{"id":"74d1f9a3-c2e2-48a7-8ac0-b1d587e42b74","arxiv_id":"2505.15287","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A pipeline that turns sparse multi-view RGB images into a claimed 1,150-scene synthetic event dataset using 3D Gaussian Splatting rendering plus a stochastic event simulator.","lead":"GS2E builds a large synthetic event-camera dataset by reconstructing real scenes with 3D Gaussian Splatting and rendering virtual camera trajectories into events. It targets event-vision tasks like 3D reconstruction and deblurring, where real event data is scarce and expensive to capture.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The threshold choice is load-bearing and internally inconsistent: §3.6 adopts c∈[0.2,0.5] while App. B and §4 use c=1 to hide 3DGS floater artifacts, with no calibration against real DVS sensors; the sim-to-real generalization claim is unsupported.","rationale":"The reader's weakest assumption correctly identifies the threshold realism assumption as the load-bearing point. My stress-test agrees but sharpens it into a concrete internal inconsistency: the main text's claim of adaptive, data-driven threshold modeling in §3.6 is contradicted by Appendix B and Section 4, where c = 1 is chosen specifically to hide 3DGS rendering floaters. Since the dataset is the primary contribution, the actual event-generation parameters determine its benchmark value. The paper has real strengths: the 3DGS reconstruction quality is quantitatively reported (PSNR 29.8 on MVImgNet), the trajectory interpolation is well specified, and the scale of the proposed dataset is substantial. However, these strengths support the pipeline's feasibility, not the specific claim that the released events are sensor-realistic enough for sim-to-real transfer. A direct calibration or transfer experiment would settle whether the threshold choice matters in practice. Because the reader already assigned CONDITIONAL, my finding does not change the verdict; it strengthens the condition by giving it a precise, testable form.","tokens_in":18351,"tokens_out":3224,"duration_ms":30161,"concrete_test":"Regenerate a subset of, say, 50 GS2E scenes with both c = 0.4 and c = 1 using the released pipeline, train E-NeRF and Event-3DGS on each version, and evaluate both on real event sequences from DSEC (or a controlled lab capture) using reconstruction metrics and event-rate/contrast-distribution comparisons against real sensor recordings. If c = 1 matches or outperforms c = 0.4 on the real-data transfer evaluation, the concern is not decisive; if c = 0.4 transfers better, the released threshold is a genuine domain gap and should be recalibrated. Also inspect the released dataset metadata to verify which threshold was actually used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that GS2E is a high-fidelity benchmark whose synthetic events transfer to real DVS sensors. That claim rests on the event streams being faithful to real sensor behavior, and the paper does not establish this. Section 3.6 states that 'we adopt c ∈ [0.2, 0.5]' to balance detail preservation and temporal coherence, while Appendix B states that 'the GS2E dataset was simulated with the parameter setting c = 1' because lower thresholds expose 3DGS floater artifacts, and Section 4 sets ΘON = ΘOFF = 1 as default. These statements are mutually inconsistent, and neither choice is validated against real event-sensor data. Section 4.2 explicitly limits the synthetic-to-real evaluation to qualitative DSEC visual comparisons and notes that the Event Quality Score is unavailable. Consequently, the 'physically-informed contrast threshold modeling' contribution is not actually exercised at the released threshold, and the 'superior generalization' claim lacks quantitative sim-to-real support. This is load-bearing: if models trained on GS2E depend on artifact-suppressed, artificially sparse events, their performance on real sensors may reflect threshold mismatch rather than dataset quality.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GS2E, a synthetic event-stream data generator. It reconstructs static scenes from sparse multi-view RGB images using 3D Gaussian Splatting, generates dense camera trajectories via velocity-aware B-spline interpolation, renders image sequences along those trajectories, and feeds them into the DVS-Voltmeter stochastic event simulator. The stated contributions are a large-scale dataset (the paper variously says 1050, 1150, or 1900 scenes), a physically-informed contrast-threshold model, and superior generalization of models trained on GS2E for event-based 3D reconstruction, deblurring, and video reconstruction. The experiments report reconstruction metrics, a qualitative comparison against DSEC real-world event data, downstream task evaluations, and an ablation of trajectory interpolation. The appendix discloses limitations including the inability to model overexposure, underexposure, or high dynamic range.","tokens_in":18640,"tokens_out":6446,"duration_ms":55591,"significance":"If the synthetic-to-real transfer claim were quantitatively established, GS2E