Pith. sign in

REVIEW 3 major objections 5 minor 43 references

Blind 3D reconstruction from extreme motion blur is possible by coupling deep topology recovery, probabilistic densification, and continuous camera trajectories.

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 ↗

PRISM3D bootstraps 3D Gaussian Splatting from extreme motion blur via VGGSfM initialization, MCMC densification, and Bézier trajectories, with an event-assisted extension that sets new SOTA.

T0 review reviewed 2026-07-11 challenge →

load-bearing objection Solid engineering that actually closes the extreme-blur 3DGS gap; the VGGSfM residual-error worry is real but secondary to the empirical gains and ablations. the 3 major comments →

arxiv 2607.03855 v1 pith:LQBSHQKL submitted 2026-07-04 cs.CV

PRISM3D: Probabilistic Refinement and Robust Initialization for Physically Consistent Scene Modeling under Extreme Motion Blur

Gopi Raju Matta , Reddypalli Trisha , Vemunuri Divya Madhuri , Kaushik Mitra This is my paper
classification cs.CV
keywords blind 3D reconstructionextreme motion blur3D Gaussian Splattingdeep Structure-from-MotionMarkov Chain Monte Carlo densificationBézier camera trajectoriesevent-assisted deblurringPRISM3D-E Benchmark
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.

The reading

Extreme camera shake turns ordinary photographs into low-pass smears that destroy the high-frequency corners and edges classical Structure-from-Motion needs, so conventional 3D Gaussian pipelines never even start. PRISM3D claims that a deep dense tracker can still recover a sparse but globally consistent skeleton of poses and points; that skeleton is then densified by treating Gaussians as samples from a scene distribution and updating them with stochastic gradient Langevin dynamics, while the blur itself is explained by integrating radiance along continuous SE(3) Bézier trajectories. The same pipeline accepts optional event streams only to improve the initial skeleton, never for photometric supervision. The result is a fully blind, event-agnostic method that reconstructs sharp novel views from inputs so degraded that classical SfM fails completely, together with a new multi-modal benchmark that pairs extreme blur with synchronized events.

Core claim

A sparse topological skeleton recovered by deep dense tracking is already sufficient to bootstrap high-fidelity 3D Gaussian Splatting from extreme motion blur, provided the subsequent densification is performed probabilistically via MCMC and the image-formation model uses continuous Bézier trajectories rather than linear or cubic splines.

What carries the argument

The coupled triple of VGGSfM deep topology recovery, 3DGS-MCMC densification under Stochastic Gradient Langevin Dynamics, and continuous SE(3) Bézier exposure paths that jointly optimize geometry and camera momentum against the observed blurry images.

Load-bearing premise

That an off-the-shelf deep tracker trained on ordinary sharp images still yields a usable topological skeleton even when the input frames have been averaged across as many as eleven consecutive sharp frames.

What would settle it

Run the identical VGGSfM-plus-MCMC-plus-Bézier pipeline on a new set of real extreme-blur sequences whose ground-truth sharp geometry is known independently; if registration collapses or novel-view PSNR falls below that of a pure event-based baseline, the claim that the sparse skeleton is sufficient is falsified.

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

Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. PRISM3D addresses blind 3D scene reconstruction from extremely motion-blurred images by coupling three components: (i) robust initialization via VGGSfM deep dense tracking to recover global topology where classical SfM fails, (ii) probabilistic densification of the resulting sparse/noisy geometry via 3DGS-MCMC (SGLD), and (iii) continuous SE(3) Bézier trajectory modeling of the physical blur formation process. The authors further introduce PRISM3D-E, which uses event streams only to synthesize EDI pseudo-sharp priors for VGGSfM initialization (not photometric supervision), and contribute the PRISM3D-E Benchmark of extreme-blur RGB paired with synthetic events. Experiments on eight synthetic scenes and five real DAVIS346 sequences, with systematic blur-level sweeps (1–11 frames), trajectory ablations, and component ablations, report new state-of-the-art novel-view and deblurring metrics for both the RGB-only and event-assisted pipelines, with strong efficiency claims (~7 min, ~1.55 GiB).

