SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting

Huajian Huang; Hui Cheng; Longwei Li; Sai-Kit Yeung; Tristan Braud; Yajie Zhao; Yingshu Chen

REVIEW 4 major objections 7 minor 1 cited by

SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read SC-OmniGS shows that 360-degree radiance fields can be reconstructed without precomputed camera poses, by jointly optimizing Gaussians, poses, and a learned distortion model.

desk verdict Useful engineering contribution with a real identifiability gap: poses and the per-pixel distortion field are jointly optimized, but the distortion model is never independently validated against known distortions. read the letter →

arxiv 2502.04734 v1 pith:WVC273EB submitted 2025-02-07 cs.CV cs.GR

Huajian Huang , Yingshu Chen , Longwei Li , Hui Cheng , Tristan Braud , Yajie Zhao , Sai-Kit Yeung This is my paper · ORCID

classification cs.CVcs.GR

keywords omnidirectionalGaussiansplattingself-calibrationcameraposeoptimizationpanoramicradiancefieldsdifferentiablemodelequirectangularprojectionnovelviewsynthesis

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

This paper aims to show that a 360-degree radiance field can be reconstructed end-to-end without accurate precomputed camera poses. SC-OmniGS treats each equirectangular image as a whole sphere, jointly optimizing 3D Gaussians, camera extrinsic poses, and a learned per-scene distortion model by minimizing a weighted spherical photometric loss. The authors report that this recovers high-quality novel views from noisy poses or even from all cameras initialized at the origin, and that the learned camera model improves results on real consumer-grade 360-degree captures over a fixed spherical projection. If the claim holds, it removes a major practical bottleneck: scene capture no longer needs a separate structure-from-motion calibration stage.

What carries the argument

The load-bearing object is the differentiable omnidirectional Gaussian rasterizer and its pose-gradient derivation. Instead of splatting onto a plane, 3D Gaussians are projected onto a unit sphere with $\phi_o(x) = (W/2\pi \cdot \arctan2(x,z)+W/2,\; H/\pi \cdot \arcsin(y/d)+H/2)$, and the Jacobian $J_o$ of this projection carries the splat into the 2D Gaussian kernel. The paper then applies the chain rule to the rendered color and kernel with respect to the 7-dimensional pose $T'=[q|t]$, giving closed-form gradients for quaternion and translation updates. A second component, the learnable camera model $\Theta := S \cdot f_t + S \odot D$, resamples input panoramas to remove real-world distortion. The optimization objective $L = L_{\mathrm{wsp}} + L_{\mathrm{aniso}}$ weights each pixel by its spherical solid-angle density and penalizes elongated Gaussians, keeping the joint calibration stable.

What would settle it

Render a synthetic equirectangular image with a known distortion field, initialize SC-OmniGS with exact poses, and compare the learned $\Theta$ with the injected field; if $\Theta$ does not reproduce the known distortion to pixel-level accuracy, the photometric-only calibration has not been shown to identify real distortion.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that omnidirectional camera self-calibration and Gaussian radiance field optimization can be folded into one differentiable pipeline. The rasterizer projects 3D Gaussians onto a unit sphere and then to equirectangular coordinates, and the authors derive the pose gradients for this projection analytically, representing each camera as a quaternion-plus-translation vector $T'=[q|t]$ to avoid singularities. Alongside poses, a generic omnidirectional camera model $\Theta = S \cdot f_t + S \odot D$ learns per-ray angle distortion from photometric loss alone, and undistorted training images are obtained by bicubic resampling. With a spatially weighted loss that compensates for oversampling near the poles and an anisotropy regularizer, the system jointly optimizes Gaussians, poses, and the camera model in about 30 minutes per scene, matching or exceeding earlier methods trained with known poses in the reported experiments.

Load-bearing premise

The load-bearing premise is that the per-pixel distortion model, learned only from photometric loss, captures true optical distortion rather than quietly absorbing errors in camera pose or scene geometry.

Editorial extensions

If this is right

If the reported results hold, 360-degree scene capture can skip a dedicated structure-from-motion calibration pass: the same pipeline handles accurate, noisy, or absent pose priors.
Jointly learning a distortion model should improve any downstream pipeline that assumes an ideal spherical camera, since real stitched panoramas deviate from that model.
Pose estimates are themselves a product of the optimization, so SC-OmniGS can output calibrated camera trajectories usable for navigation or SLAM initialization.
The method's robustness to random Gaussian initialization suggests it can bootstrap from a single depth map or even no geometric prior at all in the tested indoor and egocentric settings.

