{"id":"53b0b210-8f59-420b-8d5d-1f721dc80151","arxiv_id":"2412.11866","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MTGNet fuses coarse voxel events and fine-grained point cloud events with adaptive diffusion to improve event-based motion deblurring.","lead":"This paper combines two event data formats, voxels and point clouds, in one neural network to sharpen motion-blurred images. On standard event camera benchmarks the method reports better deblurring quality than previous approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gain may come from unmatched density-based cropping; retrain with uniform cropping to verify.","rationale":"I read the paper in good faith. The architectural idea is coherent: combining coarse voxel events with fine-grained point events through coordinate-aware mapping and diffusion is plausible, and the ablation study does show that the point branch and AFDM contribute under the paper's own training protocol. The theoretical motivation in Equations (5)-(9) is heuristic but not the main basis for the claim. The load-bearing issue is the empirical comparison itself. The density-based cropping described in Section IV-A is an explicit, asymmetric training-data selection rule: it is used for MTGNet and not reported for the baselines. Since the central claim is 'outperforms current state-of-the-art,' the comparison must be protocol-matched. If the cropping rule is responsible for a substantial part of the gain, the SOTA conclusion is not established even though the components may still be useful. The paper provides no code and does not specify whether test-time evaluation uses crops or full frames, so the reader cannot independently separate architecture from data selection. The concrete retraining test directly targets this ambiguity. I therefore agree with the reader's weakest assumption and see no reason to change the CONDITIONAL verdict; the concern warrants verification before the claim is accepted.","tokens_in":15880,"tokens_out":5474,"duration_ms":51631,"concrete_test":"Retrain MTGNet from scratch with uniform random 512x512 cropping (same architecture, same b=30, same loss weights, same optimizer schedule), and evaluate under the same test protocol on Ev-REDS and HS-ERGB. If PSNR/SSIM fall to or below the best baselines, or the reported margin shrinks to noise, the density-based cropping is the load-bearing source of the SOTA claim. A complementary check is to train one strong baseline, e.g., STCNet, with the same density-based cropping; if it then matches or exceeds MTGNet, the architecture is not the differentiator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Table I, Section IV-B) is that MTGNet is state of the art on Ev-REDS and HS-ERGB. The comparison is only valid if all methods see comparable training data. Section IV-A describes a 'density-based random cropping strategy' that is applied to MTGNet: training patches are centered only in regions where event density exceeds 80%, with symmetric expansion to 512x512. Baselines are said to be trained under their own protocols, which normally use uniform random crops. This introduces a protocol mismatch. If high-event-density patches are systematically easier, or if uniform crops expose baselines to many sparse hard patches that MTGNet never sees during training, a model trained under the density-based schedule can report higher PSNR/SSIM/LPIPS for reasons unrelated to the proposed architecture. The paper does not report test-time cropping, does not ablate the cropping rule, and releases no code, so the reported gains (e.g., +0.26 dB over STCNet and +0.37 dB over TRMD on Ev-REDS) cannot currently be separated from the data-selection effect. The ablations in Table II keep the cropping rule fixed, so they show component usefulness, not comparison fairness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes MTGNet, a network for event-based motion deblurring that fuses two event representations: a spatially dense, temporally coarse voxel representation and a spatially sparse, temporally fine point-cloud representation. A Coarse Temporal Fusion Branch processes the blurry image and voxel events, while a Fine-grained Point Branch uses an Aggregation and Mapping Module (AMM) to align point-cloud features to the image/voxel feature planes and an Adaptive Feature Diffusion Module (AFDM) to diffuse the sparse point features. The authors provide a Taylor-expansion argument (Eqs. (5)-(9)) that finer temporal sampling reduces the deblurring error, and they report state-of-the-art PSNR/SSIM/LPIPS on Ev-REDS and HS-ERGB, with qualitative results on MS-RBD. Ablations on HS-ERGB support the contribution of the voxel branch, the point-cloud branch with AFDM, and the loss-weight setting.","tokens_in":16123,"tokens_out":13169,"duration_ms":109902,"significance":"If the empirical claims were established under a controlled protocol, this paper would make a valuable contribution to event-based deblurring by demonstrating that point-cloud event representations, previously used mostly for high-level tasks, can be integrated with voxel representations for low-level restoration. The AMM and AFDM modules are reasonable and address genuine challenges of sparsity and resolution mismatch. The paper's analytic derivation, while in need of correction, correctly points toward the benefit of temporally fine event data. The ablation study