{"id":"cc1164c2-43ef-4787-b14f-a7a56ba11dbb","arxiv_id":"2501.15808","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"BDHNet, a bioinspired hybrid SNN-ANN network with neuron configuration and blur-region masking, achieves state-of-the-art event-based motion deblurring on GoPro, REBlur, and MS-RBD.","lead":"A hybrid neural network that combines spiking and artificial neurons uses image features to tune spike generation and a learned blur mask to focus attention, improving event-based motion deblurring. The method reports state-of-the-art PSNR on GoPro, REBlur, and MS-RBD, but code and error bars are not provided.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world SOTA claim is under-supported: MS-RBD has no quantitative comparison, and the only real-world table margin over MAENet on REBlur is 0.04 dB with no error bars.","rationale":"The reader's weakest assumption was that the comparison with baselines is fair and that reported differences are meaningful, specifically noting the 0.04 dB margin over MAENet on REBlur. I agree with that assessment, but I see an even more direct evidentiary problem: the paper explicitly claims SOTA on MS-RBD in the contributions and conclusion, yet no quantitative MS-RBD results appear anywhere in the manuscript. This means the real-world SOTA claim is not fully specified in the paper's own data. The GoPro result is more substantial, but the abstract and conclusion emphasize both synthetic and real-world performance, so a SOTA claim that depends on a single-run 0.04 dB real-world margin is fragile. The proposed test—repeated-seed evaluation on REBlur plus a full MS-RBD comparison table—would settle whether the concern lands. If the margin disappears or MS-RBD numbers are not best, the claim reduces to competitive rather than state-of-the-art. If the margin persists and MS-RBD numbers are provided and favorable, the conditional acceptance is justified. This does not impugn the authors' integrity; it is a request for evidence commensurate with the strength of the claim. The conditional verdict already given by the reader is appropriate, so I do not recommend changing it.","tokens_in":11411,"tokens_out":5445,"duration_ms":51633,"concrete_test":"Run BDHNet and the strongest event-based baselines (at least MAENet, EIFNet, STCNet) on REBlur with 3 random seeds under the protocol of Section 4.2, reporting mean and standard deviation of PSNR/SSIM. In the same submission, add a full quantitative MS-RBD comparison table for all compared methods using an identical training/evaluation protocol. If the 0.04 dB REBlur margin does not persist across seeds, or if BDHNet is not ranked first on MS-RBD, the state-of-the-art claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BDHNet achieves state-of-the-art event-based motion deblurring on synthetic and real-world datasets. The GoPro result (37.04 vs 36.45 for STCNet) is a meaningful margin, but the real-world portion of the claim rests on two weak pillars. First, the contribution list and conclusion state SOTA on MS-RBD, yet Section 4 reports no quantitative comparison on MS-RBD: there is no PSNR/SSIM table for BDHNet or any baseline on that dataset, only qualitative figures (Fig. 5 and Fig. 6). Second, on REBlur, the only real-world dataset with numbers, the fine-tuned PSNR is 38.50 versus MAENet's 38.46, a 0.04 dB gap. The paper reports a single training run and provides no standard deviation, confidence interval, or repeated-seed results. In image restoration, seed-level variation is typically at least an order of magnitude larger than 0.04 dB. Even the no-fine-tune REBlur margin (36.01 vs 35.75 for EIFNet) is only 0.26 dB, which is still within plausible run-to-run noise. Thus the manuscript's own data do not establish that BDHNet is state-of-the-art in real-world conditions. This is an evidentiary gap between the stated claim and the reported measurements, not a dispute over consensus. If the 0.04 dB gap disappears under repeated runs, or if an MS-RBD quantitative comparison places BDHNet below a baseline, the claim should be downgraded from 'significantly outperforms' to 'competitive.