Pith. sign in

REVIEW 6 major objections 6 minor 13 references

A splat-based pipeline can reconstruct dense, deblurred 3D scenes from RGB-D video that is too blurry for standard pose estimation.

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

2026-08-01 19:32 UTC pith:MU4LRYXG

load-bearing objection Useful RGB-D integration that removes COLMAP from splat-based deblurring, but the robustness claim needs an initialization sensitivity study and the evaluation is thinner than it should be. the 6 major comments →

arxiv 2607.16926 v1 pith:MU4LRYXG submitted 2026-07-18 cs.CV

Splat-based 3D Scene Reconstruction with Extreme Motion-blur

Hyeonjoong Jang , Dongyoung Choi , Donggun Kim , Woohyun Kang , Min H. Kim This is my paper
classification cs.CV
keywords 3D reconstructionmotion blurRGB-DGaussian splattingcamera pose estimationdepth alignmentimage deblurringSLAM
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.

Motion blur from low light and fast camera movement normally breaks 3D reconstruction because pose estimation from sharp feature matches cannot start. This paper aims to establish that an RGB-D video with extreme blur is sufficient on its own: rough relative poses come from combining optical flow with depth alignment, and then a dense cloud of fixed-scale 3D Gaussians seeded from the depth maps is refined together with all camera poses by comparing rendered depth to measured depth. Once poses and geometry are globally consistent, the paper models blur as the average of sharp virtual exposures along the camera's path during each frame's exposure time, letting the same framework deblur the texture. If right, this means dim-light handheld capture can produce sharp, dense, globally consistent scene reconstructions without sharp frames or an external pose initializer, and the paper backs this with a new synthetic and real RGB-D dataset of extreme motion blur plus ablations showing the depth-guided refinement and fixed Gaussian scale are the components that matter.

Core claim

The paper's claim is that extreme motion blur in an RGB-D sequence does not have to defeat 3D reconstruction. The authors argue that by first recovering rough camera motion between consecutive frames from a combination of optical flow and depth alignment, then refining all camera poses and the 3D geometry jointly inside a splat-based scene representation where 3D Gaussians are initialized densely from the depth maps and optimized against a rendered-depth loss, one obtains a globally consistent point cloud and camera trajectory even when no sharp color frame is available. They then treat each blurred frame as the average of sharp renders from a set of virtual camera poses interpolated across

What carries the argument

The central object is a scene representation built from 3D Gaussian splats: a dense cloud of small Gaussian blobs, each with a position, opacity, and fixed scale, initialized from the input depth maps and rendered by rasterization. The argument runs through three coupled mechanisms. First, local pose initialization: optical flow is compared with pixels reprojected through the depth maps, plus a depth-consistency term, before point-to-plane alignment of the two point clouds. Second, global pose and geometry refinement: camera poses, Gaussian positions, and opacities are jointly optimized against a rendered-depth-versus-measured-depth loss, with ambiguous Gaussians pruned, so the whole sequenc

Load-bearing premise

The load-bearing premise is that the initial per-frame poses computed by optical flow plus depth alignment are close enough to the true camera trajectory that the later depth-only global refinement can converge to it; if the starting trajectory is far off, the refinement has nothing to recover from, and the paper gives no sensitivity analysis for that situation.

What would settle it

Take a synthetic RGB-D sequence with known camera trajectory and blur; perturb the initial relative pose estimates from the flow-plus-depth stage by rotations of a few degrees or translations of a few percent of scene scale, then check whether the depth-supervised Gaussian refinement still returns to the true trajectory. The central claim predicts graceful recovery from moderate initialization error; a sharp failure beyond a small perturbation would mark the method as initialization-dependent despite the joint refinement.

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

If this is right

  • Dense, globally consistent scenes can be reconstructed from RGB-D input without any sharp color frame or precomputed initial poses, since local alignment uses flow plus depth and global refinement uses only rendered depth.
  • Deblurring is tied to geometry: the blurred frame is explained as the average of sharp renders from virtual poses along the exposure path, so the restored texture is consistent with the recovered point cloud instead of being a blind per-image restoration.
  • Depth-supervised joint optimization of Gaussian positions and camera poses reduces drift and loop-closure artifacts that accumulate when camera motion is estimated only from consecutive frame pairs.
  • Keeping the Gaussian scale fixed is what makes depth measurements informative for pose; allowing scale to float weakens pose accuracy and hurts deblurring, as the ablations show.
  • The method produces denser and better-aligned point clouds than comparison SLAM and reconstruction methods under extreme blur, and those point clouds translate directly into higher novel-view deblur quality.

