Pith. sign in

REVIEW 2 major objections 4 minor 39 references

JADE-GS restores sharp 3D scenes from blurry frames and events by learning a per-pixel blend of two deblurring priors, with no sharp ground truth needed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 01:38 UTC pith:F7Y24SUT

load-bearing objection A genuinely novel, well-ablated method for per-pixel allocation between analytical and learned event-based deblurring priors in 3DGS, whose central claim about the router's spatial choices is plausible but never directly validated against per-pixel oracle quality. the 2 major comments →

arxiv 2607.14990 v2 pith:F7Y24SUT submitted 2026-07-16 cs.CV

JADE-GS: Joint Allocation of Deblurring Evidence for Event-Assisted 3D Gaussian Splatting

Haoyu Fu , Jiafeng Huang , Yuchen Wang , Shengjie Zhao This is my paper
classification cs.CV
keywords event camerasmotion deblurring3D Gaussian Splattingevent double integrallearned restorationpixel-wise routingnovel view synthesisperceptual quality
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

JADE-GS sets out to prove that motion-blurred 3D reconstruction can be sharpened by deciding, pixel by pixel, which of two deblurring priors to trust. The paper argues that the analytical event-inversion prior and a learned frame–event restorer fail in complementary regions, and that a uniform or globally learned weight cannot exploit that complementarity. It introduces a tiny Spatial Prior Router that reads only the blurry frame and event stream, fuses the two fixed restorations into a supervision target, and is trained without sharp references. If correct, the method reaches or beats the perceptual quality of diffusion-based alternatives at a fraction of the training cost and keeps the native real-time 3DGS rendering pipeline at inference.

Core claim

The paper's central claim is that the right way to combine analytical and learned deblurring evidence in event-assisted 3D Gaussian Splatting is spatial allocation, not averaging. A lightweight router predicts a per-pixel weight α from the original measurements and builds the fused target I_pseudo = α ⊙ I_edi + (1−α) ⊙ F(B,V(E)). Because both restorations are fixed and measurement-dependent, the router's gradient is nonzero only where the two priors disagree. The router is trained solely by two consistency signals—agreement with the mid-exposure rendering of the Gaussian scene being optimized, and agreement of the reblurred prediction with the measured blurry frame—and is discarded after opt

What carries the argument

The Spatial Prior Router is the load-bearing component: a roughly 9,300-parameter convolutional network with shallow event and image encoders and a three-layer head that outputs a single sigmoid channel α ∈ [0,1] per pixel. Its input is the blurry RGB frame and the polarity-signed temporal projection of the binned event voxel—never the two restorations—so the allocation is conditioned on the cues that determine local reliability. Its output fuses the analytical EDI reconstruction and the frozen learned restorer into one supervision target. The router is trained by two ℓ1 consistencies: render consistency against a stop-gradient mid-exposure 3DGS rendering, and reblur consistency that pushes

Load-bearing premise

The method assumes that two consistency signals—matching the scene being reconstructed and matching the measured blur after a simple reblur—can teach the router a useful per-pixel blend even though no sharp ground truth is ever seen.