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces BDHNet, a hybrid ANN-SNN encoder-decoder for event-based motion deblurring. It proposes a Neuron Configurator Module (NCM) that uses image features to set per-pixel membrane potentials and thresholds in SNN blocks, and a Region of Blurry Attention Module (RBAM) that generates an unsupervised blurry-region mask from aggregated spike features and image-based threshold maps, then uses the mask to gate cross-modal attention. Experiments compare against image-only and event-based methods on GoPro, REBlur, and MS-RBD, with ablations on GoPro. The paper claims state-of-the-art performance on all three datasets, with particularly strong claims for real-world generalization.","tokens_in":11759,"tokens_out":4658,"duration_ms":39668,"significance":"If the results are reproducible, the core architectural ideas are plausible and useful: replacing global LIF thresholds with image-conditioned per-pixel configurations and using an event-derived blur mask to gate fusion directly target the non-uniform event distribution problem. The GoPro results show a non-trivial margin (37.04 vs 36.45 PSNR over STCNet), and the GoPro ablations attribute consistent gains to NCM (+0.22 dB) and RBAM (+0.33-0.44 dB). The paper does not ship code or checkpoints, but the derivation in Section 3.1 is correct and the ablations are internally consistent. However, the real-world SOTA claim currently rests on a 0.04 dB REBlur margin and on qualitative MS-RBD evidence only; the manuscript would be materially strengthened by multi-seed statistics and by quantitative MS-RBD evaluation.","major_comments":[{"comment":"The contribution list and the conclusion state that BDHNet achieves SOTA on MS-RBD, but Section 4 provides no quantitative comparison on MS-RBD: Table 1 reports only GoPro and REBlur, and MS-RBD appears only in qualitative Figures 5 and 6. Either add a PSNR/SSIM table for MS-RBD with all compared baselines, or revise the SOTA claim to cover only the datasets with quantitative results.","section":"§4.3, Table 1, Contributions, Conclusion"},{"comment":"On fine-tuned REBlur, BDHNet's PSNR is 38.50 versus MAENet's 38.46, a 0.04 dB difference, and the paper reports no standard deviation, confidence interval, or repeated-seed runs. Since seed-level variation in image restoration is typically larger than 0.04 dB, this margin does not by itself establish superiority in real-world conditions. The same concern applies to the no-fine-tune REBlur margin (36.01 vs 35.75 for EIFNet). Please provide multi-seed mean and standard deviation (or equivalent statistics) and avoid the phrase 'significantly outperforms' unless a statistical test supports it.","section":"§4.3, Table 1"},{"comment":"The baseline evaluation protocol is under-specified. The text says methods without open-source checkpoints were trained 'under the optimal parameter settings as specified in the respective papers,' but it does not document the event voxel representation, number of temporal bins, optimizer, learning rate schedule, epochs, crop size, or random seeds used for those baselines on REBlur. Without this information or released code/checkpoints, the fairness of the 0.04 dB comparison cannot be independently verified.","section":"§4.3"}],"minor_comments":[{"comment":"There are several grammatical slips, for example 'dynamically adjusts' should be 'dynamically adjust' and 'has achieved SOTA in varying blurry conditions' is not a complete construction; a careful proofread is needed.","section":"Abstract and Introduction"},{"comment":"In Equation (14), the spike feature S is written as S(x, y, t) after the SNN output was denoted S; please clarify the spatial and temporal indexing and the exact tensor shape that is summed over time.","section":"§3.2.2"},{"comment":"The text mentions a 'PSNR Loss' with citation [8], but PSNR is normally used as a metric rather than a training loss; please specify the actual loss function (for example L1 or MSE) and how PSNR enters the training objective.","section":"§4.2"},{"comment":"Figure 7 is described as showing training loss under different neuron configurations, but the curves are not labeled in a way that lets the reader identify which configuration corresponds to which curve; please add a legend or a text description of each curve.","section":"Figure 7"},{"comment":"The sentence 'all event-based methods are all based on the raw event data produced by EFNet' is ambiguous; it should state explicitly which event representation and preprocessing each baseline and the proposed method use.