Pith. sign in

REVIEW 3 major objections 6 minor 42 references

GeMS: Efficient Gaussian Splatting for Extreme Motion Blur

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A Gaussian-splatting pipeline reconstructs sharp 3D scenes from severely motion-blurred images alone.

desk verdict A practical and largely convincing blur-robust 3DGS pipeline with a real novelty (learned SfM initialization), held back mainly by an unvalidated trajectory model. read the letter →

arxiv 2508.14682 v1 pith:P5DGOCJN submitted 2025-08-20 cs.CV

Gopi Raju Matta , Trisha Reddypalli , Vemunuri Divya Madhuri , Kaushik Mitra This is my paper
classification cs.CV
keywords 3DGaussianSplattingmotiondeblurringnovelviewsynthesiscameraposeestimationeventcamerasstructurefromextremeblurprobabilisticscenerefinement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

GeMS tries to solve a practical deadlock: reconstructing a sharp 3D scene from photographs taken with severe camera motion, where ordinary pose-estimation tools fail because blur destroys the features they match. The paper claims that by replacing brittle feature-matching pose estimation with a learned differentiable structure-from-motion module [4], treating Gaussians as samples of a scene distribution through Markov-chain Monte Carlo Gaussian splatting [5], and jointly optimizing the camera's exposure-time trajectory together with the Gaussians under a physical blur-formation loss, a 3D Gaussian scene can be recovered from blurred inputs alone. When event data is available, event-based double-integral deblurring [7] produces sharp enough images to initialize the same pipeline, and the paper reports this variant outperforms event-based radiance-field baselines and even sharp-supervised methods. Because the pipeline is built on Gaussian splatting, it also runs in minutes rather than hours, which is what would make the approach usable in robotics, handheld photography, and other real settings.

What carries the argument

The load-bearing mechanism is a differentiable blur-formation loop. Latent sharp frames are rendered from the Gaussian scene at virtual camera poses sampled along an SE(3) Bezier trajectory, and their average approximates the observed motion-blurred image; an L1 plus D-SSIM loss between the synthesized and real blur propagates gradients into both the Gaussian parameters and the trajectory control points. Two stabilizers make the loop feasible under extreme blur: a learned differentiable structure-from-motion module [4] supplies poses and point clouds that survive blur, and MCMC-based Gaussian sampling [5] replaces heuristic densification with probabilistic refinement. The named object that c

What would settle it

Generate test sequences whose blur comes from trajectories deliberately outside the paper's trajectory family—for example high-frequency jitter or a rapid segmentation of motion during one exposure—by rendering known scenes through those paths. If GeMS's reconstruction quality drops sharply on such inputs, then the method's extreme-blur claim is bounded by the assumed smooth product-of-exponentials camera path.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that severe motion blur does not need to be treated as a failure case for 3D reconstruction. It presents GeMS, a Gaussian-splatting pipeline that takes only severely motion-blurred images and produces a sharp scene plus refined camera trajectories, with no sharp reference images and no classical structure-from-motion initialization. The argument is mechanistic: each blurry image is the integral of latent sharp images along the camera's motion during exposure, so if the Gaussians and trajectory are right, rendering the Gaussians at sampled poses along a learned Bezier curve and averaging them reproduces the observed blur; the residual between tha

Load-bearing premise

The load-bearing premise is that each blurred image is produced by a camera following one smooth pose path of the exact parametric form in Eq. (8); if real camera motion during exposure deviates from that form, the synthesized blur used for supervision is systematically biased.

