Pith. sign in

REVIEW 2 major objections 4 minor 36 references

Learned dual-fisheye calibration turns imperfect 360° captures into seamless rendered views.

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 ↗

Training 3D Gaussian splatting with a learnable dual-fisheye distortion model, then turning it off at inference, renders seamless 360-degree novel views from imperfect panoramas.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid practical calibration for 360° 3DGS, but the angular-distortion component is under-validated and the 'seamless' claim rests on an identifiability assumption. the 2 major comments →

arxiv 2508.20080 v1 pith:2DFUM7BQ submitted 2025-08-27 cs.CV cs.GR

Seam360GS: Seamless 360{\deg} Gaussian Splatting from Real-World Omnidirectional Images

Changha Shin , Woong Oh Cho , Seon Joo Kim This is my paper
classification cs.CV cs.GR
keywords 360-degree imagingdual-fisheye cameras3D Gaussian splattingcamera calibrationnovel view synthesisomnidirectional renderingstitching artifacts
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

The paper claims that the seams, overlaps, and distortions in consumer 360° panoramas can be compensated during 3D Gaussian splatting training rather than requiring perfectly stitched, pre-calibrated inputs. It models a dual-fisheye camera as two offset optical centers plus a smoothly varying, per-ray rotation field, and optimizes these calibration parameters together with the scene Gaussians using only the imperfect captured images as supervision. At test time the calibration is switched off, so the model renders an ideal, seamless 360° view. On real Ricoh360 and 360Roam datasets the method reports higher PSNR and SSIM and lower LPIPS than prior 360° rendering baselines, while using fewer Gaussians. If correct, this means consumer cameras can produce accurate, artifact-free novel views without raw fisheye capture or per-device manual calibration.

Core claim

The paper's central claim is that the two failure modes of consumer dual-fisheye 360° cameras—the physical separation of the front and back optical centers and the angular distortion introduced by stitching—can be modeled and corrected inside a 3D Gaussian splatting optimization. Instead of treating input panoramas as perfectly calibrated spherical images, the method re-expresses each ideal Gaussian in front- and back-camera coordinates using learned translations ΔT_F and ΔT_B, then rotates it by a learned per-ray angular distortion ΔR_F or ΔR_B (Eqs. 4–9). Training jointly fits both the scene and these calibration variables to the imperfect captures through a photometric loss; at inference

What carries the argument

The central mechanism is a distortion-aware dual-fisheye Gaussian transformation. Each 3D Gaussian is first translated from an ideal center to the learned front/back fisheye centers via ΔT_F and ΔT_B, then rotated by a smooth, learnable angular-distortion grid ΔR(θ, φ) sampled bilinearly in spherical coordinates, producing camera-specific Gaussians that are rasterized separately and merged into a complete 360° panorama. During training this reproduces the artifact patterns of real captures, so the photometric loss drives both the scene and the calibration variables; during inference the calibration is switched off, making the original ideal Gaussians render a seamless view.

Load-bearing premise

The load-bearing premise is that a single learned translation per fisheye plus a smooth per-ray rotation field can fully represent the misalignment of a real dual-fisheye panorama; if the camera also has radial distortion, rolling-shutter effects, or nonlinear stitching errors, the fitted rotation can absorb scene-geometry errors and produce visually seamless but geometrically wrong views.

What would settle it

Render a known 3D calibration object (for example a checkerboard rig or a straight-edged corridor) through the learned dual-fisheye model and inspect the merged 360° output: if straight 3D lines bend across the front/back seam, or if the recovered ΔR changes substantially when the scene content changes while the camera is fixed, the angular-distortion field is overfitting to scene geometry rather than recovering true lens distortion. A second check is to compare depth maps rendered with and without the calibration module against a structure-from-motion or LiDAR reference.

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

If this is right

  • Seamless 360° novel views become obtainable from ordinary consumer panoramas, without raw fisheye capture or per-view manual alignment.
  • Calibration variables are optimized only during training, so inference stays real-time; the paper reports roughly 30% faster test-time rendering than OmniGS while using fewer Gaussians.
  • The angular-distortion grid learned for one scene transfers to other scenes from the same camera, suggesting a device-level calibration can be reused.
  • The same pipeline works on both equirectangular panoramas and raw dual-fisheye inputs, with a small quality gain on raw fisheye, broadening its applicability.
  • Ablation results show both calibration components matter, with angular-distortion correction contributing more than lens-gap translation; disabling both costs about 1 dB PSNR on the 360Roam indoor scenes.

