{"id":"fe4b5faf-c1a3-4d8e-b76c-84c22297c20d","arxiv_id":"2412.07293","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"EventSplat achieves real-time novel view synthesis from event-only camera streams by supervising 3D Gaussian Splatting with accumulated event differences, event-to-video-guided initialization, and spline-interpolated poses.","lead":"A new method lets fast-moving event cameras, which record brightness changes instead of normal images, build 3D scenes that can be rendered in real time. It combines an event-to-video prior, smooth camera trajectory estimates, and 3D Gaussian Splatting to beat earlier event-based NeRFs in speed and visual quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'state-of-the-art quality' claim is not established because main experiments omit event-based 3DGS baselines; the only comparison is an appendix table on 4 grayscale scenes with no error bars.","rationale":"The reader's verdict is CONDITIONAL with medium correctness risk, and I agree with that overall assessment. The clear strength of the paper is the rendering-speed advantage: 3DGS rasterization yields millisecond-scale render times versus hundreds-to-thousands of milliseconds for the NeRF baseline, and this is demonstrated consistently. The load-bearing weak point is the 'state-of-the-art quality' component of the central claim. The main experiments compare only against a NeRF method and a naive E2VID+3DGS pipeline; the cited event-based 3DGS methods are not part of the headline comparison. The appendix's EV-GS comparison is too thin to support a state-of-the-art claim: four scenes, grayscale only, no error bars, no per-scene breakdown. This is not merely a presentation issue, because the claim is precisely about being state-of-the-art among event-only methods. I also noted the mathematical inconsistency in Eq. (4): the formula writes E(a,b) as the integral of log(I'(t)) dt, whereas event cameras measure changes in log intensity, i.e., the integral of d/dt log I(t) dt; the implemented loss (L2-L1) uses the correct quantity, so this is a presentation flaw rather than a fatal one. The pose-interpolation assumption identified by the reader is real and stated explicitly in Sec. 5.1, but it is a scope condition common to event-based NVS methods and does not directly threaten the rendering claim when poses are available. The missing-baseline issue, by contrast, bears directly on whether the paper delivers what it promises, and it can be settled by running the published implementations on identical data. Therefore, the verdict should remain CONDITIONAL: the method is plausible and fast, but the state-of-the-art quality claim is contingent on a fair, complete comparison that the paper does not yet provide.","tokens_in":15046,"tokens_out":12091,"duration_ms":137298,"concrete_test":"Obtain the official implementations of EV-GS (Ref [70]), Event3DGS (Ref [72]), and E2GS (Ref [11]); run EventSplat and all baselines on the same EventNeRF synthetic scenes and the EDS/TUM-VIE real sequences used in this paper, with identical train/test splits and the same per-scene linear color transform. Report per-scene PSNR, SSIM, LPIPS, and rendering time for every method, along with error bars across repeated runs. If EventSplat is not statistically better on a majority of scenes and metrics, revise the abstract's 'state-of-the-art' claim to 'comparable quality at real-time speed.'","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim, as stated in the abstract and Sec. 1, is that EventSplat achieves 'state-of-the-art quality using only event data' while being orders of magnitude faster to render. The main quantitative comparisons (Table 1 for synthetic, Table 2 for real scenes) benchmark only Robust-e-NeRF and E2VID+3DGS. The paper cites several event-based 3DGS methods, including EV-GS [70], Event3DGS [72], and E2GS [11], but does not include them in the main tables. The only direct comparison to an event-based 3DGS method appears in Appendix B.2, Table 4, which reports mean PSNR and SSIM over just four synthetic scenes (chair, ficus, hotdog, mic) from the EventNeRF dataset, in grayscale, with no per-scene values, no error bars, and no rendering-time comparison. Additionally, even against the selected NeRF baseline, the synthetic-scene mean PSNR is 28.14 for EventSplat versus 28.19 for Robust-e-NeRF, so the 'higher visual fidelity' assertion does not hold uniformly across metrics. Because the headline contribution is explicitly positioned as state-of-the-art among event-only view synthesis methods, the omission of concurrent event-based 3DGS baselines from the primary evaluation is a load-bearing gap: if those methods perform comparably or better on the same data, the quality claim must be weakened to 'comparable quality at real-time speed,' even though the rendering-speed advantage is solid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EventSplat, a 