Editorial extensions

If this is right

  • If the claims hold, handheld and vehicle-mounted cameras can be used for 3D reconstruction under long exposures without a sharp reference frame or a separate pose-estimation step.
  • The same joint trajectory-geometry optimization should make deblurring and novel view synthesis stable under increasingly severe blur, since each module corrects the other's errors.
  • When event sensors are present, event data can serve as an initialization aid rather than as supervision, lowering the hardware and data requirements for event-based reconstruction.
  • Because the pipeline is Gaussian-splatting-based, the reported order-of-magnitude speed and memory savings over radiance-field baselines would make blurred-input reconstruction practical on a single GPU.
  • The released synthetic event dataset would let other deblurring and reconstruction methods be evaluated under controlled extreme blur.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same integral-blur loop may extend to rolling-shutter distortion, defocus blur, or scene motion during exposure, since those are also temporal integrals over a latent sharp sequence, though each needs its own forward model.
  • Editorial inference: the product-of-exponentials trajectory in Eq. (8) is not the standard SE(3) Bezier blend (exponentials add only for commuting motions), so the reported advantage over linear and spline paths should be tested on independently measured real camera trajectories before generalizing.
  • Editorial inference: if the learned structure-from-motion module is the main source of blur tolerance, then pairing it with a blur-aware tracker could narrow the gap between the no-event and event-assisted variants.
  • Editorial inference: the comparison suggests event data may be most valuable at the initialization stage; a testable extension is to feed other blur-tolerant pose estimators into the same MCMC-Gaussian loop and measure how much of GeMS-E's gain comes from initialization versus joint optimization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes GeMS, a 3D Gaussian Splatting framework that reconstructs sharp 3D scenes directly from severely motion-blurred images without COLMAP or sharp-image supervision. It combines VGGSfM for blur-robust SfM initialization, 3DGS-MCMC for probabilistic Gaussian densification, and joint optimization of camera trajectories and Gaussian parameters using a physics-based blur formation model. GeMS-E extends this with event-based EDI deblurring, using the deblurred images only for initialization while supervising with the original blurry images. Experiments on synthetic ExBluRF and real E2NeRF datasets report state-of-the-art PSNR/SSIM/LPIPS, large speedups over NeRF baselines, and introduce a synthetic event dataset EveGeMS.

Significance. If the central claims hold, this is a practically important step: it removes the unrealistic sharp-image/COLMAP initialization requirement of existing deblurring pipelines and extends 3DGS to extreme motion blur. The modular ablations (Table II, Figures 11-12) are systematic, and the reported efficiency gains (Tables V-VI) are striking. The EveGeMS dataset is also a useful community resource. The main weaknesses are mathematical: the SE(3) trajectory parameterization in Eq. (8) is not justified, the gradient derivation in Eq. (11) is formally incorrect, and the sharp-supervised comparison is not apples-to-apples. These issues are fixable and do not, in my view, invalidate the empirical contribution, but they need to be addressed before the paper can be accepted.