Where Pith is reading between the lines

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

  • Because the calibration is fitted with photometric loss alone, the learned ΔR may absorb geometric errors in textureless regions; the paper itself reports reduced calibration performance there, so 'seamless' should be read as a visual property first, with geometric fidelity needing a separate check.
  • The near-identical ΔR across scenes from the same camera suggests a practical extension the paper does not test: pre-train the calibration module on a few scenes, freeze it, and then train new scenes faster.
  • The synthetic gaps are sampled within ±2 cm, so the model's behavior for larger baseline errors or for non-rotational distortion (for example rolling-shutter effects) is untested; extending the calibration with more intrinsic parameters is a natural next step.
  • Since the reported ground truth itself contains stitching artifacts, part of the measured gain may come from outputting a cleaner image than the reference; an evaluation against a high-quality panoramic capture or a laser scan would better separate calibration accuracy from image smoothing.
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

2 major / 4 minor

Summary. The paper proposes Seam360GS, a 3D Gaussian Splatting framework for 360° novel-view synthesis that explicitly models artifacts of consumer dual-fisheye cameras. It translates Gaussians to per-camera centers with learnable vectors ΔT_F, ΔT_B and applies a learned per-ray angular distortion field ΔR_F, ΔR_B when rasterizing front/back views; the two views are then concatenated to reproduce the imperfect input panoramas. At inference the distortion module is disabled, yielding an 'ideal' seamless panorama. Experiments on a synthetic dataset, EgoNeRF/Ricoh360, and 360Roam show consistent improvements over four baselines in PSNR/SSIM/LPIPS, and an ablation confirms both components contribute. A synthetic experiment validates the learned inter-camera translation against ground truth.

Significance. If the calibration is physically meaningful, the method is practically valuable: it addresses a real, under-studied problem—stitching artifacts in consumer 360° cameras—within the efficient 3DGS paradigm, with no extra inference cost. The paper is well specified, uses standard benchmarks and baselines, and includes a synthetic calibration test and component ablations. The main limitation is that the angular distortion field is never validated against ground truth, so the central 'seamless and geometrically accurate' claim rests on an identifiability assumption. This is a correctness risk that should be resolved before publication.

major comments (2)
  1. [§3.2, Eqs. (8)–(9), and §4.2, Table 1] The angular distortion field ΔR is load-bearing for the paper's 'flawless seamless rendering' claim, but it is learned only under a photometric loss against distorted panoramas. Table 1 validates the inter-camera translation ΔT against synthetic ground truth, but no experiment measures the error in ΔR. Because the Gaussian scene geometry is optimized simultaneously, many (ΔR, scene) pairs can produce nearly identical distorted training views: the model may absorb scene-reconstruction errors into ΔR and then remove them at inference, yielding a smooth but geometrically false 'ideal' view. The cross-scene consistency of ΔR reported in Supplementary Sec. B is consistent with a camera property but does not establish that it is the true physical distortion. Please add a synthetic experiment with known angular distortions and report recovery error for ΔR, or evaluate the undistorted renderings
  2. [§4.3, Tables 2–4] The core empirical claim is an average PSNR gain of ~0.9–1.0 dB over the strongest baseline, but no error bars, multiple seeds, or significance tests are reported. The real-world ground truth itself contains stitching artifacts, so the reported metrics measure fidelity to imperfect images, not correctness of the seamless rendering. Also, the OP43DGS results on Ricoh360 come from a checkpoint saved when training terminated due to out-of-memory (Sec. 4.1); this is disclosed but should be highlighted in the table and considered when interpreting the ranking. A dedicated evaluation around the seam region and a geometric check would substantially strengthen the paper.
minor comments (4)
  1. [§4.1, Implementations] The text says 'OP43DGS [22]' but OP43DGS is reference [29]; [22] is OmniGS. Please correct the citation.
  2. [§4.3, Table 2] The text claims the method uses 'fewer Gaussians' than baselines. This is true on Synthetic and 360Roam, but on Ricoh360 Ours (621K) is slightly above OmniGS (619K). Please qualify the claim.
  3. [§3.1, Eq. (3)] The simplified rendering equation writes ∥p − π(m_i)∥²_W_i as a Mahalanobis distance, but W_i is not defined in the simplified formulation. A one-sentence definition or a pointer to [16] would avoid confusion.
  4. [§4.2, Table 1] The comparison with OmniNeRF mixes input formats: ours consumes stitched ERP panoramas, while OmniNeRF consumes raw dual-fisheye images. The large MAE gap may be partly attributable to the input representation, not only to the calibration mechanism. Please state this explicitly or add an ERP-input variant of OmniNeRF.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical comparison against external baselines; calibration parameters are fitted, not predictions derived from the target.