Where Pith is reading between the lines

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

  • The paper does not report how much error in the initial flow-plus-depth poses the global refinement tolerates; a controlled noise experiment would map that basin and is the clearest untested threat to the method's generality.
  • The depth-only refinement is invariant to shifting every pose and every Gaussian by one rigid transform, so a globally wrong initial trajectory cannot be repaired; adding an inertial prior or a coarse global reference is a natural extension the paper leaves implicit.
  • The exposure model assumes one linear camera path per frame; extending it to rolling-shutter or hand-jitter patterns would require higher-order exposure trajectories and is not covered by the reported experiments.
  • The initial alignment depends on a learned optical-flow network; whether the pipeline survives blur patterns outside that network's training distribution is open, and a cross-dataset blur test would directly probe it.

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

6 major / 6 minor

Summary. This paper addresses 3D scene reconstruction from RGB-D sequences under extreme motion blur. The pipeline consists of three stages: (i) local pose estimation between consecutive frames by optimizing optical-flow, photometric, and depth-consistency losses and then applying point-to-plane ICP; (ii) global refinement of poses and geometry by depth-supervised training of fixed-scale 3D Gaussians with an opacity regularizer; (iii) deblurring by optimizing start/end exposure poses and comparing the input image with the average of virtual views rendered along the interpolated exposure trajectory. The method is evaluated on three synthetic scenes with ground-truth poses and on real Azure Kinect sequences, with code and dataset to be released.

Significance. If the claims hold, the method would fill a practical need: RGB-D reconstruction in low-light and fast-motion conditions where COLMAP and feature-based methods fail. The depth-only global refinement with fixed-scale Gaussians is an interesting idea, and the planned public release of code and dataset is a clear positive. However, the evidence is currently limited: only three synthetic scenes are evaluated quantitatively, no error bars or multiple runs are reported, the most relevant deblurring baselines (BAD-Gaussians, Deblur-GS) are compared only qualitatively, and the pose-metric tables are inconsistent. The robustness claim regarding initialization is not demonstrated. The significance is therefore conditional on additional validation.