provides clear evidence for the architectural components. However, the main empirical claim is currently undermined by a training-protocol confound (density-based cropping applied only to the proposed method), and no code is provided. These issues prevent the paper from establishing its central claim in its present form.","major_comments":[{"comment":"The density-based random cropping strategy is applied only to MTGNet, while the comparison methods are said to be trained 'under the optimal parameter settings as specified in the respective papers.' If baselines use uniform cropping, the training distributions differ: MTGNet sees only patches centered in regions with >80% event density, which may be systematically easier than the uniformly cropped patches used for the baselines. This confound undermines the state-of-the-art claim in Table I. The paper should either (i) train all methods with the same cropping strategy, (ii) train MTGNet with uniform random cropping, or (iii) provide an ablation showing that the reported ranking is invariant to the cropping rule. Without one of these, the gains of +0.26 dB over STCNet and +0.37 dB over TRMD on Ev-REDS cannot be reliably attributed to the proposed architecture.","section":"Section IV-A (Implementation details; density-based random cropping) and Section IV-B (Comparisons with…"},{"comment":"The discretization-error analysis is mathematically incorrect. For the left Riemann sum used in Eq. (8), the error between the integral and the sum is O(Δt) in general, not O(Δt^2) as claimed in Eq. (9). Specifically, the Euler-Maclaurin formula gives a leading term proportional to Δt [f(t0)+f(tn)], which is absent from Eq. (9). If the authors intended a trapezoidal or higher-order quadrature rule, that rule must be stated and used consistently. The qualitative conclusion that finer temporal sampling improves deblurring remains plausible, but the derivation as written needs to be corrected or replaced.","section":"Section III-A, Eqs. (8)-(9)"}],"minor_comments":[{"comment":"Reference [15] is cited for both 'DS-Deblur' and 'EIFNet'; these appear to be different methods, so the citation list or the text should be corrected to cite the appropriate paper for DS-Deblur.","section":"Section II-C2 (Related Work)"},{"comment":"The check-mark layout in Table II is ambiguous in the provided format; for example, row 4 could be read as 'Image + Voxel + Cloud' or 'Image + Cloud + AFDM.' Please use explicit column labels or symbols so that each configuration is unambiguous.","section":"Table II (Ablation study)"},{"comment":"The point-cloud representation in Eq. (4) discards polarity by setting z = t. The paper should discuss this design choice and its potential effect on deblurring performance, since polarity encodes the direction of intensity change.","section":"Section III-B (Pre-Processing of the Events)"},{"comment":"The paper does not state whether code or trained models will be released. Given the multi-representation pipeline and the dataset-specific preprocessing, providing code/checkpoints would aid reproducibility.","section":"General"},{"comment":"The figure captions contain the fragment '/searc', which appears to be a typo for '/search'.","section":"Figs. 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"The paper may be salvageable with additional experiments that match the cropping protocol and correct the discretization-error derivation. Please ask the authors to report multiple seeds as well. I would also encourage the editors to require disclosure of code/checkpoints if the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper brings point cloud event representation into event-based deblurring for the first time, fusing it with a voxel branch via coordinate-assisted mapping (AMM) and an adaptive diffusion module (AFDM). That is a genuinely new architectural combination, and the ablation study does a decent job showing that the point cloud branch helps only when AFDM is present. The Taylor expansion motivating finer temporal sampling is hand-wavy but the conclusion is standard and reasonable.\n\nThe quantitative claim is state of the art on Ev-REDS and HS-ERGB, with modest but consistent gains. The main problem is the training protocol. The paper uses a density-based random cropping strategy for MTGNet, where patches are centered only in regions with event density above 80%, while baselines are trained under their own uniform cropping protocols. This is a real mismatch. If high-density patches are systematically easier, the reported margins could come from data selection rather than architecture. The paper does not ablate the cropping rule, report test-time cropping, or release code, so the SOTA claim is not yet verified. The loss-weight tuning on HS-ERGB (Table III) is another leak on that dataset, though the gains on Ev-REDS are less affected.\n\nThe ablations are useful and internally consistent, but they keep the cropping fixed, so they support component usefulness, not comparison fairness. Single-run metrics without variance also make the margins hard to interpret. The writing is readable; some typos but no conceptual confusion.