major comments (3)
  1. [IV.A.3, Eq. (8)] Eq. (8) defines T_t as a product of exponentials exp(w_j log T_j). For non-commuting SE(3) elements this is not the standard Bézier interpolation, and no derivation or reference supports this specific form. The blur-formation model and the joint optimization rely on this parameterization, but Fig. 8 only compares capacity (linear vs spline vs Bézier), not fidelity to true camera motion. On the ExBluRF synthetic scenes, ground-truth sharp sub-frame poses are available; please quantify the pose error of the fitted trajectory against those poses, and state how the control points T_j are initialized/constrained relative to the VGGSfM poses. Without such a test, the claimed advantage over sharp-supervised baselines may be due to extra degrees of freedom rather than accurate motion modeling.
  2. [IV.A.5, Eq. (11)] Eq. (11) is not a correct chain rule for the forward model: C_i is rendered from θ and T_i, so the gradient of L w.r.t. T must include the direct term ∂B_k/∂C_i ∂C_i/∂T. The factor ∂θ/∂T is not present in the computation graph (θ and T are jointly optimized independent variables); if it is meant as an implicit-function correction, that requires derivation. Please correct Eq. (11) and describe what the autodiff implementation actually computes. This is load-bearing because the joint trajectory-geometry optimization is a central contribution.
  3. [V.B, Table I] The headline claim of a 1 dB PSNR gain over ExBluRF*/BAD-Gaussians* compares GeMS-E, which receives event streams, with methods that receive only RGB sharp images for initialization. Events contain temporal information that sharp stills do not, so the comparison conflates input modality with supervision. Add a controlled ablation feeding sharp images (or EDI-deblurred images) into the same initialization and optimization pipeline to isolate the source of the gain. Also, all quantitative results are single-run; please report mean±std over multiple seeds, since several gaps in Table I are below 0.5 dB.
minor comments (6)
  1. [References [2] and [27]] These two references are the same 3D Gaussian Splatting paper with different metadata; please consolidate to avoid duplication.
  2. [Eq. (13)] The notation for event bins is inconsistent (t', τ', and the index set need cleanup). Please define all quantities explicitly.
  3. [Table II] The row 'w/ MCMC + w/ EDI + w/ COLMAP' shows COLMAP succeeding after EDI deblurring, which seems to contradict the paper's statement that COLMAP fails under severe blur. Please clarify the exact input conditions under which COLMAP is run in this ablation.
  4. [Figure 8] The number of virtual cameras and the trajectory representation are varied in the same plot, making it hard to isolate each effect. Consider separate plots for n and for the trajectory model.
  5. [Section V.B] The baseline 'EDI+3DGS' is not described in the experimental setup: what SfM/initialization does it use, and how is it trained? Please clarify.
  6. [Section V.C] Real-world comparisons are qualitative only. A no-reference or event-based quantitative metric would strengthen the practical claim, or the authors should explicitly state why no metric is applicable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; GeMS/GeMS-E is an empirical pipeline evaluated against external baselines and does not reduce to its inputs by construction.

full rationale

I walked the claimed derivation chain: GeMS composes VGGSfM for SfM initialization, 3DGS-MCMC for Gaussian optimization, and a joint trajectory-geometry optimization using a physics-based blur formation model (Eqs. 6-8), then compares against external sharp-supervised and event-based baselines on ExBluRF synthetic data and E2NeRF real data. No step defines a prediction in terms of the quantity it is supposed to predict, and no fitted parameter is renamed as a prediction. The trajectory parameterization in Eq. (8) is an unvalidated ansatz and Eq. (11) contains an incorrect chain rule, but these are correctness risks, not circularity: the blur model depends on the trajectory assumption, yet the assumption is not derived from or equivalent to the output, and it is directly testable against ground-truth sub-frame poses. All load-bearing components (VGGSfM, 3DGS-MCMC, EDI, BAD-Gaussians) are external prior work with independent provenance; there is no self-citation chain and no uniqueness theorem imported from the authors. The only author-created asset, EveGeMS, is generated with ESIM from ground-truth sharp frames, which is standard practice and does not collapse any claim, especially since the same findings hold on the real E2NeRF dataset. Therefore the central claims are empirically supported rather than definitionally forced.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical, so the load-bearing inputs are modeling assumptions and hand-set hyperparameters rather than fitted physical constants. The main concern is the non-standard trajectory interpolation in Eq. (8), which is stated without proof and directly affects the blur synthesis used for supervision.