would address a real bottleneck in event-vision research: the scarcity of large, multi-view-consistent event datasets with aligned RGB frames. The pipeline design is plausible and the choice to build on 3DGS plus a stochastic event simulator is well motivated. The paper also makes a concrete dataset release, benchmarks several downstream tasks, and provides a useful comparison table of existing event-based 3D datasets. However, the current evidence does not support the central claim of superior generalization: the contrast threshold, which is the key physical parameter, is specified inconsistently and is chosen to hide rendering artifacts rather than calibrated to real sensors; the synthetic-to-real evaluation is only qualitative; and the experiments lack comparisons against prior synthetic event data generators or error bars. Significance is therefore conditional on substantial additional validation.","major_comments":[{"comment":"The contrast threshold, the key physical parameter of event simulation, is specified inconsistently. Section 3.6 states that 'we adopt c ∈ [0.2, 0.5]' after an empirical sweep, Appendix B states that 'the GS2E dataset was simulated with the parameter setting c = 1' because lower thresholds expose 3DGS floater artifacts, and Section 4 states 'ΘON = ΘOFF = 1 as default.' These settings are mutually incompatible, and neither is calibrated against real DVS sensor data; c = 1 is explicitly chosen to suppress rendering artifacts. Because the paper's claimed contribution of 'physically-consistent contrast threshold modeling' depends on this parameter, please reconcile the statements, report which threshold was actually used for the released scenes, and provide a quantitative sensitivity analysis or a calibration against real sensor statistics (for example, the contrast-threshold distributions in Stoffregen et al. or DSEC event rates).","section":"§3.6, Appendix B, §4 (Implementation details)"},{"comment":"The claim of 'superior generalization capabilities' is not supported by quantitative evidence. Section 4.2 states that the Event Quality Score is not publicly available and that the synthetic-to-real evaluation is limited to 'qualitative evaluations' and 'visual comparisons' against DSEC. No quantitative domain-gap metric is reported anywhere in the paper. Please add measurable evidence of sim-to-real transfer, for example: event-rate and contrast-histogram statistics on real versus synthetic streams, background-activity and noise comparisons, or a transfer experiment in which models trained on GS2E are evaluated on real event data (DSEC or MVSEC) and compared with models trained on prior synthetic data.","section":"§4.2"},{"comment":"The experiments do not compare GS2E against existing synthetic event data generators. Tables 1 and 2 report downstream task metrics for methods trained or evaluated on GS2E, but there is no head-to-head comparison with events generated by ESIM, v2e, Vid2E, or DVS-Voltmeter on ordinary rendered videos. Without such baselines, the abstract's claim that GS2E overcomes the limitations of prior event datasets is untestable. In addition, all table entries are single point estimates without error bars; please report means and standard deviations across scenes or seeds, and add a data-source ablation.","section":"Tables 1-2"},{"comment":"The reported dataset scale is internally inconsistent: the Abstract and Table 3 say 'over 1150 scenes'; Section 3.3 selects 1,000 MVImgNet scenes and 50 DL3DV scenes (1050 total); Section 4 and Appendix C say 'we choose and render 1.8k scenes from MVImageNet and 100 scenes from DL3DV' (1900 total). Since dataset scale is one of the stated contributions, please correct the numbers and make the scene-selection protocol precisely match the released dataset.","section":"Abstract, §3.3, §4, Appendix C, Table 3"},{"comment":"The Abstract claims 'diverse motion and lighting conditions,' but Appendix E explicitly concedes that the pipeline 'inherits the photometric constraints of 3D Gaussian Splatting' and 'may not faithfully replicate extreme illumination conditions such as overexposure or underexposure,' including low-light and high-dynamic-range scenes. Because event cameras are most advantageous exactly in these regimes, this limitation is directly relevant to the benchmark's fidelity claim. Please either add experiments or rendering modifications that cover HDR or overexposure cases, or explicitly scope the dataset claims to moderate illumination.","section":"Abstract vs Appendix E"}],"minor_comments":[{"comment":"There are several typos and inconsistent terms that should be fixed: 'interplementation verision' in Section 4 and Appendix C; 'ourdoor' in Section 3.3; 'off-the-shell' in Section 3.1; and inconsistent spelling of 'DVS-V oltmeter' versus 'DVS-Voltmeter.'","section":"§4, Appendix C, §3.1, §3.3"},{"comment":"The ablation of trajectory interpolation methods is presented only qualitatively; please add quantitative metrics (for example, trajectory smoothness or downstream reconstruction PSNR) or explicitly state that the comparison is qualitative.","section":"§4.4, Figure 6"},{"comment":"References [52] and [53] appear to be the same paper (DA4Event); the duplicate should be removed. Also, Table 3 is difficult to read in the provided version because of merged columns, missing spaces, and layout artifacts such as 'GS2E SubmissionSynthetic'; please reformat it.","section":"References and Table 3"},{"comment":"In the speed-list interface, the notation ΔT is used but T is never defined; please define all variables used in the speed-curve construction.