3D Gaussian Splatting method trained directly from event camera streams for novel view synthesis. The training signal is a comparison between an accumulated event image D, computed by summing polarities over a randomly sampled sub-trajectory, and a rendered log-difference image D_hat, obtained by rasterizing the Gaussian scene at the start and end poses of that sub-trajectory, remosaicing the resulting color images, and taking the log difference. The method uses an event-to-video model to generate images that initialize Structure from Motion, and cubic spline interpolation to estimate camera poses at arbitrary event times. Experiments on seven synthetic scenes and five EDS real scenes compare against Robust-e-NeRF and E2VID+3DGS, reporting that EventSplat renders at 2.5–6 ms per frame, roughly two orders of magnitude faster than the NeRF baseline, with better mean SSIM and LPIPS on synthetic scenes and better PSNR, SSIM, and LPIPS on real scenes. Ablations show that event-to-video guided initialization and cubic spline interpolation each improve quality.","tokens_in":15379,"tokens_out":5031,"duration_ms":56725,"significance":"If validated, this work is significant: it is among the first to adapt 3D Gaussian Splatting to event-only supervision, and it demonstrates a clear and large rendering-speed advantage over event-based NeRF methods, an important practical bottleneck. The core speed result is robust and well supported by the reported timings. The ablation study supports the two main design choices. However, the broader claim of 'state-of-the-art quality' is not yet established, because the main experimental comparison omits concurrent event-based 3DGS baselines, and because the evaluation protocol appears to use test-time reference data for a linear color transform, which can inflate reported metrics. The paper is well written and the method is simple enough to reproduce, but the evaluation and the precision of the claims need strengthening before the headline conclusions can be accepted.","major_comments":[{"comment":"The primary evaluation compares only Robust-e-NeRF and E2VID+3DGS. The paper cites several event-based 3DGS methods (EV-GS, Event3DGS, E2GS) in Sec. 2.3, but the only direct comparison is Appendix Table 4, which reports mean PSNR and SSIM over just four grayscale synthetic scenes, with no per-scene values, no error bars, and no rendering-time comparison. Because the abstract and Sec. 1 claim state-of-the-art quality among event-only view synthesis methods, this omission is load-bearing. The authors should either include these baselines in the main tables with per-scene results and variance, or substantially weaken the claim to 'comparable quality at real-time speed.'","section":"Sec. 5.2, Tables 1–2, Appendix B.2"},{"comment":"The inference-time linear color transform is described as being 'designed to adjust our predictions' and later as requiring 'evaluation data as a reference.' If the scale and offset are optimized on the test views, the reported PSNR, SSIM, and LPIPS values are not a strict hold-out evaluation, because the test reference is used to fit a per-scene affine transform before computing the metrics. This can overstate reconstruction quality. Please clarify whether the transform is fit only on training views; if it uses test views, provide results without this fitting or with the transform estimated from training data only.","section":"Sec. 5.3 and Sec. 5.6"},{"comment":"The mean synthetic PSNR for EventSplat is 28.14 dB versus 28.19 dB for Robust-e-NeRF, so the abstract's blanket claim of 'higher visual fidelity' is not uniformly supported. The claim should be restricted to the metrics and settings where the improvement holds (SSIM and LPIPS on synthetic, all metrics on real scenes), and the PSNR deficit on synthetic scenes should be discussed, for example in terms of event accumulation noise, remosaicing, or the linear color transform.","section":"Table 1 and Sec. 5.4.1"},{"comment":"The method assumes constant-rate poses sampled at high frequency and uses cubic spline interpolation to assign poses at event times. The ablation in Table 3 shows a modest PSNR gain from adding cubic spline interpolation (18.75 to 18.86), but there is no sensitivity analysis to pose sampling rate or pose noise. Since misalignment between the accumulated event image D and the rendered log-difference image D_hat directly corrupts the supervision signal, please report experiments with coarser pose streams or otherwise quantify sensitivity to this core input assumption.","section":"Sec. 4.4 and Sec. 5.1"}],"minor_comments":[{"comment":"The notation E(a,b) := ∫_a^b log(I'(t)) dt is unusual: the accumulated event image D approximates the integral of the time derivative of log intensity, i.e., log I(b) − log I(a). Please clarify the notation, for instance by writing d/dt log I(t) in the