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's claimed real-world SOTA is the main risk. I do not see a circularity problem: the design and ablations are consistent with the reported GoPro margin, and the modules are not derived from the target metric. The manuscript would be acceptable after the authors either provide MS-RBD metrics and multi-seed statistics or temper the superiority claims. The lack of code and detailed baseline protocol amplifies the fairness concern, though it is not by itself grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: BDHNet is a genuine, well-engineered architecture and the GoPro result is a real step forward, but the paper overclaims SOTA on real-world data. The only REBlur margin is 0.04 dB and MS-RBD never appears in a quantitative table.\n\nWhat's actually new: the Neuron Configurator Module (NCM) initializes membrane potential and threshold from image features, and the Region of Blurry Attention Module (RBAM) produces an unsupervised blurry mask that gates cross-modal attention. These are sensible extensions of known hybrid SNN/ANN ideas, not a conceptual breakthrough. The ablations show each module adds a consistent 0.2–0.4 dB on GoPro, and the full model reaches 37.04 PSNR versus 36.45 for STCNet. That is a meaningful, if modest, improvement on the synthetic benchmark.\n\nThe soft spots are evidentiary. The contribution list and conclusion claim SOTA on MS-RBD, but Table 1 only reports GoPro and REBlur; MS-RBD appears only in qualitative figures. That is an unsupported claim in the paper's own text. On REBlur fine-tuned, BDHNet beats MAENet by 0.04 dB (38.50 vs 38.46). Without error bars or repeated runs, that difference is indistinguishable from training noise. The no-fine-tune gap (36.01 vs 35.75) is 0.26 dB, which could also be seed variation. The baseline protocol says methods without open-source checkpoints were trained \"under the optimal parameter settings as specified in the respective papers,\" but the paper omits the actual training protocol, event preprocessing details, and seeds. No code, no error bars, no MS-RBD numbers: the real-world SOTA claim is not reproducible from the manuscript.\n\nNone of this invalidates the GoPro result or the architecture. The math in Section 3.1 is fine, the ablations are informative, and the citations look fair. The issues are missing evidence, not misrepresentation.\n\nWho this is for: people working on event-based deblurring or hybrid SNN/ANN image restoration. It deserves a serious referee, mainly to demand the MS-RBD quantitative comparison, error bars or repeated seeds, and a documented baseline protocol. I would send it out rather than desk-reject, with the expectation that the \"significantly outperforms\" framing will need to be tempered to \"competitive\" on real-world data unless the missing evidence arrives.","headline":"A solid hybrid SNN/ANN deblurrer with a real GoPro gain, but the real-world SOTA claim rests on a 0.04 dB margin and a missing MS-RBD table.","tokens_in":12281,"tokens_out":2043,"would_cite":true,"duration_ms":19144,"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 proposes BDHNet, a dual-drive ANN-SNN network that reaches state-of-the-art event-based motion deblurring by reconfiguring spiking neurons and masking blurry regions before cross-modal attention.","keywords":["event-based motion deblurring","spiking neural networks","hybrid ANN-SNN","cross-modal attention","visual attention mechanism","neuron configuration","blurry region mask","event cameras"],"falsifier":"Run the identical training configuration on GoPro with BDHNet and the top baselines (EFNet, EIFNet, STCNet, MAENet) across several random seeds and report mean and standard deviation of PSNR; if BDHNet's advantage over MAENet and STCNet shrinks below its 0.2 to 0.6 dB margins or reverses, the state-of-the-art claim would not survive. An additional event-sparsity sweep that removes NCM should show a growing performance gap as events become sparser; if the gap stays flat, the module is not doing the claimed work.","tokens_in":11221,"feed_emoji":"⚡","tokens_out":5806,"duration_ms":48928,"temperature":0.7,"pith_summary":"This paper tries to show that event-based motion deblurring can be