What would settle it

A direct test: take a scene with sharp ground truth, compute the oracle per-pixel choice between the two restorations, and compare it with the learned α map. If the router disagrees with the oracle on a large fraction of pixels where the two restorations differ, the consistency losses are not tracking local reliability. An even simpler check is to inflate event noise in a held-out region and see whether α moves toward the learned restorer in exactly that region; if it does not, the router has not learned to allocate by evidence quality.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On the real benchmark, perceptual distance (LPIPS) drops from the strongest baseline's 0.0496 to 0.0382, a 23% reduction, while PSNR and SSIM also take the top spot; the ordering holds on every scene.
  • A learned global scalar weight collapses to α ≈ 1.4 × 10⁻⁵, effectively abandoning the analytical branch everywhere; only a per-pixel map keeps both priors contributing, so spatial allocation is the mechanism behind the gain.
  • Training fits in roughly 0.75 hours at 4.6 GB peak memory on a single GPU, versus about 10 hours for a diffusion-based prior, and inference runs at native 3DGS speed (about 110 FPS) because the router and restorers are removed.
  • The perceptual advantage previously obtained from a diffusion prior is reached without one, so the method offers a non-generative route to sharp event-assisted reconstruction.
  • The learned α map is spatially non-uniform and does not follow a simple event-density rule, suggesting the router encodes a task-specific judgment about where physical event evidence should dominate.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: because both restorations depend only on the measurements, the same allocation idea should work with any pair of complementary restorers; a testable extension is to swap in a different learned restorer or a diffusion-denoised candidate and see whether the router re-learns a sensible map.
  • Inference: the reblur consistency term is biased toward the analytical branch, since EDI inverts the same exposure model; a calibrated reblur or a third consistency signal from the RGB frame alone might shift allocations and would be a direct way to test whether the router is truly tracking reliability rather than the bias.
  • Inference: the paper leaves open whether a hand-designed map could replace the router; a cheap experiment comparing the learned map to an oracle per-pixel selection from sharp ground truth would quantify how much of the gain comes from perfect allocation, and could motivate a more compact router.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. JADE-GS proposes a method for event-assisted 3D Gaussian Splatting deblurring. It uses two complementary restoration priors — an analytical EDI inversion and a frozen EFNet-based learned refiner — and learns a lightweight Spatial Prior Router that predicts a per-pixel weight α fusing the two restorations into an auxiliary supervision target for the 3D scene. The router is trained without a sharp reference by minimizing render consistency with the scene's mid-exposure rendering and reblur consistency with the measured blurry frame; it is discarded at inference. The paper reports state-of-the-art or competitive results on EvDeblur-CDA VIS and EvDeblur-Blender, with particularly strong perceptual metrics, lower training cost than diffusion-based DiET-GS, and thorough ablations including uniform vs per-pixel allocation, individual consistency losses, a learned global scalar, a per-view free map baseline, no-sigmoid parameterization, multi-seed runs, and extended training trajectories.

Significance. The contribution is timely: it shows that a diffusion prior is not necessary to obtain the perceptual gains in event-assisted 3DGS, and that a cheap, measurement-conditioned per-pixel blending of analytical and learned priors can improve reconstruction. The paper is unusually transparent: a fixed 25k iteration budget is used throughout and justified with extended trajectories; three-seed variance is reported; the routing branch has only 9,314 parameters; all major settings are listed. These are genuine strengths. The main reservations are (i) the SOTA comparison rests on quoted baseline numbers rather than independent reproduction, and (ii) the claim that α encodes 'local prior reliability' is supported only indirectly; the paper's own Appendix F.c concedes the reblur term is biased toward EDI. If the requested diagnostics confirm the allocation interpretation, the paper would be a solid contribution; even without them, the empirical result is of interest.

major comments (2)
  1. [III-C / Eq. (6)-(9), App. F.c, App. A] The central mechanism is not yet demonstrated. The reblur term B_mid(I_pseudo,V)-B is, as the paper acknowledges, systematically biased: since I_edi is obtained by inverting the same EDI model, it has small reblur residual wherever the inversion is self-consistent, regardless of whether I_edi is closer to the true sharp image. The render term uses sg(Imid), where Imid is a rendering of a scene that is itself optimized against the router's previous fused targets; it is a moving and partly circular supervisory signal. The ablations in Table II and App. J convincingly show that per-pixel, measurement-conditioned α outperforms uniform/global and per-view free maps, ruling out naive overfitting. However, none of the experiments establishes that high α occurs where I_edi is locally more accurate than F, which is what 'spatial evidence allocation' means. I request: (i) compare learned α to orac
  2. [IV-B / Table I] The state-of-the-art numbers are quoted from [13] under 'the same public protocol', and JADE-GS is reported as a single run (multi-seed only in the supplement). Since JADE-GS intentionally reuses the DiET-GS/Ev-DeblurNeRF backbone, even small differences in loss weighting, warm-up schedules, or pose refinement could affect the comparison. The 'leading' claim thus rests on a literature comparison that the reader cannot audit. Please either re-run the public baselines in the same code/GPU and report seed variance, or clearly label the table as 'quoted from [13]' and make the code/checkpoints available. The efficiency table (Table IX) also mixes hardware and iteration budgets; the speed/memory comparisons should be qualified as indicative.
minor comments (4)
  1. [Eq. (7) / Eq. (6)] The scale-free reblur surrogate and the unit weights in Eq. (6) are not subject to any sensitivity ablation. Since the reblur term is known to be EDI-biased, please report at least a coarse sweep of c and of the loss-weight ratio, or justify the fixed values analytically.
  2. [Fig. 6 caption] The caption misspells 'blurlabelequipments'; the table row uses 'blurlabequipment'. Please correct for consistency.
  3. [Abstract / Table I] The abstract's 'leading perceptual quality on both benchmarks' should be read against JADE-GS being second on MUSIQ on CDA VIS. A one-sentence qualification in the experimental section would help.
  4. [Table VII / App. F.e] warmup and simul_start are inherited from DiET-GS and not rescaled to the 25k budget; this is disclosed, but a sentence in Sec. IV stating why this does not disadvantage the method would help.