integrand.","section":"Eq. (4)"},{"comment":"The sentence 'To fully utilize real-world event camera data, undistortion operation is performed as the accumulated image D is computed' is vague. Please state explicitly whether undistortion is applied to the event coordinates, to the accumulated image D, or to the rendered images, and where in the pipeline it occurs.","section":"Sec. 4.5"},{"comment":"TUM-VIE sequences are listed as part of the real-world evaluation and appear in qualitative figures (Fig. 4 and Fig. 9), but Table 2 reports only the five EDS scenes. Please add quantitative results for the TUM-VIE sequences or state explicitly why they are excluded from the numeric comparison.","section":"Sec. 5.1.2 and Table 2"},{"comment":"The comparison with EV-GS reports only mean PSNR and SSIM over four scenes. Please provide per-scene values, standard deviations, and a description of the evaluation protocol (e.g., whether the same color transform and test views were used for both methods), so that the comparison is reproducible and statistically meaningful.","section":"Appendix B.2, Table 4"},{"comment":"The metrics paragraph contains a typo ('It's worth noting') and a somewhat redundant explanation of why perceptual metrics are used. Consider tightening this paragraph and moving the linear color transform description to the experimental setup section, where it belongs.","section":"Sec. 5.3"}],"recommendation":"major_revision","confidential_remarks":"The core technical contribution is sound and the rendering-speed advantage is convincing. However, the evaluation falls short of the paper's stated 'state-of-the-art quality' claim: the main tables omit the event-based 3DGS baselines that the paper itself cites, and the test-time linear color transform may be using evaluation data as a reference, which could inflate the reported numbers. Both issues are fixable within the scope of a revision, but they are load-bearing for the central claim. I recommend major revision, with the expectation that the authors either add the missing baselines and a clean evaluation protocol or carefully re-scope the claims to 'real-time rendering with quality comparable to event-based NeRF methods.'"},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core result here is solid: EventSplat makes event-only novel view synthesis real-time, with rendering times in the low milliseconds versus hundreds to thousands of milliseconds for the NeRF baselines. That speed advantage is clearly demonstrated and is the paper's real contribution. The method itself is a sensible combination of existing pieces — random event accumulation into log-difference images, log-difference supervision of 3DGS, Bayer remosaicing for color event cameras, event-to-video guided SfM initialization, and cubic spline pose interpolation. The ablation shows the initialization and spline interpolation each help, and the design choices are well motivated. I believe the method works and is worth building on.\n\nThe soft spots are mostly about what the paper claims versus what it compares against. The abstract and intro say state-of-the-art quality using only event data, but the main tables only compare against Robust-e-NeRF and E2VID+3DGS. The paper cites EV-GS, Event3DGS, and E2GS, and the only direct event-based 3DGS comparison is a small appendix table on four grayscale synthetic scenes with no error bars, no per-scene values, and no rendering-time comparison. That is a real gap in the evidence for the quality claim. On the synthetic benchmark against the selected NeRF baseline, the mean PSNR is actually slightly lower (28.14 vs 28.19), so the claimed higher fidelity is not uniformly true even there. The inference-time linear color transform fitted to reference frames is also a mild weakness, though the paper discloses it in the limitations. The assumption of high-frequency constant-rate poses is stated clearly, so the reader knows the operating conditions.\n\nNone of this is load-bearing for the speed contribution, which stands on its own. But the paper would be stronger if the authors tempered the quality claim to 'comparable or better perceptual quality at real-time speed' and added the missing event-based 3DGS baselines with error bars to the main evaluation. As is, the work is important enough and sound enough to warrant serious peer review; a referee should push on the evaluation completeness rather than the method's validity. I'd cite it for the speed result and would bring it to a reading group on event-based 3DGS.","headline":"Real-time event-only 3DGS is a genuine, well-demonstrated advance, but the state-of-the-art quality claim overreaches the evidence: the main tables omit the event-based 3DGS baselines the paper itself cites.","tokens_in":15930,"tokens_out":1203,"would_cite":true,"duration_ms":27539,"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":"EventSplat trains 3D Gaussian Splatting from