{"id":"19767b06-9322-4921-a6a1-a16d6bfe4ac4","arxiv_id":"2412.19370","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"From one blurred image and an event stream, BeSplat jointly learns 3D Gaussian splats and an SE(3) Bezier camera trajectory, yielding sharp novel views at high speed.","lead":"BeSplat builds a sharp 3D scene model from one motion-blurred photo plus a stream of event-camera brightness changes, and it estimates the camera path during the exposure. If it works as claimed, it offers fast deblurring and view synthesis for AR, robotics, and photography.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BeSplat's reported results rely on COLMAP initialization from ground-truth sharp images, so the headline claim of recovery from only a blurry image and event stream is not demonstrated.","rationale":"The reader's weakest_assumption identified exactly the same issue: the pipeline uses COLMAP on corresponding sharp images for initialization, which contradicts the stated input of only a blurry image plus events. This is the single most load-bearing concern because the paper's central claim is about what inputs are sufficient for the recovery. The concern is directly evidenced by the paper's own Section 5.1. It is not a matter of external consensus or style; it is an internal inconsistency between the claimed problem setting and the implemented method. A concrete ablation replacing the privileged initialization with an unprivileged one would settle the matter. The reader's REJECT verdict is appropriate, and I find no reason to change it. I set verdict_should_be to UNCHANGED because my stress-test does not alter the reader's conclusion.","tokens_in":12149,"tokens_out":3036,"duration_ms":28387,"concrete_test":"Rerun the full BeSplat pipeline on the synthetic Livingroom sequence with identical losses and hyperparameters, but initialize Gaussian points and control knots in two ways: (a) a random point cloud within a bounding box and an identity/constant trajectory, and (b) COLMAP applied only to the input blurry image (if it produces a reconstruction at all). Report PSNR, SSIM, LPIPS, and the error of the estimated trajectory against the ground-truth trajectory. If quality degrades substantially relative to Table 1, or if the optimization diverges, the reported results depend on the privileged sharp-image initialization and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 (Implementation Details) states: 'To obtain the initial sparse point cloud and camera pose for both real and synthetic datasets, structure-from-motion (COLMAP) is applied to the corresponding sharp images, following the approach in [8].' This is the load-bearing defect. The abstract and introduction claim the method recovers a sharp Gaussian-splat scene and the camera trajectory from a single motion-blurred image and its corresponding event stream. However, COLMAP on the corresponding sharp images supplies the initial 3D point cloud and camera poses, which are exactly the geometric quantities the task claims to recover. The method therefore has access to privileged information that the stated input protocol excludes. No experiment or ablation is provided showing that the optimization converges to comparable quality without this sharp-image initialization, and no code is released to test it independently. Because the central claim is about the input modality, this is not a minor implementation detail; it invalidates the paper's headline claim as written. Secondary issues (e.g., trajectory accuracy never quantitatively measured) are also present, but the initialization assumption is the most load-bearing concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents BeSplat, a 3D Gaussian Splatting method that aims to recover a sharp radiance field and the camera motion trajectory from a single motion-blurred image and its corresponding event stream. The method models the camera trajectory with a learnable SE(3) spline, renders virtual sharp views with 3DGS, averages them to synthesize the blurry image via a motion-blur formation model, and adds an event consistency loss. Experiments on synthetic BeNeRF and real E2NeRF datasets report competitive deblurring PSNR/SSIM/LPIPS compared with BeNeRF and learning-based deblurring baselines, with roughly 4x faster training and 6x lower GPU memory usage.","tokens_in":12390,"tokens_out":7660,"duration_ms":64861,"significance":"If fully supported, BeSplat would be a meaningful contribution as the first 3DGS-based method for single-image motion deblurring with event streams, and the efficiency gains over NeRF-based BeNeRF are substantial. The qualitative results on real data and the explicit event loss are strengths. However, the central claim is conditional on the use of COLMAP initialization from sharp images, which is outside the stated input protocol, and the trajectory recovery is never quantitatively validated. These issues must be resolved before the contribution can be fairly assessed.","major_comments":[{"comment":"The implementation initializes