Circularity Check

2 steps flagged

Router training is partially self-referential: the reblur term favors the EDI branch by construction and the render term is a moving-target loop; external GT benchmarks and ablations provide independent grounding.

specific steps
  1. self definitional [Section III-C, Eqs. (6)-(7); Appendix F.c]
    "Because I_edi inverts the same event-based exposure model, reblur consistency tends to favor the analytical branch where that inversion is self-consistent, and is not a neutral selector between the two candidates. ... the analytical reconstruction is itself obtained by inverting the same exposure model, so wherever that inversion is accurate its residual is small by construction and the term pushes α up."

    I_edi = Φ_edi(B,E) is defined as the algebraic inverse of the event-exposure model, and Eq. (7) defines B_mid using the same cumulative event voxel ΔE_{k,mid}. Therefore B_mid(I_edi,V) ≈ B holds wherever the inversion is self-consistent, regardless of whether I_edi is closer to the true sharp image. The reblur term of Eq. (6) thus rewards α→1 by construction, and the router's allocation is partly fitted to EDI's self-consistency rather than to independent evidence about local sharpness. The paper explicitly acknowledges this bias, but still uses the term as one of the two signals claimed to 'anchor the map to the sensor'.

  2. other [Section III-C, Eqs. (6) and (9)]
    "L_route = ||I_pseudo − sg(I_hat_mid)||_1 + ||B_mid(I_pseudo, V(E)) − B||_1 ... L_tgt = ℓ_rec(I_hat_mid, sg(I_pseudo)) + ℓ_rec(g(I_hat_mid), g(sg(I_pseudo)))"

    The render-consistency term trains the router to match I_hat_mid, but the scene is itself optimized by L_tgt to match the same fused target I_pseudo. Thus I_hat_mid is not an independent witness: it has been pulled toward the router's previous outputs, so the render term can reinforce whatever allocation the router already produced (e.g., α≈1 if the scene has been trained mostly on I_edi). Stop-gradient blocks router gradients through the scene but does not remove the dependency of I_hat_mid on past I_pseudo. This is a bootstrapping loop, only partially anchored by the inherited EDI/blur objectives and by the biased reblur term.

full rationale

JADE-GS is not a case of pure circularity: the final novel-view results are evaluated on external ground-truth benchmarks against published baselines, the router is discarded at inference, and there are no load-bearing self-citations or imported uniqueness theorems. However, the unsupervised training of the Spatial Prior Router contains two non-independent signals. First, the reblur consistency loss is constructed from the same event-exposure model that defines I_edi, so it favors the analytical branch wherever that branch is self-consistent, not necessarily where it is correct; the paper admits this bias (Appendix F.c). Second, the render consistency loss compares the router output to a scene rendering that is itself trained with the router's fused target, creating a moving-target loop. These issues weaken the claim that α learns true local prior reliability without sharp supervision, and the ablations mainly show that per-pixel α is better than uniform or per-view free α, not that α tracks the regions where EDI is actually better. Nevertheless, the scene retains inherited supervision from the measured blurry frames and EDI, external benchmarks provide independent validation, and the authors are transparent about the reblur bias. The circularity is partial, not total: score 4.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The method rests on inherited domain assumptions (EDI model, exposure integral, static scene) and several hand-set hyperparameters. The router's self-supervised training loop introduces a potential bootstrapping circularity that is mitigated by the reblur anchor but not fully removed.