Significance. If the claims hold, the work removes a practical bottleneck that has forced prior 3DGS/NeRF deblurring methods to rely on sharp-image COLMAP oracles. The combination of deep topology recovery, MCMC densification, and continuous Bézier blur modeling is a coherent and well-motivated response to the initialization–optimization gap, and the PRISM3D-E Benchmark fills a genuine multi-modal evaluation hole. The systematic blur-severity analysis, trajectory-model ablations, and clear separation of oracle vs. blind baselines are strengths that make the empirical case falsifiable and useful to the community. The efficiency numbers further increase practical impact for robotics and handheld capture.

major comments (3)
  1. Sec. 4.4 / Tables 4–5: VGGSfM is shown to register all views where COLMAP collapses, but translation RMSE rises from 0.18 m (blur-3) to 0.72 m (blur-11). No rotation error, absolute trajectory error after BA, or final-vs-initial pose residual after joint MCMC+Bézier optimization is reported. Without these, it remains possible that SGLD densification and trajectory fitting largely compensate for residual pose error rather than converging to a physically correct scene. Please add (i) rotation/ATE statistics for VGGSfM under the same blur sweep and (ii) pose residual before vs. after joint optimization on the synthetic data where GT poses exist.
  2. Sec. 4.2 and real-data evaluation (Fig. 12 / Table 3): On the five real DAVIS346 sequences there is no independent pose ground truth and no quantitative novel-view metrics against held-out sharp views (only qualitative deblurring and efficiency). The central claim of physically consistent reconstruction under extreme real blur therefore rests almost entirely on visual inspection. At minimum, report multi-view photometric consistency on held-out real frames, or an external pose estimate (e.g., from a secondary sensor or COLMAP on EDI-deblurred frames) so that residual pose error can be quantified outside the synthetic regime.
  3. Sec. 3.1 / Eq. (3) and Sec. 4.5: The continuous SE(3) Bézier model is a core differentiator over linear/cubic splines, yet the degree M (and the choice of 9 control points) is fixed without a sensitivity study beyond the n-virtual-camera sweep in Fig. 8. Because the physical-consistency claim hinges on this parameterization, please report (or move from supplement) an ablation on control-point count / degree and confirm that the reported 0.24–0.59 dB gains over splines remain stable under modest changes of M.
minor comments (5)
  1. Table numbering is inconsistent: the main quantitative table is labeled Table 1 in the body but referred to as Table 8 in the text and later as Table 8 again in the supplement; figures similarly jump between Fig. 3/4 and Fig. 11/12. Please renumber for a single coherent sequence.
  2. Eq. (6) writes ∂L/∂T with an extra ∂Θ/∂T factor that appears to duplicate the chain already present in ∂Ci/∂Θ; clarify the intended Jacobian path (pose → latent images → blur) or correct the expression.
  3. Implementation details (Sec. 4.1) list free parameters (n=15, 9 Bézier points, pose lr 1e-3, 13 event bins, SGLD τ/k) that are only partially ablated; a short hyper-parameter table in the supplement would improve reproducibility.
  4. Related Work (Sec. 2.3) correctly distinguishes oracle vs. blind methods, but a brief note on concurrent event-assisted 3DGS works (EvaGaussians, E2GS) regarding whether they also fail under the 9–11-frame regime would strengthen the novelty claim.
  5. Fig. 1 caption and Algorithm 1 are helpful; ensure the algorithm’s “optional event stream” branch matches the main-text statement that EDI images are never used for photometric supervision.

Circularity Check

0 steps flagged

No circularity: empirical pipeline with external components, ablations, and independent GT evaluation; claims do not reduce to inputs by construction.

full rationale

PRISM3D couples three external, independently published modules (VGGSfM for sparse topology, 3DGS-MCMC/SGLD densification, continuous SE(3) Bézier trajectories) and optimizes them jointly against the observed blurry images via a standard photometric loss (Eqs. 4–6). All quantitative claims (PSNR/SSIM/LPIPS tables, robustness curves, ablations in Tables 2/6/7/10) are measured against held-out sharp ground-truth images or independent baselines (Restormer, E2NeRF, BAD-Gaussians*, etc.). Hyper-parameters (n=15 virtual cameras, 9 Bézier points, learning rates) are selected by explicit ablation rather than fitted to the final reported metric. No equation equates a reported quantity to a quantity defined from the same data; no uniqueness theorem or ansatz is imported from the authors’ prior work; the “first to bootstrap 3DGS from extreme blur” claim is an empirical novelty statement, not a derivation. The pipeline is therefore self-contained against external benchmarks and exhibits zero circular reduction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on three external algorithmic components treated as black boxes (VGGSfM, 3DGS-MCMC, EDI), a continuous-exposure image-formation model, and a handful of hand-chosen hyper-parameters. No new physical constants or particles are invented; the free parameters are ordinary optimization knobs whose values are justified by ablation rather than by fitting the final leaderboard metric.