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is confirmed on reading: the contrast-threshold statements in §3.6, Appendix B, and §4 are mutually inconsistent, and the released threshold is chosen to suppress 3DGS floaters rather than calibrated to real sensors. This is load-bearing for the sim-to-real claim. The dataset release and pipeline are plausible and potentially useful, but the paper needs quantitative domain-gap evidence, baseline comparisons against prior synthetic data, error bars, and a corrected and consistent statement of dataset scale before it can support the claimed conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read on GS2E: it is the first large-scale multi-view event dataset built by reconstructing real scenes with 3DGS and then simulating events with DVS-Voltmeter. That artifact alone is worth something: event vision needs more geometry-consistent, pose-annotated data, and the pipeline is a reasonable way to get it. The trajectory interpolation and the use of a stochastic event simulator are competent, if not novel individually. Credit where due: the released benchmark, at 1150 scenes, is bigger than prior synthetic event datasets for 3D reconstruction, and the downstream experiments show the data can train event-based 3D reconstruction and deblurring models that behave sensibly.\n\nThe soft spots are serious, and the stress-test note lands. The paper contradicts itself on the contrast threshold: Section 3.6 says c in [0.2, 0.5] is adopted; Appendix B says the released dataset was simulated with c=1 because lower thresholds expose 3DGS floaters; Section 4 sets both ON and OFF thresholds to 1. That is not a minor typo. The contrast threshold is the key physical parameter in the simulation, and the 'physically-informed contrast threshold modeling' contribution is not what generated the released data. The choice c=1 was driven by hiding rendering artifacts, not by matching a real sensor. The authors need to own this: either calibrate the threshold against real event data or drop the claim that the simulation is sensor-consistent.\n\nRelated, the 'superior generalization' claim is not backed by the experiments. Section 4.2 has only qualitative visual comparisons against DSEC, the EQS metric is unavailable, and there is no quantitative domain-gap metric, no comparison with prior synthetic event data (ESIM, v2e, Vid2E), and no error bars. For a paper whose core value is a training dataset, that is a big gap.\n\nMinor: scene counts are inconsistent (abstract says 1150, Section 3.3 says 1050, implementation details imply around 1900). The text also has duplicated sentences and typos ('interplementation verision'), which suggests haste.\n\nWho is this for? Event-vision researchers who need a large, pose-annotated synthetic dataset. If the authors fix the threshold story and add real comparisons, this could be a solid benchmark paper. As is, I would not rely on the released events as real-sensor-faithful.\n\nRecommendation: send it to peer review, but with a clear request for major revision. A serious referee should not desk-reject it; the dataset is new and potentially useful. But the authors need to resolve the threshold contradiction, release the artifacts, and provide quantitative evidence for sim-to-real transfer before it is accepted.","headline":"The dataset idea is new and worth building on, but an unresolved contrast-threshold contradiction and missing quantitative sim-to-real evidence currently undercut the paper's central claims.","tokens_in":19192,"tokens_out":2706,"would_cite":false,"duration_ms":22463,"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":"The paper claims that reconstructing static scenes with 3D Gaussian Splatting and then simulating events through a stochastic DVS voltage model produces large-scale, geometry-consistent event streams that transfer to real event-based 3D…","keywords":["event camera","3D Gaussian Splatting","synthetic dataset","event-based 3D reconstruction","contrast threshold","sim-to-real transfer","novel view synthesis","DVS-Voltmeter"],"falsifier":"Record a static real scene with a real DVS event camera while the camera follows a known trajectory, reconstruct the same scene with 3DGS from synchronized RGB frames, and run the paper's pipeline to simulate events along that trajectory. Compare the simulated stream with the real one, through event-rate statistics, contrast-threshold histograms, or by training a downstream 3D reconstruction model on GS2E and testing it on this real capture. If the modeled events are no better at predicting the