free parameters (6)
  • Reblur contrast scale c in Eq. (7) = 1 (with per-window v_max normalization)
    Hand-set constant; the paper calls the reblur operator a 'scale-free event-consistency surrogate' rather than a radiometrically calibrated reblur.
  • Routing loss weights (render/reblur) = 1 / 1
    Unit coefficients in Eq. (6), chosen by hand; no sensitivity analysis reported.
  • Scene-step target weight efnet_weight = 1
    Hyperparameter in Table VII; weight of the fused-target loss in the scene step.
  • D-SSIM mixture lambda_dssim = 0.3
    Inherited from DiET-GS/3DGS; not tuned for this method.
  • Warmup and simul_start = 1500 / 7000
    Inherited constants from DiET-GS 100k schedule, applied unchanged under the 25k budget, so auxiliary losses activate proportionally earlier.
  • Event voxel temporal bins T = 6
    Hand-selected bin count for the exposure window; the mid-exposure boundary falls after the third bin.
axioms (6)
  • domain assumption Exposure formation model of Eq. (1): the blurry frame is the time-average of latent sharp views over the exposure interval, approximated by nine sampled poses with a mid-exposure reference.
    Invoked in §III-A; inherited from prior event-guided works [12], [13] and used throughout the method.
  • domain assumption Event generation follows the Event Double Integral (EDI) model: events trigger on log-intensity changes crossing a fixed contrast threshold, and analytical inversion yields I_edi.
    Invoked in §III-B; the analytical branch assumes this model; sensitivity to threshold mismatch is acknowledged.
  • domain assumption The frozen EFNet prior, pretrained on GoPro, generalizes to the target datasets and provides a valid complementary restoration estimate at the mid-exposure reference.
    Invoked in §III-B; no fine-tuning on target scenes; relies on cross-domain generalization of EFNet.
  • ad hoc to paper The reblur operator of Eq. (7) with normalized voxel and c=1 is a sufficient consistency surrogate for the exposure model (scale-free, monotone).
    Introduced in §III-C; the paper states it is not radiometrically calibrated and has a systematic bias toward the analytical branch (Appendix F.c).
  • domain assumption The scene is static with ego-motion blur; no moving objects; synchronized frame-event capture.
    Stated in §I and Limitations; the method does not handle dynamic scenes.
  • ad hoc to paper The two consistency signals (render consistency and reblur consistency) are sufficient to train the router without sharp supervision, and the coupled scene-router optimization converges to a useful solution.
    The core self-supervision claim, supported by ablations but not theoretically analyzed (§III-C).
invented entities (1)
  • Spatial Prior Router R_theta no independent evidence
    purpose: Predicts per-pixel allocation alpha between EDI and EFNet restorations for the fused supervision target.
    A new learned module introduced by the paper; it has no falsifiable handle outside the pipeline, though its internal behavior is visualized and ablated.

pith-pipeline@v1.3.0-alltime-deepseek · 20418 in / 16749 out tokens · 172935 ms · 2026-08-04T01:38:00.128502+00:00 · methodology

0 comments
read the original abstract

Neural radiance fields and 3D Gaussian Splatting assume that each training image is a sharp and geometrically consistent observation of the scene. Motion blur violates this assumption because a single exposure integrates a continuous range of camera poses. Exposure integration also removes the temporal information needed to recover the corresponding sharp observation. Event cameras preserve this information at microsecond resolution and therefore provide a natural complement to conventional images. Existing event-assisted reconstruction methods predominantly obtain image supervision through analytical inversion of the Event Double Integral. Learned restoration from frames and events offers a second prior. Although weaker when used alone, it fails in different regions and provides complementary evidence. We present JADE-GS, which formulates the combination of these priors as spatial evidence allocation. A lightweight Spatial Prior Router predicts a pixelwise allocation using only the blurry frame and event stream, then fuses the two fixed restorations into an additional supervision target. The router is trained without a sharp reference using consistency with the scene under reconstruction and the measured exposure, and is removed after optimization. Experiments show that JADE-GS achieves leading perceptual quality on both benchmarks, attains the best fidelity on the real benchmark, and remains competitive on the synthetic one. It requires substantially lower training overhead than diffusion-based alternatives and preserves native 3DGS rendering with no generative decoding at inference.