free parameters (5)
  • number of virtual camera samples n = 15
    Chosen as n=15 after ablation over {5,10,15,20}; balances photometric accuracy against compute.
  • Bézier control-point count = 9
    Fixed at 9 control points for the SE(3) trajectory; not derived from first principles.
  • pose learning rate = 1e-3
    Set to 1e-3; Gaussian rates follow original 3DGS defaults.
  • event bins for EDI = 13
    13 bins used for PRISM3D-E event deblurring; chosen by the authors.
  • SGLD noise threshold τ and steepness k = τ=0.005, k=100
    Inherited from 3DGS-MCMC (τ=0.005, k=100) to damp noise on opaque Gaussians.
axioms (5)
  • domain assumption Motion blur is the temporal integral of latent sharp images along a continuous camera trajectory during exposure.
    Standard physical image-formation model (Eqs. 1–2); assumed throughout the loss derivation.
  • domain assumption VGGSfM recovers a sufficiently accurate global topology from extremely blurred inputs for subsequent densification.
    Empirically supported by Tables 4–5 but not proved; load-bearing for the whole pipeline.
  • domain assumption 3D Gaussians can be treated as samples from an underlying scene distribution updated by SGLD.
    Adopted wholesale from 3DGS-MCMC; enables densification of sparse VGGSfM priors.
  • ad hoc to paper Camera motion during exposure is well-approximated by a degree-M Bézier curve in SE(3).
    Chosen over linear and cubic splines after ablation (Fig. 8); not a universal physical law.
  • domain assumption Event streams generated by ESIM from latent sharp sequences are a faithful proxy for real event cameras under extreme blur.
    Used to construct the PRISM3D-E Benchmark; real DAVIS346 data is used only for qualitative/real-table validation.
invented entities (1)
  • PRISM3D-E Benchmark no independent evidence
    purpose: Paired extreme-blur RGB + synthetic event streams for multi-modal evaluation.
    New dataset contribution; independent evidence will exist once publicly released, currently promised but not yet available.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of PRISM3D: Probabilistic Refinement and Robust Initialization for Physically Consistent Scene Modeling under Extreme Motion Blur." pith.science (2026). https://pith.science/paper/LQBSHQKL

@misc{pith2026260703855,
  author       = {Pith},
  title        = {Pith review of: PRISM3D: Probabilistic Refinement and Robust Initialization for Physically Consistent Scene Modeling under Extreme Motion Blur},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQBSHQKL}},
  note         = {Machine review of arXiv:2607.03855}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We address the inverse problem of blind 3D scene reconstruction from extremely motion-blurred images, a scenario where traditional Structure-from-Motion (SfM) pipelines fail. Existing approaches typically circumvent this bottleneck by relying on impractical sharp-image supervision. In this work, we introduce PRISM3D, a unified framework enabling robust reconstruction directly from severely degraded inputs. To overcome the lack of a reliable starting point, we propose a Robust Initialization strategy utilizing deep dense tracking method (VGGSfM) to recover global topology where feature matching fails. To the best of our knowledge, we are the first to effectively leverage this paradigm to bootstrap 3D Gaussian Splatting from extreme motion blur. However, while robust, this initialization yields sparse and noisy geometry that causes deterministic optimization to diverge. To resolve this, we propose a coupled solution driven by probability and physics: we adopt a probabilistic formulation for geometric densification via Markov Chain Monte Carlo (MCMC) to robustly populate the sparse priors, while simultaneously modeling physical image formation via continuous Bezier Trajectories. Furthermore, while PRISM3D establishes a highly robust standalone pipeline, the availability of complementary event streams offers an opportunity to push the reconstruction fidelity further. To exploit this, we introduce PRISM3D-E, a multi-modal (RGB + Events) extension that seamlessly integrates high-temporal-resolution events as structural priors to maximize geometric recovery. Because existing datasets lack paired event streams under such severe degradation, we concurrently contribute the PRISM3D-E Benchmark to facilitate rigorous evaluation. Extensive experiments demonstrate that both our standalone RGB framework and its multi-modal extension establish new state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2607.03855 by Gopi Raju Matta, Kaushik Mitra, Reddypalli Trisha, Vemunuri Divya Madhuri.