the sparse point cloud and camera poses by running COLMAP on the corresponding sharp images. This supplies the 3D geometry and camera trajectory initialization that the method claims to recover, so the experimental protocol does not match the stated problem of a single blurry image plus event stream. No ablation is provided to show that the optimization converges to comparable quality without this initialization. This is a load-bearing discrepancy for the abstract's claim and must be addressed with either a revision of the problem statement or experiments that do not use sharp-image COLMAP.","section":"Section 5.1, Implementation Details"},{"comment":"The abstract and introduction claim accurate recovery of the camera motion trajectory, yet the paper reports no trajectory error metrics (e.g., ATE, RPE, RMSE) anywhere. Tables 1-3 contain only image reconstruction metrics and efficiency numbers. Since the trajectory is a core output of the method, and a wrong trajectory could still fit the blurry input if the Gaussians overfit, the trajectory-recovery claim is currently unsupported.","section":"Section 5.3, Tables 1-3"},{"comment":"The text describes the trajectory as a 7th-order Bézier curve, but Eq. (7) gives the 4x4 basis matrix from the De Boor-Cox formula for a cubic B-spline, and Eq. (8) uses four consecutive control knots. This internal inconsistency means the implemented trajectory model is not the one described. The authors should clarify the actual spline order and update the text, or correct the equations; this matters for the trajectory ablation in Table 4.","section":"Section 3.2, Eq. (7)"},{"comment":"The paper claims novel view synthesis, but all quantitative evaluations are performed on sharp frames along the same trajectory that is fitted to the blurry image and events. No held-out camera poses are used for evaluation, so the reported PSNR/SSIM/LPIPS measure the model's fit to training views rather than its ability to synthesize novel views. The paper should either evaluate on held-out views or qualify the claim as deblurring along the estimated trajectory.","section":"Sections 5.2-5.3, Fig. 1"}],"minor_comments":[{"comment":"Table 1 reports quantitative results for only two of the five synthetic sequences (Livingroom and Tanabata), while Fig. 3 shows qualitative results for all five; results for Outdoorpool, Pinkcastle, and Whiteroom are missing.","section":"Table 1"},{"comment":"The memory and training time comparisons report deterministic values without standard deviations or multiple runs, and the efficiency comparison is only against BeNeRF, not against other 3DGS-based deblurring methods.","section":"Tables 2 and 3"},{"comment":"The implementation says 19 sample points are used along the trajectory, while the ablation in Fig. 5 varies the virtual camera count n as 7, 11, and 15; the relationship between the sample count and the virtual camera count n in Eq. (9) is not defined.","section":"Section 5.1 and Fig. 5"},{"comment":"The event formation and loss in Eqs. (10)-(15) normalize both the captured and synthesized event images, which removes the absolute contrast scale; the paper does not discuss how this normalization interacts with the contrast threshold C in the event generation model.","section":"Section 4.2"}],"recommendation":"reject","confidential_remarks":"The COLMAP-on-sharp-images issue is not a minor implementation detail: it changes the task from single-input recovery to an initialization-assisted fit. If the authors can demonstrate success without sharp-image COLMAP and add trajectory metrics, the paper might be resubmitted; as it stands, the manuscript does not support the main claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: BeSplat is a legitimate engineering contribution. The novelty is limited – it transplants BeNeRF's event loss and Deblur-GS's trajectory model into 3DGS – but that integration is useful, and the efficiency numbers look real: roughly 4x faster training and 6x lower memory than BeNeRF, at a modest quality cost (e.g., 35.14 vs 37.11 PSNR on Livingroom). The ablations on virtual camera count and trajectory choice are a plus, and the real-data results are a sensible check.\n\nThe soft spots are real, and one is load-bearing. Section 5.1 says COLMAP is applied to the \"corresponding sharp images\" for both synthetic and real datasets. That supplies the initial sparse point cloud and camera pose. The paper's central claim is recovery from a single blurry image plus events; using sharp images for initialization is privileged information that the problem statement excludes. No ablation or experiment shows the method still works without it. This is not a minor implementation detail; it changes what the paper demonstrates. Also, the trajectory description is inconsistent: the text calls it a 7th-order Bezier while Eq. (7) is the cubic B-spline basis (De Boor-Cox). The ablation compares against \"cubic B-Spline\" but their own model looks like the same thing with a different name. Finally, trajectory accuracy is asserted but never measured – no ATE or RMSE anywhere. That matters because the event loss is supposed to constrain motion.