free parameters (6)
  • Number of virtual camera poses n = 15
    Chosen by ablation in Section V.D.3 (Figure 8) over {5, 10, 15, 20} to balance blur modeling and efficiency.
  • Bezier degree M and control point count = 9 control points
    Set in implementation details; no ablation reported for the number of control points, so it is chosen by hand.
  • Number of event bins for EDI = 13
    Set in implementation details; no ablation reported for this value.
  • Camera pose learning rate = 1e-3
    Set in implementation details; chosen by hand, different from the original 3DGS learning rate.
  • Number of training iterations = 7000
    Fixed for all experiments and comparisons; chosen by hand.
  • Data factor = 2
    Downsampling factor for training data; chosen by hand.
assumptions (4)
  • domain assumption Blur formation is a pure integral of latent sharp images along a single camera trajectory with known exposure time (Eq. 6-7).
    This ignores depth-dependent blur, rolling shutter, and unknown exposure time. It is used directly in the loss for both GeMS and GeMS-E.
  • domain assumption VGGSfM produces reliable camera poses and point clouds from severely motion-blurred images.
    Empirically validated on synthetic blur levels in Section V.D.1, but the real-world extreme cases rely on this property; if VGGSfM fails, GeMS without events loses its initialization.
  • domain assumption 3DGS-MCMC's SGLD sampling converges to a good scene distribution from noisy initialization.
    Adopted from [5]; the paper's robustness experiments (Figure 11) support this for the tested blur levels.
  • ad hoc to paper The product-of-exponentials formula in Eq. (8) is a valid interpolation of the camera trajectory.
    This formula is introduced in this paper and is not the standard SE(3) Bezier interpolation; no derivation or citation is provided for this specific form.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GeMS: Efficient Gaussian Splatting for Extreme Motion Blur." pith.science (2026). https://pith.science/paper/P5DGOCJN

@misc{pith2026250814682,
  author       = {Pith},
  title        = {Pith review of: GeMS: Efficient Gaussian Splatting for Extreme Motion Blur},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P5DGOCJN}},
  note         = {Machine review of arXiv:2508.14682}
}
read the original abstract

We introduce GeMS, a framework for 3D Gaussian Splatting (3DGS) designed to handle severely motion-blurred images. State-of-the-art deblurring methods for extreme blur, such as ExBluRF, as well as Gaussian Splatting-based approaches like Deblur-GS, typically assume access to sharp images for camera pose estimation and point cloud generation, an unrealistic assumption. Methods relying on COLMAP initialization, such as BAD-Gaussians, also fail due to unreliable feature correspondences under severe blur. To address these challenges, we propose GeMS, a 3DGS framework that reconstructs scenes directly from extremely blurred images. GeMS integrates: (1) VGGSfM, a deep learning-based Structure-from-Motion pipeline that estimates poses and generates point clouds directly from blurred inputs; (2) 3DGS-MCMC, which enables robust scene initialization by treating Gaussians as samples from a probability distribution, eliminating heuristic densification and pruning; and (3) joint optimization of camera trajectories and Gaussian parameters for stable reconstruction. While this pipeline produces strong results, inaccuracies may remain when all inputs are severely blurred. To mitigate this, we propose GeMS-E, which integrates a progressive refinement step using events: (4) Event-based Double Integral (EDI) deblurring restores sharper images that are then fed into GeMS, improving pose estimation, point cloud generation, and overall reconstruction. Both GeMS and GeMS-E achieve state-of-the-art performance on synthetic and real-world datasets. To our knowledge, this is the first framework to address extreme motion blur within 3DGS directly from severely blurred inputs.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 32 canonical work pages

  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,” Aug. 2020

    work page 2020

  2. [2]

    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 TOG, vol. 42, no. 4, July 2023. [Online]. Available: https: //repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    work page 2023

  3. [3]

    Structure-from-motion Revisited,

    J. L. Schonberger and J.-M. Frahm, “Structure-from-motion Revisited,” inCVPR, 2016. [Online]. Available: https://github.com/colmap/colmap

    work page 2016

  4. [4]

    Vggsfm: Visual geometry grounded deep structure from motion,

    J. Wang, N. Karaev, C. Rupprecht, and D. Novotny, “Vggsfm: Visual geometry grounded deep structure from motion,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21 686–21 697

    2024

  5. [5]

    3d gaussian splatting as markov chain monte carlo,

    S. Kheradmand, D. Rebain, G. Sharma, W. Sun, Y .-C. Tseng, H. Isack, A. Kar, A. Tagliasacchi, and K. M. Yi, “3d gaussian splatting as markov chain monte carlo,”Advances in Neural Information Processing Systems, vol. 37, pp. 80 965–80 986, 2025

    work page 2025

  6. [6]

    BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting,

    L. Zhao, P. Wang, and P. Liu, “BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting,” inECCV. Springer, 2024

    work page 2024

  7. [7]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 6820–6829

    work page 2019

  8. [8]

    Global structure-from-motion revisited,

    L. Pan, D. Bar ´ath, M. Pollefeys, and J. L. Sch ¨onberger, “Global structure-from-motion revisited,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 58–77

    2024

Show all 42 references
  1. [9]

    From coarse to fine: Robust hierarchical localization at large scale,

    P.-E. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk, “From coarse to fine: Robust hierarchical localization at large scale,” inCVPR, 2019

    2019

  2. [10]

    SuperGlue: Learning feature matching with graph neural networks,

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “SuperGlue: Learning feature matching with graph neural networks,” inCVPR, 2020

    2020

  3. [11]

    Pixel- perfect structure-from-motion with featuremetric refinement,

    P. Lindenberger, P.-E. Sarlin, V . Larsson, and M. Pollefeys, “Pixel- perfect structure-from-motion with featuremetric refinement,” inPro- ceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 5987–5997

    2021

  4. [12]

    Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction,

    J. Reizenstein, R. Shapovalov, P. Henzler, L. Sbordone, P. Labatut, and D. Novotny, “Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 901–10 911

    2021

  5. [13]

    Image matching across wide baselines: From paper to practice,

    Y . Jin, D. Mishkin, A. Mishchuk, J. Matas, P. Fua, K. M. Yi, and E. Trulls, “Image matching across wide baselines: From paper to practice,”International Journal of Computer Vision, vol. 129, no. 2, pp. 517–547, 2021

    2021

  6. [14]

    A multi-view stereo benchmark with high- resolution images and multi-camera videos,

    T. Schops, J. L. Schonberger, S. Galliani, T. Sattler, K. Schindler, M. Pollefeys, and A. Geiger, “A multi-view stereo benchmark with high- resolution images and multi-camera videos,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 3260– 3269

    2017

  7. [15]

    Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,

    J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan, “Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,” inCVPR, 2021

    2021

  8. [16]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-nerf 360: Unbounded anti-aliased neural radiance fields,” inCVPR, 2022

    2022

  9. [17]

    Alignerf: High-fidelity neural radiance fields via alignment- aware training,

    Y . Jiang, P. Hedman, B. Mildenhall, D. Xu, J. T. Barron, Z. Wang, and T. Xue, “Alignerf: High-fidelity neural radiance fields via alignment- aware training,”CVPR, 2023

    2023

  10. [18]

    Diver: Real- time and accurate neural radiance fields with deterministic integration for volume rendering,

    L. Wu, J. Y . Lee, A. Bhattad, Y . Wang, and D. Forsyth, “Diver: Real- time and accurate neural radiance fields with deterministic integration for volume rendering,” 2022. 14

    2022

  11. [19]

    NeRF++: Analyzing and improving neural radiance fields,

    K. Zhang, G. Riegler, N. Snavely, and V . Koltun, “NeRF++: Analyzing and improving neural radiance fields,”arXiv:2010.07492, 2020

    2010 arXiv

  12. [20]

    Autoint: Automatic integration for fast neural volume rendering,

    D. B. Lindell, J. N. Martel, and G. Wetzstein, “Autoint: Automatic integration for fast neural volume rendering,” inCVPR, 2021

    2021

  13. [21]

    Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,

    C. Reiser, S. Peng, Y . Liao, and A. Geiger, “Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,” inICCV, 2021

    2021

  14. [22]

    Plenoctrees for real-time rendering of neural radiance fields,

    A. Yu, R. Li, M. Tancik, H. Li, R. Ng, and A. Kanazawa, “Plenoctrees for real-time rendering of neural radiance fields,” inICCV, 2021

    2021

  15. [23]