full rationale

The paper's central claim is that jointly optimizing 3D Gaussian parameters with translation vectors ΔT and angular distortion grids ΔR yields seamless 360° novel views from imperfect dual-fisheye panoramas. The load-bearing evaluation is Table 2, an empirical comparison of PSNR/SSIM/LPIPS against external baselines (EgoNeRF, OP43DGS, ODGS, OmniGS) on real-world datasets. This is not circular: the reported numbers are measured against held-out ground-truth images, not derived from the fitted parameters by construction. The calibration variables ΔT and ΔR are learned parameters, and the 'seamless' output at inference is obtained by omitting them; this is the intended inversion of the fitted artifact model, not a prediction of a quantity that was itself used as the fit target. The synthetic calibration assessment (Sec. 4.2, Table 1) measures recovery of a known inter-camera gap from synthetic data; although the forward model resembles the data-generation setup, the ground-truth gap is external to the optimization and the MAE is a genuine error measure, so this is a self-consistency check rather than a circular reduction. There are no load-bearing self-citations: the authors do not cite their own prior work to justify the core model, and the cited baselines are external. The Limitations paragraph (Sec. 5) notes degraded performance in textureless regions, which is a stated weakness, not a circular step. The identifiability concern that ΔR might absorb scene-geometry errors under photometric loss alone is a correctness/validity risk about whether the learned distortion is the true camera distortion, but it is not a circularity in the paper's equations: no Eq. X is defined in terms of Eq. Y, and no fitted parameter is renamed as a prediction. Accordingly, no circular steps are exhibited, and the score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central claim depends on two fitted structures (camera-center translations and per-ray rotation grids) that are learned purely from photometric consistency on training views. They are not derived from the camera's optical design, and no independent geometric ground truth is used on real scenes. All other components (3DGS, rasterizer) come from prior work.

free parameters (2)
  • dT_F, dT_B (per-camera translation vectors) = not reported; synthetic MAE 0.115 cm vs ground truth
    Learned 3D translations of the front/back camera centers (Eqs. 4-5), fit by photometric loss during training; no real-camera ground-truth values reported.
  • dR_F, dR_B (angular distortion grids) = not reported as numbers; resolutions tested 16x32 to 128x256
    Learnable rotation fields indexed by spherical direction (theta, phi), fit to training views; no independent verification of their physical meaning.
axioms (5)
  • standard math 3DGS rendering equations (Eqs. 1-3) correctly model the scene
    Adopted from the original 3DGS paper; assumed as background.
  • domain assumption OmniGS 360-degree rasterizer correctly projects Gaussians onto the unit sphere
    Used as the rendering backend; its geometric correctness is taken from the cited work [22].
  • ad hoc to paper Dual-fisheye artifacts can be modeled by a per-camera 3D translation plus a smooth per-ray rotation field (Eqs. 4-9)
    Central modeling choice; not derived from lens optics and not validated geometrically on real cameras.
  • domain assumption The angular distortion grids are smooth, enforced by total variation loss
    Needed to prevent noisy dR; the smoothness prior is assumed to match real distortion behavior.
  • domain assumption Input ERP panoramas are a valid representation for training
    The method operates on equirectangular images, which are assumed to preserve the distortion structure.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Seam360GS: Seamless 360{\deg} Gaussian Splatting from Real-World Omnidirectional Images." pith.science (2026). https://pith.science/paper/2DFUM7BQ

@misc{pith2026250820080,
  author       = {Pith},
  title        = {Pith review of: Seam360GS: Seamless 360\deg Gaussian Splatting from Real-World Omnidirectional Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DFUM7BQ}},
  note         = {Machine review of arXiv:2508.20080}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

360-degree visual content is widely shared on platforms such as YouTube and plays a central role in virtual reality, robotics, and autonomous navigation. However, consumer-grade dual-fisheye systems consistently yield imperfect panoramas due to inherent lens separation and angular distortions. In this work, we introduce a novel calibration framework that incorporates a dual-fisheye camera model into the 3D Gaussian splatting pipeline. Our approach not only simulates the realistic visual artifacts produced by dual-fisheye cameras but also enables the synthesis of seamlessly rendered 360-degree images. By jointly optimizing 3D Gaussian parameters alongside calibration variables that emulate lens gaps and angular distortions, our framework transforms imperfect omnidirectional inputs into flawless novel view synthesis. Extensive evaluations on real-world datasets confirm that our method produces seamless renderings-even from imperfect images-and outperforms existing 360-degree rendering models.