major comments (6)
  1. [Sec. 3.2 / Table 3] The global refinement (Eq. 5) uses only depth and opacity terms and is non-convex, yet the paper demonstrates it only from near-accurate initial poses: Table 3's 'w/o refinement' row shows ATE 0.123 m, RPE (trans) 0.007 m, RPE (rot) 0.047 deg on Bedroom. No sensitivity analysis is provided for larger initial pose errors, and no comparison to ground-truth initialization is given. Since the abstract and introduction claim that a precise initial pose is not required, this is a load-bearing gap. The authors should perturb the Sec. 3.1 poses to measure the basin of convergence and discuss the textureless-case failure they concede in the Supplementary (Fig. 6).
  2. [Sec. 4.1 / Supplementary Sec. 4] The quantitative evaluation does not include the most relevant deblurring baselines: BAD-Gaussians [49] and Deblur-GS [45] appear only in qualitative comparisons (Supplementary Fig. 3). To support the claim of outperforming existing deblurring approaches, the paper should report PSNR/SSIM/LPIPS for these methods on the synthetic scenes, e.g., by initializing them with ground-truth poses or with the proposed estimated poses. Without this, the central deblurring claim is not properly tested.
  3. [Table 1 vs Table 3 vs Supplementary Table 2] The pose metrics for what appears to be the same configuration are inconsistent. For Bedroom with the full pipeline, Table 1 reports ATE 0.102 m, RPE (trans) 0.006 m, RPE (rot) 0.092 deg; Table 3 reports 0.103, 0.012, 0.195; Supplementary Table 2 reports 0.049, 0.004, 0.057. The authors should clarify which numbers are correct, whether the evaluation protocol differs (e.g., trajectory alignment method), and correct the discrepancies.
  4. [Sec. 4.1 / Table 2] The claim that the method achieves high accuracy across all metrics is not supported by the Office row: DROID-SLAM obtains better PSNR (25.680 vs 25.649) and better LPIPS (0.144 vs 0.160). The paper overstates the result. The narrative should be revised to a more nuanced claim, with an analysis of why the method is less competitive in this scene.
  5. [Eq. (1) / Sec. 3.1] The depth consistency loss in Eq. (1) is insufficiently specified. The paper does not define the SE(3) convention for exp(ω)^T x, nor does it derive why PjHj is measured against -exp(ω)^T x. As written, this is difficult to reproduce and the geometric correctness is unclear. Please provide a step-by-step derivation defining coordinate frames and the exponential map convention.
  6. [Sec. 4] The quantitative evaluation is based on three synthetic scenes, each seemingly a single run, without error bars or statistical significance. The differences in Table 2 are sometimes very small (e.g., Office PSNR difference is 0.03 dB), so single-run results are not convincing. The authors should run multiple seeds/trajectories and report variance.
minor comments (6)
  1. [Implementation details] The text 'data processing partially adapted from CF-3DGS [14]' is unclear: reference [14] is 3D Gaussian Splatting, not CF-3DGS. Please correct the citation.
  2. [Sec. 4.1] NICE-SLAM, Point-SLAM, and SplaTAM are mentioned in the text but omitted from Table 1. The main text should state which methods failed and why; the supplementary explanation is not enough for a reader of the main paper.
  3. [Eqs. (5) and (8)] The loss weights λ_D, λ_o, and λ_B are never reported. Please provide the actual values used in the experiments for reproducibility.
  4. [Sec. 3.2] The choice of voxel size s and the fixed Gaussian scale s/2 is not justified. Please report the value used and, if possible, include a sensitivity study.
  5. [Sec. 3.1] The mask definition 'M_t(p_i) = D_t(p_i) > 0' is written without a formal condition; clarify that the mask is 1 where depth is valid and 0 otherwise.
  6. [Table 3] In Table 3, the 'w/o depth loss' and 'Scale fix + refinement' rows have identical pose metrics. The text explains this indirectly, but readers may be confused; please add a note that the deblur depth loss does not affect the pose-refinement stage.

Circularity Check

0 steps flagged

No circularity: the pipeline is self-contained and evaluated against external ground truth; robustness caveats are correctness concerns, not circular reasoning.

full rationale

The derivation chain in this paper is self-contained and externally checked. Camera poses are initialized from optical flow + ICP (Sec. 3.1) and refined by depth/opacity supervision (Eqs. 3–5); neither of these steps defines the target quantity in terms of itself. The deblur model (Eq. 6) is a standard exposure-averaging assumption, and the deblur losses (Eqs. 7–8) follow BAD-Gaussians [49], an external prior work, not a self-citation. The quantitative evaluation is against ground-truth camera trajectories and ground-truth novel views on a new dataset (Tables 1–2, Sec. 4.1), so the reported metrics are not manufactured by a fitted parameter renamed as a prediction. The only load-bearing caveats are robustness concerns: Sec. 3.2's depth-only refinement is non-convex and the Supplementary Limitations section (Fig. 6) concedes textureless-scene failures; the paper also does not perturb the Sec. 3.1 initializations to test their basin of convergence. These are correctness/robustness limitations, not circularity. The 'CF-3DGS [14]' implementation note is a citation-formatting inconsistency and does not support any inference. No circular step is present; score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The method leans on standard building blocks (Gaussian Splatting, optical flow, ICP) plus two domain assumptions specific to this setting: a linear-SE(3) exposure model and reliable metric depth. The main ad hoc choices are the fixed Gaussian scale and the underspecified baseline-depth loss. No new physical entities are introduced.