real events or the resulting reconstruction than events from a standard video-driven simulator, the central sim-to-real claim fails.","tokens_in":18118,"feed_emoji":"⚡","tokens_out":8826,"duration_ms":68716,"temperature":0.7,"pith_summary":"GS2E claims that a static 3D Gaussian Splatting reconstruction of a real scene is enough of a foundation to generate large-scale, geometry-consistent event-camera data. The pipeline reconstructs each scene from sparse multi-view RGB images, renders dense image sequences along velocity-controlled virtual camera trajectories, and feeds them to the DVS-Voltmeter stochastic event simulator. A contrast threshold is chosen to match real sensor behavior while suppressing rendering artifacts. The paper argues that the resulting dataset of over 1,150 scenes provides the viewpoint diversity and temporal density that video-driven synthetic datasets lack, and that models trained on it transfer to real event data for 3D reconstruction. If this holds, high-quality event-vision training data can be produced from ordinary multi-view photo collections instead of specialized hardware or dense video.","feed_headline":"Gaussian splatting generates a 1,150-scene event-camera dataset","feed_subtitle":"Real-world scenes reconstructed in 3D feed a stochastic event simulator to produce better event-vision training data.","key_machinery":"The load-bearing mechanism is the pairing of 3D Gaussian Splatting with a stochastic DVS voltage model. 3DGS represents the scene as a set of anisotropic 3D Gaussians with optimized opacity and radiance, rendered by differentiable rasterization, which turns sparse RGB views into a photorealistic radiance field that can be re-rendered from arbitrary poses. DVS-Voltmeter then treats each pixel's voltage as $\\Delta V_d(t) = \\mu \\Delta t + \\sigma W(\\Delta t)$ and fires an event when the stochastic process crosses ON or OFF thresholds. The two are joined by a contrast threshold $c$: the paper sweeps $c \\in [0.25, 1.5]$ and settles on $c = 1$, a value that preserves event detail while hiding 3DGS floater artifacts. Velocity-controlled cubic B-spline trajectory interpolation provides the motion that turns rendered frames into temporally dense event streams.","core_discovery":"The central claim is that events rendered from 3DGS-reconstructed scenes, rather than from dense RGB video or graphics-engine assets, combine photorealistic appearance with strict geometric consistency. Starting from sparse multi-view images with known poses, GS2E trains a 3D Gaussian Splatting radiance field, then uses cubic B-spline interpolation with a velocity profile to create dense camera trajectories, including novel-view mini-trajectories. The rendered frames drive DVS-Voltmeter, which models each pixel's photovoltage as Brownian motion with drift and emits ON/OFF events when contrast thresholds are crossed, with inverse-Gaussian timestamp sampling. The paper reports that this setup yields event streams that align with scene structure, support non-uniform motion, and, when used to train event-based 3D reconstruction, deblurring, and video reconstruction models, generalize to real event data better than streams synthesized from dense RGB video.","pith_inferences":["If 3DGS truly decouples scene geometry from sensor simulation, the same reconstructed scenes could be reused to generate other sensor modalities, such as optical flow, depth, or alternative event-camera noise models, making 3DGS a general synthetic-data generator.","The paper chooses $c = 1$ partly to hide Gaussian floater artifacts rather than from direct sensor calibration; a natural extension is per-scene or per-sensor contrast-threshold distributions measured from real DVS recordings, which could close more of the sim-to-real gap.","The static-scene assumption is the clearest boundary of the method; extending the pipeline to dynamic scenes would require a deformable or 4D Gaussian representation, a step the paper leaves for future work.","If quantitative event-realism metrics such as EQS become publicly available, the paper's qualitative DSEC comparison could be replaced by a direct numerical check of simulated event realism."],"forward_implications":["Training event-based 3D reconstruction models on GS2E should yield better generalization to real event streams than training on events synthesized from dense RGB video, as the paper's DSEC comparisons and reconstruction experiments indicate.","Because each 3DGS scene can be re-rendered along any smooth trajectory, GS2E can generate multiple multi-view event streams per scene without new hardware capture, supporting controlled evaluations under different motion speeds.","The dataset's pairing of blurry RGB frames, sharp poses, and temporally dense events makes it usable for event-based deblurring and video reconstruction, not only 3D reconstruction.","The velocity-controlled interpolation lets researchers test models under mild, medium, and strong camera motion while holding the scene geometry fixed, isolating motion effects from scene effects."