Figures

Figures reproduced from arXiv: 2607.14990 by Haoyu Fu, Jiafeng Huang, Shengjie Zhao, Yuchen Wang.

Figure 1
Figure 1. Figure 1: Perceptual error (LPIPS ↓) versus training GPU memory (GB ↓) on the real EvDeblur-CDAVIS benchmark. JADE-GS occupies the desirable low-error, low-memory corner: it trains in under 5 GB while matching or beating the perceptual quality of every baseline. Frame-only methods keep memory low but incur high error; diffusion-based methods reach competitive error only at several times the memory. JADE-GS resolves … view at source ↗
Figure 1
Figure 1. Figure 1: The left panel shows reconstruction from blurry frames and event [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of JADE-GS. The 2D teacher ingests the blurry frame, an event-voxel grid and an EDI initializa [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Complementary restoration priors. Standalone EDI preserves several [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the learned edge-texture division. Top: event-sensitive feature activations obtained by zero [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of JADE-GS. The analytical EDI branch and the frozen EFNet refiner [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on the five CDAVIS scenes (one scene per row). From left to right: blurry input, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on the five CDAVIS scenes (one scene per row). From left to right: blurry input, EDI+GS, EFNet+GS, DiET-GS, DiET-GS++, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The two priors used in JADE-GS. EDI preserves sharper boundaries but drifts and adds noise; EFNet pro [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Learned routing behavior on a CDAVIS scene. Left: the restored [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Training loss curves for all five CDAVIS scenes. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the learned routing weight [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: PSNR, SSIM, and LPIPS evaluation curves for all five CDAVIS scenes. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training loss curves for all five CDAVIS scenes. The losses have flattened by roughly 20k–25k iterations on every scene, which is the criterion used [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    NeRF: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing scenes as neural radiance fields for view synthesis,” inEuropean Conference on Computer Vision (ECCV), 2020, pp. 405–421

    2020

  2. [2]

    Zip-NeRF: Anti-aliased grid-based neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Zip-NeRF: Anti-aliased grid-based neural radiance fields,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 19 697–19 705

    2023

  3. [3]

    3D gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3D gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, pp. 1–14, 2023

    2023

  4. [4]

    BAD-NeRF: Bundle adjusted deblur neural radiance fields,

    P. Wang, L. Zhao, R. Ma, and P. Liu, “BAD-NeRF: Bundle adjusted deblur neural radiance fields,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 4170–4179

    2023

  5. [5]

    BAD-Gaussians: Bundle adjusted deblur gaussian splatting,

    L. Zhao, P. Wang, and P. Liu, “BAD-Gaussians: Bundle adjusted deblur gaussian splatting,” inEuropean Conference on Computer Vision (ECCV), 2024, pp. 233–250

    2024

  6. [6]

    Deblurring 3D gaussian splatting,

    B. Lee, H. Lee, X. Sun, U. Ali, and E. Park, “Deblurring 3D gaussian splatting,” inEuropean Conference on Computer Vision (ECCV), 2024, pp. 127–143

    2024

  7. [7]

    A 128×128 120 db 15µs latency asynchronous temporal contrast vision sensor,

    P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128×128 120 db 15µs latency asynchronous temporal contrast vision sensor,”IEEE Journal of Solid-State Circuits, vol. 43, no. 2, pp. 566–576, 2008

    2008

  8. [8]

    Event-based vision: A survey,

    G. Gallego, T. Delbr ¨uck, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis, and D. Scara- muzza, “Event-based vision: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 1, pp. 154–180, 2020

    2020

  9. [9]

    Asymmetric hierarchical difference-aware interaction network for event-guided motion deblur- ring,

    W. Yang, J. Wu, L. Li, W. Dong, and G. Shi, “Asymmetric hierarchical difference-aware interaction network for event-guided motion deblur- ring,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025, pp. 9265–9273

    2025

  10. [10]

    Motion- adaptive transformer for event-based image deblurring,

    S. Xu, Z. Sun, M. Zhong, C. Cao, Y . Liu, X. Fu, and Y . Chen, “Motion- adaptive transformer for event-based image deblurring,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025, pp. 6952–6960

    2025

  11. [11]

    Bringing a blurry frame alive at high frame-rate with an event camera,

    L. Pan, C. Scheerlinck, X. Yu, R. Hartley, M. Liu, and Y . Dai, “Bringing a blurry frame alive at high frame-rate with an event camera,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 6820–6829