moving event cameras alone, reaching millisecond rendering at quality above event-based NeRFs.","keywords":["event cameras","3D Gaussian splatting","novel view synthesis","event-to-video initialization","cubic spline pose interpolation","real-time rendering","log-intensity difference","high dynamic range"],"falsifier":"Take a real event sequence with dense ground-truth poses, then train the method repeatedly with the pose stream artificially downsampled (for example, to one-tenth of the original rate). If reconstruction quality stalls or collapses when interpolated poses deviate from the true trajectory, the high-rate-pose assumption is confirmed as load-bearing.","tokens_in":14828,"feed_emoji":"⚡","tokens_out":4474,"duration_ms":50402,"temperature":0.7,"pith_summary":"This paper claims that 3D Gaussian Splatting can be supervised directly by event camera data, without any RGB images, to reconstruct static scenes and render novel views in real time. The key move is to treat event streams as measurements of log-intensity change and to train the Gaussian representation by matching accumulated events to the log-difference of rendered views. The authors show this approach matches or beats existing event-based neural radiance fields in visual quality while rendering roughly an order of magnitude faster. This would matter because event cameras work where ordinary cameras blur or saturate, and real-time rendering makes event-only view synthesis practical.","feed_headline":"Event-only view synthesis now renders in milliseconds","feed_subtitle":"EventSplat supervises 3D Gaussians with accumulated events, beating event-based NeRFs in quality and speed.","key_machinery":"The central identity is the log-difference equation: over a time interval [a, b], the event camera's output approximates E(a, b) = integral of the time derivative of log intensity, which can be estimated twice—once by accumulating events into the image D, and once by rendering two views from the Gaussian scene and subtracting their log-remosaiced images to form Dhat. The training loop minimizes the discrepancy between these two approximations. Two supporting mechanisms carry the method: remosaicing of the rendered RGB image back into the Bayer pattern so that per-pixel event data can be compared against the correct color channel, and a two-stage initialization where a pretrained event-to-video model synthesizes images, structure-from-motion produces initial 3D points, and cubic spline interpolation supplies poses at arbitrary event times.","core_discovery":"On its own terms, the paper introduces a fully event-driven training pipeline for 3D Gaussian Splatting. For a random sub-trajectory of the event stream, it accumulates events into a difference image D that approximates the integral of log-intensity change, and it rasterizes the current set of 3D Gaussians at the two endpoint poses to produce a rendered log-difference image Dhat. The reconstruction loss between D and Dhat, applied only where events actually occurred, drives the Gaussian optimization. To make this work, the authors add two ingredients: an event-to-video model whose output is fed to structure-from-motion to initialize Gaussian positions, and cubic spline interpolation of the camera trajectory so that arbitrary event timestamps get accurate poses. The experiments report that on synthetic scenes the method achieves comparable PSNR and better SSIM and LPIPS than Robust-e-NeRF, and on real scenes it outperforms both Robust-e-NeRF and E2VID-plus-3DGS on all three metrics, with rendering times of a few milliseconds per frame.","pith_inferences":["A natural extension would be to treat the random sub-trajectory length as a curriculum, starting with short intervals for local geometry and lengthening them to capture global structure, which the current equal random sampling only approximates.","The log-difference supervision could be combined with joint optimization of the spline control points, potentially relaxing the assumption that constant-rate poses are accurate enough for interpolation.","The static-scene restriction suggests a direct follow-up: attach per-Gaussian motion or a continuous time axis to render dynamic scenes from event streams, which the paper explicitly leaves for future work.","The remosaicing step is a simple channel-wise masking that could be replaced by a learned or more accurate color-processing model, possibly improving color fidelity on real event cameras."],"forward_implications":["Event-only novel view synthesis becomes practical in real time, with rendering times in the single milliseconds instead of tens to hundreds of milliseconds.","The method is usable in fast-motion and high-dynamic-range conditions where conventional cameras suffer from motion blur or saturation, because events encode relative log-intensity changes.","The pretrained event-to-video prior plus structure-from-motion provides a viable initialization for 3D Gaussian Splatting without any RGB supervision, reducing the need for dense color images.","Because the supervision signal is a log-intensity difference, the reconstructed scene carries an unknown absolute intensity offset that can be corrected by a simple linear transformation at inference time."