\n\nWho is this for? Event-based vision researchers, especially those interested in alternative event representations or deblurring architectures. It deserves a serious referee, but the referee should ask for a matched-cropping baseline, variance estimates, and ideally code. As it stands, I would treat 'state of the art' as conditional.","headline":"Novel point-cloud/voxel fusion for event deblurring, but the SOTA claim is undercut by an unmatched density-based cropping protocol.","tokens_in":16654,"tokens_out":2371,"would_cite":true,"duration_ms":22005,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MTGNet fuses voxel-grid and point-cloud event representations with an aggregation-mapping module and an adaptive diffusion module, and reports state-of-the-art deblurring results on Ev-REDS, HS-ERGB, and MS-RBD.","keywords":["event camera","motion deblurring","point cloud representation","voxel representation","multi-temporal granularity","adaptive feature diffusion","cross-modal fusion","image restoration"],"falsifier":"Retrain MTGNet under the same uniform random-cropping protocol used for the baselines and re-run the comparison on Ev-REDS and HS-ERGB; if the PSNR, SSIM, and LPIPS leads disappear or reverse, the paper's state-of-the-art claim is not established. A second check: apply the density-based cropping to the baselines as well and see whether MTGNet still leads.","tokens_in":15682,"feed_emoji":"⚡","tokens_out":5745,"duration_ms":43532,"temperature":0.7,"pith_summary":"This paper argues that event-based motion deblurring is held back by how events are represented: voxel grids preserve spatial density but blur the timing of events, while point clouds keep microsecond timestamps but are spatially sparse. It proposes MTGNet, which combines both representations in one network, with two new modules that map sparse point features onto image space and adaptively diffuse them. On the Ev-REDS and HS-ERGB benchmarks the method reports the best PSNR, SSIM, and LPIPS among the compared deblurring methods, and it generalizes to the real-world MS-RBD dataset without adaptation. The reason this matters is that event cameras record motion at microsecond resolution, and a deblurring method that actually uses that fine timing could recover sharper edges and textures in fast-moving scenes.","feed_headline":"Fusing point-cloud and voxel events sets new deblurring records","feed_subtitle":"MTGNet keeps raw microsecond event timestamps, diffuses sparse point features, and tops Ev-REDS and HS-ERGB benchmarks.","key_machinery":"The load-bearing object is the pair of event representations treated as complementary: the voxel grid $V \\in \\mathbb{R}^{h\\times w\\times b}$, which is dense in space but coarse in time, and the point cloud $P \\in \\mathbb{R}^{b\\times m\\times 3}$, which keeps each event's original timestamp as the z-coordinate and is fine in time but sparse in space. The paper motivates the fusion with a Taylor-expansion argument (Equation 9) showing that the error between the estimated sharp image and the true one shrinks as the temporal step $\\Delta t$ decreases. Two modules carry the fusion: AMM performs coordinate-assisted mapping from point features to the image plane, and AFDM adaptively diffuses the sparse mapped features with Gaussian weights whose standard deviation is predicted from the coarse fusion features. The diffusion step acts as feature-level interpolation, which matters when the event sensor has lower resolution than the RGB sensor.","core_discovery":"The central claim is that the temporal granularity of the event representation is a first-order factor in deblurring quality, and that no single representation is sufficient: voxel-based events give dense spatial features but coarse time bins, and point-cloud events give exact timestamps but almost no spatial density. MTGNet therefore fuses both, using a Coarse Temporal Fusion Branch built on a multi-scale U-Net for images and voxel events, and a Fine-grained Point Branch that treats events as points (x, y, t). The Aggregation and Mapping Module groups points by farthest-point sampling and k-nearest neighbors, aggregates each group's features by attention, fuses across time bins with an LSTM, and maps the result back to the image plane using the group centroids. The Adaptive Feature Diffusion Module then estimates a per-position Gaussian diffusion range from the fused coarse features and spreads each sparse point feature over that neighborhood. The paper reports that this combination sets new state-of-the-art numbers on synthetic and real datasets, with the largest gains coming from the diffusion step.","pith_inferences":["The paper's theoretical argument that finer $\\Delta t$ always helps suggests a testable scaling rule: holding the network fixed, deblurring quality should improve monotonically as the number of time bins $b$ increases, until spatial sparsity dominates. The paper does not run this sweep.","The density-based cropping strategy could be masking part of the gain: if the same cropping were applied to the baselines, the reported state-of-the-art margins might shrink or reverse. This is the main risk to the claim.","The point branch's LSTM temporal fusion and the Gaussian diffusion could be reused as a plug-in module for other event-vs-frame fusion tasks such as interpolation, super-resolution, or high-dynamic-range imaging, not just deblurring.","A direct comparison against a voxel-only variant with the same total parameter count and the same cropping protocol would isolate the contribution of the point branch; the paper's ablations vary both the representation