flow consistency inequality, Eq. (5), used to detect occlusion.","marker":"[40]"},{"why":"Is the strongest compared baseline (STFAN), against which DAN reports its PSNR gains.","marker":"[16]"},{"why":"Supplies SPyNet, the smaller optical flow estimator used in one variant of the pipeline.","marker":"[38]"},{"why":"Supplies LiteFlowNet, the optical flow estimator whose variant gives slightly better results than SPyNet in Table 2.","marker":"[39]"},{"why":"Provides the patch-based aggregation approach and the real blurry video dataset used for qualitative evaluation.","marker":"[6]"},{"why":"Is the EDVR baseline using deformable convolution, compared in Table 1.","marker":"[9]"},{"why":"Supplies the GOPRO dataset on which DAN reports 33.003 dB.","marker":"[18]"},{"why":"Supplies the U-Net backbone used inside the ABDN for the full version.","marker":"[43]"}],"fun_headline_variants":["Video deblurring with combined deconvolution and aggregation wins both benchmarks","Two-stage deblurring with frame aggregation beats prior art on video benchmarks","Non-local preclearing plus aligned deconvolution and reliability fusion for video deblurri","New video deblurring net: pre-clear, align, deblur, then fuse by sharpness"],"cache_read_input_tokens":13952,"weakest_assumption_plain":"The entire chain depends on the assumption that after the preprocessing network, optical flow between frames is accurate enough to warp neighbors to the central frame, and that the forward-backward consistency test correctly identifies every occluded region; when flow is wrong on textureless or heavily blurred areas, misaligned pixels can be amplified by the reliability-map fusion.","fun_headline_variants_meta":{"raw":{"variants":["Video deblurring with combined deconvolution and aggregation wins both benchmarks","Two-stage deblurring with frame aggregation beats prior art on video benchmarks","Non-local preclearing plus aligned deconvolution and reliability fusion for video deblurring","New video deblurring net: pre-clear, align, deblur, then fuse by sharpness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000348,"raw_usage":{"total_tokens":1907,"prompt_tokens":951,"completion_tokens":956,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":861}},"tokens_in":567,"tokens_out":956,"duration_ms":8402,"temperature":1.0,"reasoning_tokens":861,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:47:57.724681+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full DAN on a video containing large textureless regions and heavy motion blur, then replace the estimated optical flow with ground-truth flow obtained from an ultra-high-speed capture; if the PSNR jumps by far more than the 0.37 to 0.44 dB margin over STFAN, the claimed result is dominated by flow accuracy rather than by the network design.","supporting_citations":[{"cited_title":"Deep video deblurring for hand-held cameras","cited_arxiv_id":null,"evidence_quote":"Supplies the Adobe240 video deblurring dataset, the training/test split, and the DBN baseline compared in Table 1."},{"cited_title":"Unflow: Unsupervised learning of optical flow with a bidirectional census loss","cited_arxiv_id":null,"evidence_quote":"Supplies the forward-backward optical flow consistency inequality, Eq. (5), used to detect occlusion."},{"cited_title":"Spatio-temporal filter adaptive network for video deblurring","cited_arxiv_id":null,"evidence_quote":"Is the strongest compared baseline (STFAN), against which DAN reports its PSNR gains."},{"cited_title":"Optical flow estimation using a spatial pyramid network","cited_arxiv_id":null,"evidence_quote":"Supplies SPyNet, the smaller optical flow estimator used in one variant of the pipeline."},{"cited_title":"Video deblurring for hand-held cameras using patch-based synthesis","cited_arxiv_id":null,"evidence_quote":"Provides the patch-based aggregation approach and the real blurry video dataset used for qualitative evaluation."},{"cited_title":"Edvr: Video restoration with enhanced deformable convolutional networks","cited_arxiv_id":null,"evidence_quote":"Is the EDVR baseline using deformable convolution, compared in Table 1."},{"cited_title":"Deep multi-scale convolutional neural network for dynamic scene deblurring","cited_arxiv_id":null,"evidence_quote":"Supplies the GOPRO dataset on which DAN reports 33.003 dB."}],"review_version":1}