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation, differentiable rasterizer, and the optimization and densification backbone that the method adapts to event supervision.","marker":"[26]"},{"why":"Provides the primary event-based NeRF baseline, Robust-e-NeRF, the synthetic dataset used for evaluation, and the symmetric contrast-threshold initialization settings.","marker":"[42]"},{"why":"The event-to-video model whose synthetic images are fed to structure-from-motion to initialize the Gaussian positions in the guided initialization.","marker":"[56]"},{"why":"ESIM, the event camera simulator used to generate the synthetic event streams for evaluation.","marker":"[55]"},{"why":"The EDS real-world dataset, from which five sequences are used to evaluate the method against baselines on real event data.","marker":"[24]"},{"why":"The TUM-VIE dataset, from which two mocap sequences are used for real-world evaluation.","marker":"[30]"},{"why":"EventNeRF, an earlier event-based NeRF approach and dataset source used for the grayscale synthetic comparison against EV-GS.","marker":"[59]"}],"fun_headline_variants":["Event-only splatting renders novel views in milliseconds","Gaussian splatting on events alone outpaces NeRF rendering","EventSplat turns event streams into real-time 3D views","Millisecond view synthesis from moving event cameras","Events alone drive 3D Gaussian rendering faster than NeRFs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method relies on the camera poses being sampled at a high constant rate so that cubic spline interpolation yields accurate poses for any event timestamp; if the pose stream is too coarse or drifts, the accumulated event image and the rendered log-difference image will be misaligned and training degrades.","fun_headline_variants_meta":{"raw":{"variants":["Event-only splatting renders novel views in milliseconds","Gaussian splatting on events alone outpaces NeRF rendering","EventSplat turns event streams into real-time 3D views","Millisecond view synthesis from moving event cameras","Events alone drive 3D Gaussian rendering faster than NeRFs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2119,"prompt_tokens":887,"completion_tokens":1232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":1149}},"tokens_in":503,"tokens_out":1232,"duration_ms":11482,"temperature":1.0,"reasoning_tokens":1149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:56:39.365053+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real event sequence with dense ground-truth poses, then train the method repeatedly with the pose stream artificially downsampled (for example, to one-tenth of the original rate). If reconstruction quality stalls or collapses when interpolated poses deviate from the true trajectory, the high-rate-pose assumption is confirmed as load-bearing.","supporting_citations":[{"cited_title":"3D Gaussian Splatting for Real-Time Radiance Field Rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting representation, differentiable rasterizer, and the optimization and densification backbone that the method adapts to event supervision."},{"cited_title":"Robust e-NeRF: NeRF from Sparse & Noisy Events under Non-Uniform Motion,","cited_arxiv_id":null,"evidence_quote":"Provides the primary event-based NeRF baseline, Robust-e-NeRF, the synthetic dataset used for evaluation, and the symmetric contrast-threshold initialization settings."},{"cited_title":"High speed and high dynamic range video with an event camera","cited_arxiv_id":null,"evidence_quote":"The event-to-video model whose synthetic images are fed to structure-from-motion to initialize the Gaussian positions in the guided initialization."},{"cited_title":"ESIM: an open event camera simulator","cited_arxiv_id":null,"evidence_quote":"ESIM, the event camera simulator used to generate the synthetic event streams for evaluation."},{"cited_title":"Event-aided direct sparse odometry","cited_arxiv_id":null,"evidence_quote":"The EDS real-world dataset, from which five sequences are used to evaluate the method against baselines on real event data."},{"cited_title":"TUM-VIE: The TUM Stereo Visual-Inertial Event Dataset","cited_arxiv_id":null,"evidence_quote":"The TUM-VIE dataset, from which two mocap sequences are used for real-world evaluation."},{"cited_title":"EventNeRF: Neural radiance fields from a single colour event camera","cited_arxiv_id":null,"evidence_quote":"EventNeRF, an earlier event-based NeRF approach and dataset source used for the grayscale synthetic comparison against EV-GS."}],"review_version":1}