\n\nOn the positive side, the paper is honest about the small quality gap to BeNeRF, positions itself as an efficiency alternative, and doesn't overclaim beyond the initialization issue – except that the initialization issue is exactly what undercuts the task framing.\n\nWho this is for: people working on event-based 3D reconstruction, or on making deblurring practical for robotics and AR. The paper deserves a serious referee: it's a plausible combination with decent experiments and a real efficiency story. But as written, the headline claim is not demonstrated. The COLMAP initialization must be addressed, either by removing the sharp-image dependency or by reframing the contribution as \"deblurring with a known pose prior.\" My recommendation: reject in current form, but encourage a revision that either eliminates the sharp-image initialization or explicitly treats it as an input, and adds trajectory evaluation metrics.","headline":"An efficient 3DGS port of BeNeRF's event-plus-blur loss, with a real efficiency win, but the single-blurry-image claim is undercut by sharp-image COLMAP initialization.","tokens_in":12942,"tokens_out":2466,"would_cite":false,"duration_ms":21909,"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":"BeSplat reconstructs a sharp, renderable 3D scene from a single motion-blurred image plus its event stream, jointly estimating the camera trajectory during exposure.","keywords":["Gaussian Splatting","motion deblurring","event camera","novel view synthesis","camera trajectory estimation","Bézier curve","radiance field","single image deblurring"],"falsifier":"Run the identical optimization on the BeNeRF synthetic sequences with the Gaussian point cloud initialized from random depth in the camera frustum and the trajectory initialized to the identity pose; if the resulting PSNR falls below the learning-based baselines in Table 1, the claim that a single blurry image plus events alone recovers the sharp scene is not supported.","tokens_in":11962,"feed_emoji":"🎥","tokens_out":7229,"duration_ms":57631,"temperature":0.7,"pith_summary":"This paper seeks to show that a sharp 3D scene representation can be recovered from inputs as limited as one motion-blurred photograph and the event stream recorded during that exposure. The method, BeSplat, represents the scene as 3D Gaussian splats rather than a neural radiance field, and it models the camera motion during the exposure as a differentiable Bézier curve in SE(3). By minimizing a photometric loss against the blurred image and an event loss against the accumulated events, the optimization recovers both the latent sharp scene and the camera trajectory. If the claim holds, it turns a single blurry frame plus events into enough information for real-time novel-view synthesis, deblurring, and motion estimation in one pass.","feed_headline":"A blurry frame plus an event stream yields a sharp 3D scene","feed_subtitle":"BeSplat estimates the camera trajectory and renders novel sharp views, training about four times faster than its NeRF-based predecessor.","key_machinery":"The load-bearing mechanism is the differentiable combination of 3D Gaussian Splatting with a Bézier-curve camera trajectory in the special Euclidean group SE(3). The trajectory is defined by learnable control knots; at each sampled timestamp four control knots are combined through the De Boor–Cox basis matrix to yield an interpolated pose. Virtual sharp images rendered from the Gaussians at those poses are averaged to synthesize the blurred image, and the log-intensity difference between two trajectory endpoints synthesizes the accumulated event image. Both synthesized measurements are compared with the real measurements, giving gradients that flow back through the rasterizer to update both the Gaussian parameters and the control knots.","core_discovery":"The authors claim to be the first to solve the single-blurry-image-plus-event-stream deblurring and novel-view-synthesis problem inside the 3D Gaussian Splatting framework. Their central discovery is that jointly optimizing Gaussian parameters and a 7th-order Bézier trajectory in SE(3), with 19 sampled poses during the exposure, lets the blurry-image formation model and a normalized accumulated-event loss pull out a sharp radiance field and the camera motion that produced the blur. On the BeNeRF synthetic dataset and the E2NeRF real dataset, the recovered splats render view-consistent sharp images along the trajectory, with training about four times faster and memory consumption about six times lower than the NeRF-based BeNeRF baseline.","pith_inferences":["Inference: Because the initialization relies on COLMAP run on sharp images, an immediate unproven extension is the fully constraint-satisfying setting where only a blurry frame and events are available; event-only or blur-only initialization is a natural follow-up test.","Inference: The event-loss formulation is