    2019

  12. [12]

    Mitigating motion blur in neural radiance fields with events and frames,

    M. Cannici and D. Scaramuzza, “Mitigating motion blur in neural radiance fields with events and frames,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 9286–9296

    2024

  13. [13]

    DiET-GS: Diffusion prior and event stream- assisted motion deblurring 3D gaussian splatting,

    S. Lee and G. H. Lee, “DiET-GS: Diffusion prior and event stream- assisted motion deblurring 3D gaussian splatting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 21 739–21 749

    2025

  14. [14]

    Event-aided sharp ra- diance field reconstruction for fast-flying drones,

    R. Zou, M. Cannici, and D. Scaramuzza, “Event-aided sharp ra- diance field reconstruction for fast-flying drones,”arXiv preprint arXiv:2602.21101, 2026

    arXiv 2026

  15. [15]

    Event-based fusion for motion deblurring with cross- modal attention,

    L. Sun, C. Sakaridis, J. Liang, Q. Jiang, K. Yang, P. Sun, Y . Ye, K. Wang, and L. Van Gool, “Event-based fusion for motion deblurring with cross- modal attention,” inEuropean Conference on Computer Vision (ECCV), 2022, pp. 412–428

    2022

  16. [16]

    Deblur-NeRF: Neural radiance fields from blurry images,

    L. Ma, X. Li, J. Liao, Q. Zhang, X. Wang, J. Wang, and P. V . Sander, “Deblur-NeRF: Neural radiance fields from blurry images,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 12 861–12 870

    2022

  17. [17]

    PDRF: Progressively deblurring radiance field for fast scene reconstruction from blurry images,

    C. Peng and R. Chellappa, “PDRF: Progressively deblurring radiance field for fast scene reconstruction from blurry images,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 2023, pp. 2029–2037

    2023

  18. [18]

    Motion deblurring via spatial- temporal collaboration of frames and events,

    W. Yang, J. Wu, J. Ma, L. Li, and G. Shi, “Motion deblurring via spatial- temporal collaboration of frames and events,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 6531–6539

    2024

  19. [19]

    EventNeRF: Neural radiance fields from a single colour event camera,

    V . Rudnev, M. Elgharib, C. Theobalt, and V . Golyanik, “EventNeRF: Neural radiance fields from a single colour event camera,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 4992–5002

    2023

  20. [20]

    Ev-NeRF: Event-based neural radiance field,

    I. Hwang, J. Kim, and Y . M. Kim, “Ev-NeRF: Event-based neural radiance field,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023, pp. 837–847

    2023

  21. [21]

    E 2NeRF: Event enhanced neural radiance fields from blurry images,

    Y . Qi, L. Zhu, Y . Zhang, and J. Li, “E 2NeRF: Event enhanced neural radiance fields from blurry images,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 13 254– 13 264

    2023

  22. [22]

    EV-GS: Event-based gaussian splatting for efficient and accurate radiance field rendering,

    J. Wu, S. Zhu, C. Wang, and E. Y . Lam, “EV-GS: Event-based gaussian splatting for efficient and accurate radiance field rendering,” inIEEE International Workshop on Machine Learning for Signal Processing (MLSP), 2024, pp. 1–6

    2024

  23. [23]

    E2GS: Event enhanced Gaussian splatting,

    H. Deguchi, M. Masuda, T. Nakabayashi, and H. Saito, “E2GS: Event enhanced Gaussian splatting,” inIEEE International Conference on Image Processing (ICIP), 2024

    2024

  24. [24]

    EaDeblur-GS: Event-assisted 3D deblur reconstruction with Gaussian splatting,

    Y . Weng, Z. Shen, R. Chen, Q. Wang, and J. Wang, “EaDeblur-GS: Event-assisted 3D deblur reconstruction with Gaussian splatting,”arXiv preprint arXiv:2407.13520, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    EvaGaussians: Event stream assisted gaussian splatting from blurry images,

    W. Yu, C. Feng, J. Li, J. Tang, J. Yang, Z. Tang, M. Cao, X. Jia, Y . Yang, L. Yuan, and Y . Tian, “EvaGaussians: Event stream assisted gaussian splatting from blurry images,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 24 780– 24 790