improved by borrowing two attention mechanisms from human vision: adjusting how individual spiking neurons fire based on image content, and generating a mask that highlights blurry regions to guide cross-modal fusion. The proposed BDHNet combines an ANN image branch with an SNN event branch, and on the GoPro, REBlur, and MS-RBD benchmarks it reports the best PSNR/SSIM among compared event-based methods, including 37.04/0.977 on GoPro and 38.50/0.978 on REBlur. If these results hold, hybrid ANN-SNN architectures become a stronger option for real-world deblurring where events are sparse or unevenly distributed. The modules also make the network generalize to a new dataset without fine-tuning better than the compared methods.","feed_headline":"Bioinspired hybrid network tops event-based deblurring benchmarks","feed_subtitle":"Neuron-level configuration and an unsupervised blurry mask lift GoPro PSNR to 37.04, ahead of prior event-based methods.","key_machinery":"The central mechanism is a dual-drive enhancement inserted after each encoder layer. The Neuron Configurator Module (NCM) computes $V_{\\mathrm{init}} = \\varphi_{\\mathrm{init}}(B) + \\psi_{\\mathrm{init}}(E)$ and sets the LIF firing threshold to $V'_{\\mathrm{th}} = 1 - \\sigma(V_{\\mathrm{init}})$, replacing the vanilla scalar threshold with a per-pixel tensor that persists across timesteps. The Region of Blurry Attention Module (RBAM) sums spikes over time, applies a deformable convolution whose biases are predicted from image features, min-max normalizes the result into $S_{\\mathrm{map}}$, binarizes it against an image-derived threshold map $T_{\\mathrm{map}}$, and uses the resulting mask $M$ to gate multi-head cross-attention between image features and event features. These two mechanisms carry the paper's claim that sparse-event blurry regions can be made to fire and that the resulting motion clues can be directed to the right spatial locations during fusion.","core_discovery":"The core claim is that the main obstacle in event-based motion deblurring is not the capacity of the network but the mismatch between non-uniform, redundant event data and uniform neuron and synaptic processing. BDHNet addresses this with two complementary mechanisms: the Neuron Configurator Module uses image features to set each neuron's initial membrane potential and threshold per pixel and per timestep, effectively lowering the firing barrier in blurry regions where events are sparse; the Region of Blurry Attention Module aggregates spike features with a deformable filter, compares the normalized spike map to an image-derived threshold map to produce an unsupervised blurry mask, and uses that mask to gate cross-modal attention between image and event features. With these mechanisms, the model reports state-of-the-art results on three benchmarks and shows stronger cross-dataset generalization without fine-tuning.","pith_inferences":["Because NCM only changes how neuron states are initialized and thresholds are set, the same mechanism could be dropped into other SNN-based restoration or recognition pipelines where event density is uneven, though the paper does not test this transfer.","The unsupervised blurry mask produced by RBAM could serve as a motion or blur localization signal for other tasks such as blur estimation or segmentation, but the paper does not pursue those applications.","The reported margin over MAENet on REBlur is only 0.04 dB; without repeated-seed error bars, part of the claimed advantage may be training noise, so a fair test would include statistical replicates.","A direct stress test of the two modules would artificially sparsify event streams: if NCM is doing the claimed work, its benefit should grow as events become sparser, but the paper does not report such a sweep."],"forward_implications":["On GoPro, the full model reaches 37.04 PSNR and 0.977 SSIM, ahead of the best compared event-based methods, STCNet at 36.45 and MAENet at 36.07.","On REBlur with fine-tuning it reaches 38.50 PSNR and 0.978 SSIM, and without fine-tuning it still reaches 36.01 PSNR and 0.967 SSIM, the best among event-based methods in that setting.","The pixel-level threshold configuration in NCM lets neurons fire even where event density is low, so motion cues in blurry, low-contrast regions are not lost.","The mask-guided