Figure 1
Figure 1. Figure 1: Overview of the PRISM3D Framework. We address extreme motion blur through a unified inverse problem formulation. (Top Left) For the standalone PRISM3D pipeline, we input blurry images directly. For PRISM3D-E, we incorporate complementary event streams to generate high-frequency structural priors (EDI). (Top Right) These inputs feed into our Robust Initialization stage, which utilizes deep point tracking an… view at source ↗
Figure 2
Figure 2. Figure 2: The PRISM3D-E Benchmark: We introduce a synthetic event dataset designed for extreme motion blur. While RGB frames suffer from severe degradation, the complementary event streams preserve fine structural cues. Event Generation and Formulation Event cameras asynchronously trigger an event e(x, y, τ, p) when the logarithmic brightness change exceeds a threshold Θ: p_{x,y,\tau } = \begin {cases} -1,&\log (\ma… view at source ↗
Figure 3
Figure 3. Figure 3: Results on the Synthetic Dataset: PRISM3D [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results on the Real Dataset: PRISM3D reconstructs sharp and high-quality images from severely motion-blurred real-world inputs. In contrast, existing methods struggle with artifacts, noise, loss of fine details, and text degradation. Our framework effectively restores textures and structural consistency, as evident in the insets. modal extension preserve accurate color distributions and fine-grained textur… view at source ↗
Figure 5
Figure 5. Figure 5: Synthetic Motion Blur Levels: Images generated by averaging 1 (sharp), 3, 5, 7, 9, and 11 consecutive frames from an 11-image burst at each viewpoint, illustrating the increasing severity of motion blur used for evaluation. VGGSfM Robustness: To evaluate VGGSfM’s robustness, we simulated increasing blur severity by averaging 1, 3, 5, 7, 9, and 11 consecutive frames from sharp burst sequences ( [PITH_FULL_… view at source ↗
Figure 6
Figure 6. Figure 6: MCMC robustness to various blur-corrupted point cloud initializa￾tions: PSNR comparison of 3DGS-MCMC with 3DGS across various blur point cloud initializations obtained from VGGSfM [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Impact of trajectory representations and virtual camera count. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Novel view synthesis LPIPS (↓) of PRISM3D-E with and without continuous event loss under varying sparsity regimes. Left: absolute training views. Right: relative percentage [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The PRISM3D-E Benchmark. We augment the challenging, motion-blurred RGB sequences from ExBluRF [13] with synchronized synthetic event streams. While the high-magnitude blur in the RGB frames presents a complex optimization landscape, the high-temporal-resolution event data provides critical structural and motion priors. This multi-modal integration significantly enhances the robustness of the reconstructi… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative Comparisons on the Synthetic Dataset: [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Deblurring Results on the Real Dataset: Performance on in-the-wild captures featuring complex, unknown camera trajectories. As highlighted in the zoomed￾in insets, PRISM3D successfully restores illegible text and intricate textures. In contrast, existing methods tend to over-smooth the outputs or amplify real-world sensor noise [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Novel View Synthesis Results on the Synthetic Dataset: [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative ablation study on the PRISM3D-E Benchmark (Syn [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative ablation study on the Real Dataset. [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visualization of the reconstructed scene geometry for five real-world scenes. Each row corresponds to a different scene. From left to right, the columns show: (i) input image, (ii) rendering produced by PRISM3D, (iii) rendering produced by PRISM3D-E, (iv) depth map estimated by PRISM3D, and (v) depth map estimated by PRISM3D-E. The rendered views together with the corresponding depth maps demonstrate that… view at source ↗
Figure 17
Figure 17. Figure 17: Extended Visualizations on the Synthetic Dataset (Frame A): [PITH_FULL_IMAGE:figures/full_fig_p032_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Extended Visualizations on the Synthetic Dataset (Frame B): [PITH_FULL_IMAGE:figures/full_fig_p033_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Extended Visualizations on the Real Dataset (Frame A): [PITH_FULL_IMAGE:figures/full_fig_p034_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Extended Visualizations on the Real Dataset (Frame B): [PITH_FULL_IMAGE:figures/full_fig_p035_20.png] view at source ↗