free parameters (4)
  • Loss weights λ_I, λ_C (and unstated λ_D, λ_o, λ_B) = λ_I=75, λ_C=25; others not reported
    Weights in Eqs. (2), (5), (8) trade off optical flow, depth consistency, depth, opacity, and SSIM terms; no sensitivity analysis is given.
  • Number of virtual views k = 15 (synthetic), 17 (real)
    Controls the discrete approximation of the exposure-average in Eq. (6); chosen per dataset without ablation.
  • Voxel size s and Gaussian scale s/2 = not reported
    Determines point-cloud density and fixed Gaussian scale; the fixed scale is load-bearing per the ablation, but no numeric value is given.
  • Iteration/pruning schedule = prune every 100N, reset every 200N, total 600N
    Hyperparameters of the opacity-pruning schedule in Section 3.2; no sensitivity study is reported.
axioms (6)
  • domain assumption Exposure-time motion blur equals the average of k sharp views along a linearly interpolated SE(3) path (Eq. 6).
    Underlies the entire deblur loss; real rolling-shutter or nonlinear motion breaks it. Adopted from BAD-Gaussians, not derived here.
  • domain assumption Pretrained optical flow (Unimatch) provides reliable pixel correspondences between two extremely blurry frames.
    Used to initialize local poses in Section 3.1; no failure analysis is provided.
  • domain assumption Depth maps are metric, RGB-aligned, and accurate enough to supervise global geometry (Eq. 3).
    The refinement and depth-consistency losses assume depth is ground truth; sensor noise or misalignment would bias poses.
  • ad hoc to paper The SE(3) exponential map convention in Eq. (1) is such that P_i·x and P_j·(-exp(ω)^T x) measure distances to the same baseline line.
    Under the standard camera-to-camera convention the baseline direction is -R^T x, so the loss as written may be frame-inconsistent; no convention is specified.
  • ad hoc to paper Fixing Gaussian scale to s/2 and optimizing only position/opacity resolves pose-geometry ambiguity.
    Justified only by ablation (Table 3), not by analysis of gauge freedom.
  • standard math Standard geometry and ICP convergence: point-to-plane ICP succeeds after global alignment.
    Local pose step uses point-to-plane ICP; assumes the initialization is within the convergence basin.

pith-pipeline@v1.3.0-alltime-deepseek · 15329 in / 21310 out tokens · 217602 ms · 2026-08-01T19:32:05.571609+00:00 · methodology

0 comments
read the original abstract

We propose a splat-based 3D scene reconstruction method from RGB-D input that effectively handles extreme motion blur, a frequent challenge in low-light environments. Under dim illumination, RGB frames often suffer from severe motion blur due to extended exposure times, causing traditional camera pose estimation methods, such as COLMAP, to fail. This results in inaccurate camera pose and blurry color input, compromising the quality of 3D reconstructions. Although recent 3D reconstruction techniques like Neural Radiance Fields and Gaussian Splatting have demonstrated impressive results, they rely on accurate camera trajectory estimation, which becomes challenging under fast motion or poor lighting conditions. Furthermore, rapid camera movement and the limited field of view of depth sensors reduce point cloud overlap, limiting the effectiveness of pose estimation with the ICP algorithm. To address these issues, we introduce a method that combines camera pose estimation and image deblurring using a Gaussian Splatting framework, leveraging both 3D Gaussian splats and depth inputs for enhanced scene representation. Our method first aligns consecutive RGB-D frames through optical flow and ICP, then refines camera poses and 3D geometry by adjusting Gaussian positions for optimal depth alignment. To handle motion blur, we model camera movement during exposure and deblur images by comparing the input with a series of sharp, rendered frames. Experiments on a new RGB-D dataset with extreme motion blur show that our method outperforms existing approaches, enabling high-quality reconstructions even in challenging conditions. This approach has broad implications for 3D mapping applications in robotics, autonomous navigation, and augmented reality. Both code and dataset are publicly available on https://github.com/KAIST-VCLAB/gs-extreme-motion-blur.

Figures

Figures reproduced from arXiv: 2607.16926 by Donggun Kim, Dongyoung Choi, Hyeonjoong Jang, Min H. Kim, Woohyun Kang.