    Plenoxels: Radiance fields without neural networks,

    S. Fridovich-Keil, A. Yu, M. Tancik, Q. Chen, B. Recht, and A. Kanazawa, “Plenoxels: Radiance fields without neural networks,” in CVPR, 2022

    2022

  16. [24]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”TOG, 2022

    2022

  17. [25]

    Tensorf: Tensorial radiance fields,

    A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su, “Tensorf: Tensorial radiance fields,” inECCV, 2022

    2022

  18. [26]

    Direct voxel grid optimization: Super- fast convergence for radiance fields reconstruction,

    C. Sun, M. Sun, and H.-T. Chen, “Direct voxel grid optimization: Super- fast convergence for radiance fields reconstruction,”CVPR, 2022

    2022

  19. [27]

    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,”TOG, 2023

    2023

  20. [28]

    Optical models for direct volume rendering,

    N. Max, “Optical models for direct volume rendering,”IEEE Transac- tions on Visualization and Computer Graphics, pp. 99–108, 1995

    1995

  21. [29]

    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,” inCVPR,

  22. [30]

    Exblurf: Efficient radiance fields for extreme motion blurred images,

    D. Lee, J. Oh, J. Rim, S. Cho, and K. M. Lee, “Exblurf: Efficient radiance fields for extreme motion blurred images,” inICCV, 2023

    2023

  23. [31]

    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,” inCVPR, 2023. [Online]. Available: https://wangpeng000.github.io/BAD-NeRF/

    2023

  24. [32]

    Deblur-gs: 3d gaussian splatting from camera motion blurred images,

    W. Chen and L. Liu, “Deblur-gs: 3d gaussian splatting from camera motion blurred images,”Proceedings of the ACM on Computer Graphics and Interactive Techniques, vol. 7, no. 1, pp. 1–15, 2024

    2024

  25. [33]

    E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images,

    Y . Qi, L. Zhu, Y . Zhang, and J. Li, “E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 13 254–13 264

    2023

  26. [34]

    Deblurring neural radiance fields with event-driven bundle adjustment,

    Y . Qi, L. Zhu, Y . Zhao, N. Bao, and J. Li, “Deblurring neural radiance fields with event-driven bundle adjustment,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 9262–9270

    2024

  27. [35]

    E2gs: Event enhanced gaussian splatting,

    H. Deguchi, M. Masuda, T. Nakabayashi, and H. Saito, “E2gs: Event enhanced gaussian splatting,” in2024 IEEE International Conference on Image Processing (ICIP). IEEE, 2024, pp. 1676–1682

    2024

  28. [36]

    ESIM: An Open Event Camera Simulator,

    H. Rebecq, D. Gehrig, and D. Scaramuzza, “ESIM: An Open Event Camera Simulator,” inProceedings of The 2nd Conference on Robot Learning. PMLR, Oct. 2018, pp. 969–982

    2018

  29. [37]

    Multi-stage Progressive Image Restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Multi-stage Progressive Image Restoration,” inCVPR,

  30. [38]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inCVPR, 2022

    2022

  31. [39]

    Pytorch: An Imperative Style, High-performance Deep Learning Library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antigaet al., “Pytorch: An Imperative Style, High-performance Deep Learning Library,”Advances in neural information processing systems, vol. 32, 2019. [Online]. Available: ht...

    2019

  32. [40]

    gsplat: An open-source library for gaussian splatting,

    V . Ye, R. Li, J. Kerr, M. Turkulainen, B. Yi, Z. Pan, O. Seiskari, J. Ye, J. Hu, M. Tanciket al., “gsplat: An open-source library for gaussian splatting,”Journal of Machine Learning Research, vol. 26, no. 34, pp. 1–17, 2025

    2025

  33. [2021]

    Available: https://github.com/swz30/MPRNet

    [Online]. Available: https://github.com/swz30/MPRNet

  34. [2022]

    Available: https://limacv.github.io/deblurnerf/

    [Online]. Available: https://limacv.github.io/deblurnerf/

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.