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

43 extracted references · 2 linked inside Pith

  1. [1]

    In: CVPR (2021) 4

    Barron,J.T.,Mildenhall,B.,Tancik,M.,Hedman,P.,Martin-Brualla,R.,Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: CVPR (2021) 4

    2021

  2. [2]

    In: CVPR (2022) 4

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) 4

    2022

  3. [3]

    In: ECCV (2022) 4

    Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022) 4

    2022

  4. [4]

    Proceedings of the ACM on Computer Graphics and Interactive Techniques 7(1), 1–15 (2024) 4, 5, 6, 8

    Chen, W., Liu, L.: Deblur-gs: 3d gaussian splatting from camera motion blurred images. Proceedings of the ACM on Computer Graphics and Interactive Techniques 7(1), 1–15 (2024) 4, 5, 6, 8

    2024

  5. [5]

    In: 2024 IEEE International Conference on Image Processing (ICIP)

    Deguchi, H., Masuda, M., Nakabayashi, T., Saito, H.: E2gs: Event enhanced gaussian splatting. In: 2024 IEEE International Conference on Image Processing (ICIP). pp. 1676–1682. IEEE (2024) 4

    2024

  6. [6]

    In: CVPR (2022) 4

    Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022) 4

    2022

  7. [7]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Jang, H., Choi, D., Kim, D., Kang, W., Kim, M.H.: Splat-based 3d scene reconstruc- tion with extreme motion-blur. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 26425–26434 (2025) 4

    2025

  8. [8]

    CVPR (2023) 4

    Jiang, Y., Hedman, P., Mildenhall, B., Xu, D., Barron, J.T., Wang, Z., Xue, T.: Alignerf: High-fidelity neural radiance fields via alignment-aware training. CVPR (2023) 4

    2023

  9. [9]

    International Journal of Computer Vision129(2), 517–547 (2021) 4

    Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. International Journal of Computer Vision129(2), 517–547 (2021) 4

    2021

  10. [10]

    ACM TOG42(4) (July 2023),https: //repo-sam.inria.fr/fungraph/3d-gaussian-splatting/2

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG42(4) (July 2023),https: //repo-sam.inria.fr/fungraph/3d-gaussian-splatting/2

    2023

  11. [11]

    TOG (2023) 4, 8, 19

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. TOG (2023) 4, 8, 19

    2023

  12. [12]

    Advances in Neural Information Processing Systems37, 80965–80986 (2025) 2, 4, 5, 6, 8, 19

    Kheradmand, S., Rebain, D., Sharma, G., Sun, W., Tseng, Y.C., Isack, H., Kar, A., Tagliasacchi, A., Yi, K.M.: 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems37, 80965–80986 (2025) 2, 4, 5, 6, 8, 19

    2025

  13. [13]

    In: ICCV (2023) 4, 5, 8, 21

    Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: ICCV (2023) 4, 5, 8, 21

    2023

  14. [14]

    In: CVPR (2021) 4

    Lindell, D.B., Martel, J.N., Wetzstein, G.: Autoint: Automatic integration for fast neural volume rendering. In: CVPR (2021) 4

    2021

  15. [15]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Lindenberger, P., Sarlin, P.E., Larsson, V., Pollefeys, M.: Pixel-perfect structure- from-motion with featuremetric refinement. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5987–5997 (2021) 4

    2021

  16. [16]

    arXiv preprint arXiv:2509.25191 (2025) 2, 4, 13

    Liu, Y., Luo, C., Tang, Z., Peng, J., Zhang, Z.: Vggt-x: When vggt meets dense novel view synthesis. arXiv preprint arXiv:2509.25191 (2025) 2, 4, 13

    arXiv 2025

  17. [17]

    IEEE Transactions on Visu- alization and Computer Graphics pp

    Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visu- alization and Computer Graphics pp. 99–108 (1995).https://doi.org/10.1109/ 2945.4684004

    arXiv 1995

  18. [18]

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (Aug 2020).https://doi.org/10.48550/arXiv.2003.089342, 4 PRISM3D: Robust 3D Scene Modeling under Extreme Motion Blur 17

    Pith/arXiv arXiv doi:10.48550/arxiv.2003.089342 2020

  19. [19]