Figure 1
Figure 1. Figure 1: We propose a robust 3D scene reconstruction method from RGB-D input that effectively addresses extreme motion blur. Our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our method pipeline. We begin with global alignment by comparing optical flow and reprojected pixel differences [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The illustration of our pose initialization loss. When [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: The refinement process corrects misaligned 3D Gaus [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of deblurring performance from a novel view. Our method excels in restoring high-frequency details, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of deblurring performance from novel views on real scenes captured with an Azure Kinect camera. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 1
Figure 1. Figure 1: Reconstruction results for the living room scene: [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: NICE-SLAM produces bumpy reconstructions and has [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Deblurring comparison to [45] and [49]. Comparison with Open3D reconstruction pipeline. We compare our method with that of Choi et al. [3] using the Open3D implementation. Their approach fails due to inac￾curately estimated poses caused by motion blur ( [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of (b) ours to (a) Choi et al. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison with BAGS and Deblurring 3D GS. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Failure case with a featureless scene. camera pose becomes difficult, and the outputs suffer from wavy artifacts caused by deblurring ambiguity ( [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of deblurring results using NAFNet [ [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of estimated camera trajectories: The top row displays trajectories for synthetic scenes, while the bottom row [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Our refinement step (Section 3.2 in the main paper) effectively corrects misalignments, ensuring accurate alignment while [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Reconstructed point clouds for synthetic scenes: Our method demonstrates superior performance by producing dense and [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: We show reconstructed point cloud from each method for real scenes. Our method provides dense and well-aligned point clouds. [PITH_FULL_IMAGE:figures/full_fig_p017_11.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

13 extracted references · 1 linked inside Pith

  1. [1]

    Orb-slam3: An accu- rate open-source library for visual, visual–inertial, and mul- timap slam.IEEE Transactions on Robotics, 37(6):1874– 1890, 2021

    Carlos Campos, Richard Elvira, Juan J G ´omez Rodr´ıguez, Jos´e MM Montiel, and Juan D Tard´os. Orb-slam3: An accu- rate open-source library for visual, visual–inertial, and mul- timap slam.IEEE Transactions on Robotics, 37(6):1874– 1890, 2021. 1

    2021

  2. [2]

    Simple baselines for image restoration.arXiv preprint arXiv:2204.04676, 2022

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration.arXiv preprint arXiv:2204.04676, 2022. 4

    Pith/arXiv arXiv 2022

  3. [3]

    Ro- bust reconstruction of indoor scenes

    Sungjoon Choi, Qian-Yi Zhou, and Vladlen Koltun. Ro- bust reconstruction of indoor scenes. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 5556–5565, 2015. 2

    2015

  4. [4]

    evo: Python package for the evalua- tion of odometry and slam.https://github.com/ MichaelGrupp/evo, 2017

    Michael Grupp. evo: Python package for the evalua- tion of odometry and slam.https://github.com/ MichaelGrupp/evo, 2017. 2

    2017

  5. [5]

    Splatam: Splat track & map 3d gaussians for dense rgb-d slam

    Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21357–21366, 2024. 2

    2024

  6. [6]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and An- drew J Davison. Gaussian splatting slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 18039–18048, 2024. 2

    2024

  7. [7]

    Fast point feature histograms (fpfh) for 3d registration

    Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In2009 IEEE international conference on robotics and automation, pages 3212–3217. IEEE, 2009. 1

    2009

  8. [8]

    Point-slam: Dense neural point cloud-based slam

    Erik Sandstr ¨om, Yue Li, Luc Van Gool, and Martin R Os- wald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433–18444, 2023. 2

    2023

  9. [9]

    DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras.Ad- vances in neural information processing systems, 2021

    Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras.Ad- vances in neural information processing systems, 2021. 2

    2021

  10. [10]

    Deblur-gs: 3d gaussian splat- ting from camera motion blurred images.Proc

    Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splat- ting from camera motion blurred images.Proc. ACM Com- put. Graph. Interact. Tech. (Proceedings of I3D 2024), 7(1),

    2024

  11. [11]

    Elasticfusion: Real-time dense slam and light source estimation.The Inter- national Journal of Robotics Research, 35(14):1697–1716,

    Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, An- drew J Davison, and Stefan Leutenegger. Elasticfusion: Real-time dense slam and light source estimation.The Inter- national Journal of Robotics Research, 35(14):1697–1716,

  12. [12]

    BAD- Gaussians: Bundle Adjusted Deblur Gaussian Splatting

    Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD- Gaussians: Bundle Adjusted Deblur Gaussian Splatting

  13. [13]

    Nice-slam: Neural implicit scalable encoding for slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu- jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle- feys. Nice-slam: Neural implicit scalable encoding for slam. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 12786–12796, 2022. 2 Input NAFNet Ours Figure 7. Comparison of deblurring results using NAF...

    2022