Figures

Figures reproduced from arXiv: 2508.20080 by Changha Shin, Seon Joo Kim, Woong Oh Cho.

Figure 1
Figure 1. Figure 1: (Top) An indoor scene captured by a 360° camera, with red and green boxes indicating regions affected by stitching artifacts. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our seamless 360° Gaussian Splatting rendering pipeline, which integrates fisheye camera stitching and applies lens [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparisons of novel-view synthesis results. The top portion shows images from the synthetic dataset ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on the 360Roam base. The top row displays the output of the state-of-the-art model OmniGS, fol￾lowed by our method’s results, the ground truth, and finally the seamless rendering output produced by our method. tail and fewer artifacts. Our seamless rendering pipeline, validated through rigorous ablation studies, underscores the importance of precise calibration in achieving high-fide… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on raw fisheye inputs: vanilla fisheye model [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparisons of novel-view synthesis on [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparisons of novel-view synthesis on [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparisons of novel-view synthesis on [PITH_FULL_IMAGE:figures/full_fig_p016_8.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

36 extracted references · 29 canonical work pages

  1. [1]

    http://www.insta360.com/

    Insta360 camera. http://www.insta360.com/. 1

    work page

  2. [2]

    https://www.ricoh360.com/

    Ricoh360 camera. https://www.ricoh360.com/. 1

    work page

  3. [3]

    Deep learning for omnidirectional vision: A survey and new perspectives

    Hao Ai, Zidong Cao, Jinjing Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468, 2022. 2

    Pith/arXiv arXiv 2022

  4. [4]

    360-gs: Layout-guided panoramic gaussian splatting for indoor roaming

    Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024. 2, 3, 4

    Pith/arXiv arXiv 2024

  5. [5]

    Speeded- up robust features (surf)

    Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Speeded- up robust features (surf). In Computer Vision and Image Un- derstanding, pages 346–359. Elsevier, 2008. 2

    work page 2008

  6. [6]

    Nope-nerf: Optimising neu- ral radiance field with no pose prior

    Wenjing Bian, Zirui Wang, Kejie Li, Jiawang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neu- ral radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3

    work page 2023

  7. [7]

    Automatic panoramic image stitching using invariant features

    Matthew Brown and David G Lowe. Automatic panoramic image stitching using invariant features. In International Journal of Computer Vision , pages 59–73. Springer, 2007. 2

    work page 2007

  8. [8]

    Panogrf: Generalizable spherical radiance fields for wide-baseline panoramas

    Zheng Chen, Yan-Pei Cao, Yuan-Chen Guo, Chen Wang, Ying Shan, and Song-Hai Zhang. Panogrf: Generalizable spherical radiance fields for wide-baseline panoramas. In NeurIPS, 2023. 3

    work page 2023

  9. [9]

    Bal- anced spherical grid for egocentric view synthesis

    Changwoon Choi, Sang Min Kim, and Young Min Kim. Bal- anced spherical grid for egocentric view synthesis. InCVPR,

    work page

  10. [10]

    Barron, and Pratul P

    Boyang Deng, Jonathan T. Barron, and Pratul P. Srinivasan. JaxNeRF: an efficient JAX implementation of NeRF. 2020. 6

    work page 2020

  11. [11]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981

    Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981. 2

    1981

  12. [12]

    Efros, and Xiaolong Wang

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024. 2, 3

    work page 2024

  13. [13]

    Omni-nerf: neural radiance field from 360 image captures

    Kai Gu, Thomas Maugey, Sebastian Knorr, and Christine Guillemot. Omni-nerf: neural radiance field from 360 image captures. In 2022 IEEE International Conference on Multi- media and Expo (ICME), pages 1–6. IEEE, 2022. 4, 6

    work page 2022

  14. [14]

    360roam: Real-time indoor roaming us- ing geometry-aware 360° radiance fields

    Huajian Huang, Yingshu Chen, Tianjia Zhang, and Sai- Kit Yeung. 360roam: Real-time indoor roaming us- ing geometry-aware 360° radiance fields. arXiv preprint arXiv:2208.02705, 2022. 2, 3, 6

    Pith/arXiv arXiv 2022

  15. [15]

    SC-omniGS: Self-calibrating omnidirectional gaussian splatting

    Huajian Huang, Yingshu Chen, Longwei Li, Hui Cheng, Tristan Braud, Yajie Zhao, and Sai-Kit Yeung. SC-omniGS: Self-calibrating omnidirectional gaussian splatting. In The Thirteenth International Conference on Learning Represen- tations, 2025. 2, 4, 13

    work page 2025

  16. [16]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  17. [17]