    TOG (2022) 4

    Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. TOG (2022) 4

    2022

  20. [20]

    In: European Conference on Computer Vision

    Pan, L., Baráth, D., Pollefeys, M., Schönberger, J.L.: Global structure-from-motion revisited. In: European Conference on Computer Vision. pp. 58–77. Springer (2024) 4

    2024

  21. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: 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. pp. 6820–6829 (2019) 3, 8

    2019

  22. [22]

    Advances in neural information processing systems32(2019),https://pytorch.org/8

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An Imperative Style, High- performance Deep Learning Library. Advances in neural information processing systems32(2019),https://pytorch.org/8

    2019

  23. [23]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Qi, Y., Zhu, L., Zhang, Y., Li, J.: E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13254–13264 (2023) 4, 8

    2023

  24. [24]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Qi, Y., Zhu, L., Zhao, Y., Bao, N., Li, J.: Deblurring neural radiance fields with event-driven bundle adjustment. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 9262–9270 (2024) 4, 8

    2024

  25. [25]

    In: Proceedings of The 2nd Conference on Robot Learning

    Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of The 2nd Conference on Robot Learning. pp. 969–982. PMLR (Oct 2018) 8

    2018

  26. [26]

    In: ICCV (2021) 4

    Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021) 4

    2021

  27. [27]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10901–10911 (2021) 4

    2021

  28. [28]

    In: CVPR (2019) 4

    Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: CVPR (2019) 4

    2019

  29. [29]

    In: CVPR (2020) 4

    Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020) 4

    2020

  30. [30]

    In: CVPR (2016), https://github.com/colmap/colmap2, 3, 5

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion Revisited. In: CVPR (2016), https://github.com/colmap/colmap2, 3, 5

    2016

  31. [31]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3260–3269 (2017) 4

    2017

  32. [32]

    CVPR (2022) 4

    Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CVPR (2022) 4

    2022

  33. [33]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025) 2, 4, 13

    2025

  34. [34]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Wang, J., Karaev, N., Rupprecht, C., Novotny, D.: Vggsfm: Visual geometry grounded deep structure from motion. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 21686–21697 (2024) 2, 4, 5, 6, 19

    2024

  35. [35]

    In: CVPR (2023),https://wangpeng000.github.io/BAD-NeRF/ 4 18 G

    Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: CVPR (2023),https://wangpeng000.github.io/BAD-NeRF/ 4 18 G. R. Matta et al

    2023

  36. [36]

    Wu, L., Lee, J.Y., Bhattad, A., Wang, Y., Forsyth, D.: Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering (2022) 4

    2022

  37. [37]

    Journal of Machine Learning Research26(34), 1–17 (2025) 8

    Ye, V., Li, R., Kerr, J., Turkulainen, M., Yi, B., Pan, Z., Seiskari, O., Ye, J., Hu, J., Tancik, M., et al.: gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research26(34), 1–17 (2025) 8

    2025

  38. [38]

    In: ICCV (2021) 4

    Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: ICCV (2021) 4

    2021

  39. [39]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yu, W., Feng, C., Li, J., Tang, J., Yang, J., Tang, Z., Cao, M., Jia, X., Yang, Y., Yuan, L., et al.: Evagaussians: Event stream assisted gaussian splatting from blurry images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 24780–24790 (2025) 4

    2025

  40. [40]

    In: CVPR (2022) 8

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) 8

    2022

  41. [41]

    In: CVPR (2021),https://github

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage Progressive Image Restoration. In: CVPR (2021),https://github. com/swz30/MPRNet8

    2021

  42. [42]

    arXiv:2010.07492 (2020) 4

    Zhang, K., Riegler, G., Snavely, N., Koltun, V.: NeRF++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020) 4

    Pith/arXiv arXiv 2010

  43. [43]

    In: ECCV

    Zhao, L., Wang, P., Liu, P.: BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. In: ECCV. Springer (2024) 3, 4, 5, 6, 7, 8 PRISM3D: Robust 3D Scene Modeling under Extreme Motion Blur 19 PRISM3D: Probabilistic Refinement and Robust Initialization for Physically Consistent Scene Modeling under Extreme Motion Blur — Supplementary Material — A Prelimin...

    2024

This paper was first reviewed by grok-4.5 on July 11, 2026.