    2025

  26. [26]

    E-3DGS: Event-based novel view rendering of large-scale scenes using 3D gaussian splatting,

    S. Zahid, V . Rudnev, E. Ilg, and V . Golyanik, “E-3DGS: Event-based novel view rendering of large-scale scenes using 3D gaussian splatting,” inInternational Conference on 3D Vision (3DV), 2025, pp. 926–934

    2025

  27. [27]

    EBAD-Gaussian: Event-driven bundle adjusted deblur Gaussian splatting,

    Y . Deng, Y . Wang, R. Xiao, C. Tang, J. Zhou, J. Fan, D. Xiong, J. Lv, and H. Tang, “EBAD-Gaussian: Event-driven bundle adjusted deblur Gaussian splatting,”arXiv preprint arXiv:2504.10012, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    DeblurSplat: SfM-free 3D Gaussian splatting with event camera for robust deblurring,

    P. Li, Y . Lu, P. Song, W. Guo, H. Yao, F. R. Yu, and H. Xiong, “DeblurSplat: SfM-free 3D Gaussian splatting with event camera for robust deblurring,”IEEE Transactions on Multimedia, 2026, accepted

    2026

  29. [29]

    AsyncEvGS: Asynchronous event-assisted Gaussian splatting for hand- held motion-blurred scenes,

    J. Dai, R. Jin, B. Xu, Y . Chen, L. Xu, M. Yu, T. Xue, and S. Guo, “AsyncEvGS: Asynchronous event-assisted Gaussian splatting for hand- held motion-blurred scenes,”arXiv preprint arXiv:2605.07192, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    EvFlow-GS: Event enhanced motion deblurring with optical flow for 3D gaussian splatting,

    F. An, Y . Deng, Z. Zhang, and R. Xiao, “EvFlow-GS: Event enhanced motion deblurring with optical flow for 3D gaussian splatting,”arXiv preprint arXiv:2604.22183, 2026

    Pith/arXiv arXiv 2026

  31. [31]

    DiSR-NeRF: Diffusion-guided view- consistent super-resolution NeRF,

    J. L. Lee, C. Li, and G. H. Lee, “DiSR-NeRF: Diffusion-guided view- consistent super-resolution NeRF,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 20 561–20 570

    2024

  32. [32]

    Design of an RGBW color VGA rolling and global shutter dynamic and active-pixel vision sensor,

    C. Li, C. Brandli, R. Berner, H. Liu, M. Yang, S.-C. Liu, and T. Delbruck, “Design of an RGBW color VGA rolling and global shutter dynamic and active-pixel vision sensor,” inIEEE International Symposium on Circuits and Systems (ISCAS), 2015, pp. 718–721

    2015

  33. [33]

    ESIM: An open event camera simulator,

    H. Rebecq, D. Gehrig, and D. Scaramuzza, “ESIM: An open event camera simulator,” inConference on Robot Learning (CoRL), 2018, pp. 969–982

    2018

  34. [34]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 586–595

    2018

  35. [35]

    MUSIQ: Multi- scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “MUSIQ: Multi- scale image quality transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 5148– 5157

    2021

  36. [36]

    Exploring CLIP for assessing the look and feel of images,

    J. Wang, K. C. K. Chan, and C. C. Loy, “Exploring CLIP for assessing the look and feel of images,” inProceedings of the AAAI Conference on Artificial Intelligence, 2023, pp. 2555–2563

    2023

  37. [37]

    Unsupervised event-based video reconstruction,

    G. Fox, X. Pan, A. Tewari, M. Elgharib, and C. Theobalt, “Unsupervised event-based video reconstruction,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024

    2024

  38. [38]

    4D gaussian splatting for real-time dynamic scene rendering,

    G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and X. Wang, “4D gaussian splatting for real-time dynamic scene rendering,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 20 310–20 320

    2024

  39. [39]

    Dynamic 3D gaus- sians: Tracking by persistent dynamic view synthesis,

    J. Luiten, G. Kopanas, B. Leibe, and D. Ramanan, “Dynamic 3D gaus- sians: Tracking by persistent dynamic view synthesis,” inInternational Conference on 3D Vision (3DV), 2024, pp. 800–809

    2024