    Atten- tive deep stitching and quality assessment for 360° omnidi- rectional images

    Hak Gu Kim, Heoun-Taek Lim, and Yong Man Ro. Atten- tive deep stitching and quality assessment for 360° omnidi- rectional images. IEEE Journal of Selected Topics in Signal Processing, 14(1):209–221, 2020. 2

    work page 2020

  18. [18]

    Omnistitch: Depth-aware stitching framework for omnidirectional im- ages

    Minsu Kim, Jaewon Lee, and Sungmin Park. Omnistitch: Depth-aware stitching framework for omnidirectional im- ages. In Proceedings of the ACM Multimedia Conference ,

    work page

  19. [19]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. In NeurIPS. Curran Associates, Inc., 2012. 6

    work page 2012

  20. [20]

    360fu- sionnerf: Panoramic neural radiance fields with joint guid- ance

    Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fu- sionnerf: Panoramic neural radiance fields with joint guid- ance. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023. 3

    work page 2023

  21. [21]

    Odgs: 3d scene reconstruction from omni- directional images with 3d gaussian splattings

    Suyoung Lee, Jaeyoung Chung, Jaeyoo Huh, and Ky- oung Mu Lee. Odgs: 3d scene reconstruction from omni- directional images with 3d gaussian splattings. In NeurIPS,

    work page

  22. [22]

    Omnigs: Omnidirectional gaussian splatting for fast radiance field reconstruction using omnidirectional images

    Longwei Li, Huajian Huang, Sai-Kit Yeung, and Hui Cheng. Omnigs: Omnidirectional gaussian splatting for fast radiance field reconstruction using omnidirectional images. arXiv preprint arXiv:2404.03202, 2024. 2, 3, 4, 6, 8, 12, 13, 14

    Pith/arXiv arXiv 2024

  23. [23]

    Barf: Bundle-adjusting neural radiance fields

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Si- mon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3

    work page 2021

  24. [24]

    Distinctive image features from scale- invariant keypoints

    David G Lowe. Distinctive image features from scale- invariant keypoints. International Journal of Computer Vi- sion, 60(2):91–110, 2004. 2

    work page 2004

  25. [25]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM, 65(1):99–106, 2021. 2, 3

    work page 2021

  26. [26]

    Pytorch: An im- perative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library. Ad- vances in neural information processing systems , 32, 2019. 6

    work page 2019

  27. [27]

    and Martin-Brualla, Ricardo

    Park, Keunhong and Henzler, Philipp and Mildenhall, Ben and Barron, Jonathan T. and Martin-Brualla, Ricardo. Camp: Camera preconditioning for neural radiance fields. In ACM Trans. Graph., 2023. 2, 3

    work page 2023

  28. [28]

    Sparf: Neural radiance fields from sparse and noisy poses

    Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3

    work page 2023

  29. [29]

    Wang and et al

    X. Wang and et al. op43dgs: Optimized projection for 360° gaussian splatting. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,

    work page

  30. [30]

    Self- calibrating neural radiance fields

    Christopher Choy Animashree Anandkumar Minsu Cho Yoonwoo Jeong, Seokjun Ahn and Jaesik Park. Self- calibrating neural radiance fields. In ICCV, 2021. 2, 3

    work page 2021

  31. [31]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6

    2018

  32. [32]

    2, 3, 4, 6, 12, 13, 14

    work page

  33. [33]

    Nerf–: Neural radiance fields without camera pose calibration

    Weidi Xie Min Chen Victor Adrian Prisacariu Zirui Wang, Shangzhe Wu. Nerf–: Neural radiance fields without camera pose calibration. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,

    work page

  34. [35]

    Depth-aware stitching framework for omnidirectional vision

    Wei Zhang, Ming Li, and Jian Chen. Depth-aware stitching framework for omnidirectional vision. In Proceedings of the ICCV Workshops, 2023. 2

    work page 2023

  35. [2021]

    Baseline Comparisons with Dual-Fisheye and MVG We further evaluate our method using raw dual-fisheye inputs

    2, 3 Seam360GS: Seamless 360° Gaussian Splatting from Real-World Omnidirectional Images Supplementary Material A. Baseline Comparisons with Dual-Fisheye and MVG We further evaluate our method using raw dual-fisheye inputs. Tab.5 and Fig.5 show that applying our seamless calibration al- gorithm to fisheye images yields a meaningful improvement (PSNR +0.34d...

    work page

  36. [2023]

    2, 3, 5, 6, 12, 13, 14

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.