and the presence of AFDM together."],"forward_implications":["If MTGNet's reported gains hold, the deblurring community has a new best-practice input: keep raw event timestamps rather than collapsing them into bins, and spend network capacity on mapping sparse points into image space.","The ablation results imply that point-cloud events alone are not enough: without diffusion they barely change performance, so the benefit comes specifically from fusing fine time with dense space and then densifying the point features.","On the real-world MS-RBD dataset, where no ground truth is available, the method shows qualitative generalization when trained only on Ev-REDS, suggesting the representation choice transfers across sensors with different resolutions.","Because the AFDM produces a per-image diffusion map, the same architecture could adapt to variable event density and resolution mismatch without re-training for each sensor pair."],"supporting_citations":[{"why":"Supplies the multi-scale U-Net backbone for the Coarse Temporal Fusion Branch and the image-only baseline MIMO-UNet+.","marker":"[6]"},{"why":"Introduces the Ev-REDS and MS-RBD benchmarks, the GEM baseline, and the evaluation protocol the paper follows.","marker":"[43]"},{"why":"Provides the HS-ERGB dataset of sharp videos with real events used for the main objective evaluation.","marker":"[50]"},{"why":"VID2E is the simulator used to generate synthetic events for Ev-REDS from REDS frames.","marker":"[49]"},{"why":"EVDI is a compared event-based deblurring method and a prior baseline on both datasets.","marker":"[12]"},{"why":"STCNet is the strongest prior baseline on Ev-REDS that MTGNet claims to surpass on SSIM and LPIPS.","marker":"[45]"},{"why":"TRMD is a two-stage residual event-deblurring method used as a comparison point.","marker":"[24]"},{"why":"EIFNet is a cross-modal fusion baseline the paper compares against on both benchmarks.","marker":"[15]"}],"fun_headline_variants":["Point-cloud and voxel events fuse to top deblurring tests","Event time granularity drives best deblurring accuracy","MTGNet blends event grains to sharpen motion blur","Exact event timestamps push deblurring past prior best","Adaptive diffusion of sparse point features boosts deblurring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark claim assumes that training MTGNet only on patches cropped from high-event-density regions does not give it an advantage over comparison methods trained with standard random crops.","fun_headline_variants_meta":{"raw":{"variants":["Point-cloud and voxel events fuse to top deblurring tests","Event time granularity drives best deblurring accuracy","MTGNet blends event grains to sharpen motion blur","Exact event timestamps push deblurring past prior best","Adaptive diffusion of sparse point features boosts deblurring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000331,"raw_usage":{"total_tokens":1868,"prompt_tokens":997,"completion_tokens":871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":785}},"tokens_in":613,"tokens_out":871,"duration_ms":10515,"temperature":1.0,"reasoning_tokens":785,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:29:54.946049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain MTGNet under the same uniform random-cropping protocol used for the baselines and re-run the comparison on Ev-REDS and HS-ERGB; if the PSNR, SSIM, and LPIPS leads disappear or reverse, the paper's state-of-the-art claim is not established. A second check: apply the density-based cropping to the baselines as well and see whether MTGNet still leads.","supporting_citations":[{"cited_title":"Unifying motion deblurring and frame interpo- lation with events,","cited_arxiv_id":null,"evidence_quote":"EVDI is a compared event-based deblurring method and a prior baseline on both datasets."},{"cited_title":"Rethinking coarse-to-fine approach in single image deblurring,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-scale U-Net backbone for the Coarse Temporal Fusion Branch and the image-only baseline MIMO-UNet+."},{"cited_title":"Generalizing event- based motion deblurring in real-world scenarios,","cited_arxiv_id":null,"evidence_quote":"Introduces the Ev-REDS and MS-RBD benchmarks, the GEM baseline, and the evaluation protocol the paper follows."},{"cited_title":"Time lens: Event-based video frame interpolation,","cited_arxiv_id":null,"evidence_quote":"Provides the HS-ERGB dataset of sharp videos with real events used for the main objective evaluation."},{"cited_title":"Video to events: Recycling video datasets for event cameras,","cited_arxiv_id":null,"evidence_quote":"VID2E is the simulator used to generate synthetic events for Ev-REDS from REDS frames."},{"cited_title":"Motion deblurring via spatial- temporal collaboration of frames and events,","cited_arxiv_id":null,"evidence_quote":"STCNet is the strongest prior baseline on Ev-REDS that MTGNet claims to surpass on SSIM and LPIPS."},{"cited_title":"Motion deblur by learning residual from events,","cited_arxiv_id":null,"evidence_quote":"TRMD is a two-stage residual event-deblurring method used as a comparison point."},{"cited_title":"Event-based motion deblurring with modality-aware decomposition and recomposition,","cited_arxiv_id":null,"evidence_quote":"EIFNet is a cross-modal fusion baseline the paper compares against on both benchmarks."}],"review_version":1}