cross-modal attention in RBAM improves performance by 0.44 dB over using plain addition for fusion, according to the ablation study.","The hybrid design preserves temporal event structure through the SNN branch while the ANN branch retains color and structure details needed for pixel-level restoration."],"supporting_citations":[{"why":"Supplies the MIMO multi-scale encoder-decoder backbone used for the image branch and final reconstruction.","marker":"[9]"},{"why":"Provides the GoPro and REBlur benchmarks, the EFNet cross-modal attention baseline, and the event preprocessing standard.","marker":"[27]"},{"why":"Defines the MotionSNN hybrid SNN-ANN deblurring approach whose uniform neuron configuration BDHNet improves on.","marker":"[23]"},{"why":"Gives the iterative Leaky Integrate-and-Fire neuron equations that NCM reconfigures.","marker":"[33]"},{"why":"Source of the voxel-based event representation with $b=12$ timesteps used in this work.","marker":"[34]"},{"why":"A prior approach that initializes only the first-timestep membrane potential, which NCM extends to full threshold configuration.","marker":"[1]"}],"fun_headline_variants":["Vision-inspired neuron config and blur mask beat prior event-deblur","Event deblurring: adaptive neuron thresholds and unsupervised mask lift SOTA","Adaptive neuron firing and self-taught blur mask set new deblur record","Bioinspired neuron config and mask top event-deblurring leaderboards"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim rests on the assumption that the baseline comparisons are fair, meaning the reported differences in PSNR and SSIM come from the architecture rather than from differences in training protocol, event preprocessing, or random seeds.","fun_headline_variants_meta":{"raw":{"variants":["Vision-inspired neuron config and blur mask beat prior event-deblur","Event deblurring: adaptive neuron thresholds and unsupervised mask lift SOTA","Adaptive neuron firing and self-taught blur mask set new deblur record","Bioinspired neuron config and mask top event-deblurring leaderboards"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3921,"prompt_tokens":921,"completion_tokens":3000,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2918}},"tokens_in":537,"tokens_out":3000,"duration_ms":20392,"temperature":1.0,"reasoning_tokens":2918,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:55:29.792256+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical training configuration on GoPro with BDHNet and the top baselines (EFNet, EIFNet, STCNet, MAENet) across several random seeds and report mean and standard deviation of PSNR; if BDHNet's advantage over MAENet and STCNet shrinks below its 0.2 to 0.6 dB margins or reverses, the state-of-the-art claim would not survive. An additional event-sparsity sweep that removes NCM should show a growing performance gap as events become sparser; if the gap stays flat, the module is not doing the claimed work.","supporting_citations":[{"cited_title":"Rethinking coarse-to-fine approach in sin- gle image deblurring","cited_arxiv_id":null,"evidence_quote":"Supplies the MIMO multi-scale encoder-decoder backbone used for the image branch and final reconstruction."},{"cited_title":"Motion-oriented hy- brid spiking neural networks for event-based motion deblur- ring.IEEE Transactions on Circuits and Systems for Video Technology, 2023","cited_arxiv_id":null,"evidence_quote":"Defines the MotionSNN hybrid SNN-ANN deblurring approach whose uniform neuron configuration BDHNet improves on."},{"cited_title":"Spatio-temporal backpropagation for training high- performance spiking neural networks.Frontiers in neuro- science, 12:331, 2018","cited_arxiv_id":null,"evidence_quote":"Gives the iterative Leaky Integrate-and-Fire neuron equations that NCM reconfigures."},{"cited_title":"Event-based motion deblurring with modality-aware decomposition and recomposition","cited_arxiv_id":null,"evidence_quote":"Source of the voxel-based event representation with $b=12$ timesteps used in this work."},{"cited_title":"A hybrid ann-snn architecture for low- power and low-latency visual perception","cited_arxiv_id":null,"evidence_quote":"A prior approach that initializes only the first-timestep membrane potential, which NCM extends to full threshold configuration."}],"review_version":1}