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting scene representation and differentiable rasterization that turns sparse RGB views into a re-renderable radiance field.","marker":"[32]"},{"why":"Supplies the DVS-Voltmeter stochastic event simulator, the Brownian-motion voltage model used to generate events from rendered frames.","marker":"[39]"},{"why":"Supplies the MVImgNet multi-view image collection used as the main source of object-centric scenes for reconstruction and event synthesis.","marker":"[90]"},{"why":"Supplies the DL3DV scene dataset with accurate poses and depth, used for scene-level structural richness.","marker":"[40]"},{"why":"Supplies the contrast-threshold calibration philosophy and the observation that thresholds vary across sensors, motivating the paper's threshold sweep.","marker":"[64]"},{"why":"The v2e video-driven event synthesis baseline whose limitations in viewpoint diversity and geometric consistency the paper addresses.","marker":"[24]"},{"why":"ESIM, an earlier event simulator whose fixed contrast thresholds and image/video-driven approach are contrasted with the proposed pipeline.","marker":"[56]"},{"why":"Event-3DGS is a baseline method in the paper's 3D reconstruction experiments, used to evaluate the dataset's training value.","marker":"[19]"},{"why":"E-NeRF is the event-based reconstruction baseline in the paper's experiments, used to measure model performance on GS2E data.","marker":"[33]"}],"fun_headline_variants":["Gaussian splatting generates synthetic event data from sparse views","GS2E: Gaussian splatting for high-fidelity event generation","Synthetic events from Gaussian splatting boost 3D reconstruction","Sparse views to event streams: Gaussian splatting does it","Gaussian splatting simulates events for better vision generalization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark value rests on the assumption that events generated from 3DGS-rendered images by DVS-Voltmeter with a manually chosen contrast threshold $c = 1$ are realistic enough to transfer to real DVS sensors; the paper's appendices concede that 3DGS does not model overexposure, underexposure, or high dynamic range, and that $c = 1$ was chosen in part to hide floater artifacts rather than calibrated against real sensor data.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian splatting generates synthetic event data from sparse views","GS2E: Gaussian splatting for high-fidelity event generation","Synthetic events from Gaussian splatting boost 3D reconstruction","Sparse views to event streams: Gaussian splatting does it","Gaussian splatting simulates events for better vision generalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000753,"raw_usage":{"total_tokens":3324,"prompt_tokens":891,"completion_tokens":2433,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":2343}},"tokens_in":507,"tokens_out":2433,"duration_ms":17109,"temperature":1.0,"reasoning_tokens":2343,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:20:12.237369+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a static real scene with a real DVS event camera while the camera follows a known trajectory, reconstruct the same scene with 3DGS from synchronized RGB frames, and run the paper's pipeline to simulate events along that trajectory. Compare the simulated stream with the real one, through event-rate statistics, contrast-threshold histograms, or by training a downstream 3D reconstruction model on GS2E and testing it on this real capture. If the modeled events are no better at predicting the real events or the resulting reconstruction than events from a standard video-driven simulator, the central sim-to-real claim fails.","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 rasterization that turns sparse RGB views into a re-renderable radiance field."},{"cited_title":"Dvs-voltmeter: Stochastic process-based event simulator for dynamic vision sensors","cited_arxiv_id":null,"evidence_quote":"Supplies the DVS-Voltmeter stochastic event simulator, the Brownian-motion voltage model used to generate events from rendered frames."},{"cited_title":"Reducing the sim-to-real gap for event cameras","cited_arxiv_id":null,"evidence_quote":"Supplies the contrast-threshold calibration philosophy and the observation that thresholds vary across sensors, motivating the paper's threshold sweep."},{"cited_title":"v2e: From video frames to realistic dvs events","cited_arxiv_id":null,"evidence_quote":"The v2e video-driven event synthesis baseline whose limitations in viewpoint diversity and geometric consistency the paper addresses."},{"cited_title":"Esim: an open event camera simulator","cited_arxiv_id":null,"evidence_quote":"ESIM, an earlier event simulator whose fixed contrast thresholds and image/video-driven approach are contrasted with the proposed pipeline."},{"cited_title":"E-nerf: Neural radiance fields from a moving event camera.IEEE Robotics and Automation Letters, 8(3):1587– 1594, 2023","cited_arxiv_id":null,"evidence_quote":"E-NeRF is the event-based reconstruction baseline in the paper's experiments, used to measure model performance on GS2E data."}],"review_version":1}