modular and could be attached to other explicit scene representations, which would separate the contribution of trajectory recovery from the contribution of the Gaussian parameterization.","Inference: The recovered Bézier poses could be directly compared against inertial or visual-odometry ground truth in a real capture, offering a quantitative check on whether the trajectory recovered from blur-plus-events is genuinely the physical camera motion."],"forward_implications":["The recovered scene is an explicit Gaussian point cloud, so sharp novel views can be rendered at real-time rates rather than through slow ray marching.","Training on the real dataset takes roughly one hour thirty minutes versus about six hours for BeNeRF, and GPU memory drops from 8.65 GB to 1.45 GB, making the method feasible on a single consumer GPU.","The same optimization outputs both a deblurred sequence of sharp images along the trajectory and an estimate of the camera motion that caused the blur, which could serve as a motion cue for downstream robotics or AR tasks.","The ablation shows a 7th-order Bézier trajectory reaches quality comparable to cubic B-splines while taking about one third of the training effort, so trajectory parameterization is a major efficiency lever."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting scene representation and differentiable rasterizer that BeSplat optimizes.","marker":"[22]"},{"why":"Provides the differentiable Bézier SE(3) trajectory model, the virtual-camera rendering schedule, and the pose-gradient simplification adopted for camera-motion estimation.","marker":"[8]"},{"why":"Defines the task of recovering a sharp radiance field from a single blurry image and event stream, supplies the event-loss formulation and the synthetic dataset, and is the main method being compared.","marker":"[29]"},{"why":"Provides the COLMAP structure-from-motion initialization of Gaussian points and camera poses from the corresponding sharp images.","marker":"[48]"},{"why":"Generates the synthetic event streams with the ESIM simulator for the synthetic evaluation sequences.","marker":"[44]"},{"why":"Supplies the real-world event-plus-blurry-image dataset used for the real-data experiments.","marker":"[42]"}],"fun_headline_variants":["Sharp 3D from one blurry frame and an event stream","BeSplat: single blurry image and events yield sharp 3D","First Gaussian Splatting for blurry image plus event data","Blur and events become sharp 3D via Gaussian splatting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline starts from a COLMAP reconstruction computed on sharp images of the same scene, so the actual inputs are not limited to one blurry image and its events.","fun_headline_variants_meta":{"raw":{"variants":["Sharp 3D from one blurry frame and an event stream","BeSplat: single blurry image and events yield sharp 3D","First Gaussian Splatting for blurry image plus event data","Blur and events become sharp 3D via Gaussian splatting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1453,"prompt_tokens":897,"completion_tokens":556,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":479}},"tokens_in":513,"tokens_out":556,"duration_ms":5839,"temperature":1.0,"reasoning_tokens":479,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:39:38.075623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical optimization on the BeNeRF synthetic sequences with the Gaussian point cloud initialized from random depth in the camera frustum and the trajectory initialized to the identity pose; if the resulting PSNR falls below the learning-based baselines in Table 1, the claim that a single blurry image plus events alone recovers the sharp scene is not supported.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting scene representation and differentiable rasterizer that BeSplat optimizes."},{"cited_title":"Deblur-gs: 3d gaussian splatting from camera motion blurred images","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable Bézier SE(3) trajectory model, the virtual-camera rendering schedule, and the pose-gradient simplification adopted for camera-motion estimation."},{"cited_title":"Benerf: neural radiance fields from a single blurry image and event stream","cited_arxiv_id":null,"evidence_quote":"Defines the task of recovering a sharp radiance field from a single blurry image and event stream, supplies the event-loss formulation and the synthetic dataset, and is the main method being compared."},{"cited_title":"Structure-from-motion revisited","cited_arxiv_id":null,"evidence_quote":"Provides the COLMAP structure-from-motion initialization of Gaussian points and camera poses from the corresponding sharp images."},{"cited_title":"ESIM: An Open Event Camera Simulator","cited_arxiv_id":null,"evidence_quote":"Generates the synthetic event streams with the ESIM simulator for the synthetic evaluation sequences."},{"cited_title":"E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world event-plus-blurry-image dataset used for the real-data experiments."}],"review_version":1}