Reading between the lines

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

If the learned $\Theta$ converges to true optical distortion, a camera-specific $\Theta$ could be reused across scenes from the same device, converting per-scene calibration into a transferable prior—this is not tested in the paper.
The pose-gradient derivation is not limited to equirectangular output; porting it to other sphere-based splatting renderers could make monocular omnidirectional SLAM self-calibrating.
A direct stress test would feed SC-OmniGS synthetic panoramas with an injected, known distortion field and check whether $\Theta$ recovers that field exactly; photometric loss alone may not separate distortion from pose errors.

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.

Desk Editor's Note

read the letter

Short version: this is a solid systems paper, not a breakthrough. It genuinely extends OmniGS by deriving pose gradients through the omnidirectional splatter and adding a learnable per-pixel distortion field; the experiments are broad and the gains over perturbed-pose baselines are consistent. But the headline 'no pose prior' is narrower than the abstract suggests, and the camera model itself is the soft spot: a ~3.5M-parameter per-pixel warp, shared across views, trained only against photometric loss, with no smoothness prior or independent calibration check. That is a real identifiability concern, though the paper is transparent enough that it reads as an honest limitation rather than a hidden one.

What's new: the pose-gradient derivation in Sec. 4.2 is a straightforward but useful chain-rule result, and applying it inside an omnidirectional 3D-GS rasterizer is new. The ablation in Table 3 shows pose refinement and the camera model each help, and the robustness curves in Fig. 5 are a nice addition. The paper also does the right thing by testing multiple point-cloud initializations, including random and estimated depth.

The main weaknesses, in order:

1. Camera model identifiability. Eq. 15's D is per-pixel, zero-initialized, Tanh-bounded, shared across views, and optimized purely by photometric plus anisotropy loss. Nothing anchors it to a physical distortion model. The stress-test concern is fair: the warp could absorb pose misalignment or geometry errors. The paper never validates D against known distortion (e.g., synthetic images with a known distortion field), so 'self-calibration' is doing real work the experiments don't isolate. This is addressable, not fatal.

2. The from-scratch claim. On real multi-room 360Roam, training from scratch fails for all methods, including this one — the paper's own Limitation says so. The abstract's 'even no pose prior' is therefore supported only on three synthetic single-room scenes. The claim should be scoped.

3. Engineering hygiene. No code, single-run metrics without error bars, and pose error reported only on training views (Table 7), which doesn't rule out the warp overfitting those views. These are minor-to-moderate, not disqualifying.

Overall: the math in Sec. 4.2 is fine, the citation pattern is appropriate, and the paper is honest about its limits. It deserves a serious referee; if it crossed my desk I'd send it out, with the request that the camera-model identifiability be addressed in revision.

Referee Report

4 major / 7 minor

Summary. SC-OmniGS proposes a self-calibrating omnidirectional Gaussian splatting system for 360-degree radiance field reconstruction. The core idea is to treat equirectangular images as a whole sphere, derive pose gradients inside a differentiable omnidirectional rasterizer, and jointly optimize 3D Gaussians, per-view camera poses, and a shared per-pixel omnidirectional camera model Θ = S·ft + S⊙D defined in Eq. (15). The camera model is used to resample input images into distortion-free images, and training minimizes a weighted spherical photometric loss plus an anisotropy regularizer (Eq. (19)). Experiments on the synthetic OmniBlender dataset and the real-world 360Roam dataset compare against BARF, L2G-NeRF, CamP, 3D-GS, and OmniGS under pose perturbation and from-scratch initialization. The paper reports consistent and often large improvements, especially under pose perturbation, with an additional gain on real data from the learned camera model. The from-scratch no-pose-prior claim is demonstrated only on three synthetic scenes, and the paper's Limitation paragraph concedes that from-scratch training fails on real multi-room scenes.

Significance. If the underlying identifiability concern is resolved, this is a significant contribution: it is the first omnidirectional self-calibrating 3D Gaussian splatting system, the pose-gradient derivation in Sec. 4.2 is clean and useful beyond this paper, the weighted spherical photometric loss is principled, and the reported training efficiency (about 30 minutes, real-time rendering) is attractive. The consistently large gains over perturbed-pose baselines are meaningful. However, the empirical claims currently rest on single-run metrics, no released code, and a high-capacity learned distortion field whose separation of true lens distortion from pose and geometry error is not independently validated. The no-pose-prior claim is also much narrower than the abstract suggests, since it is demonstrated only on synthetic scenes and explicitly fails on real multi-room data. The contribution is therefore promising but conditionally established.

major comments (4)

[§6, Limitation; §5.2, Table 1] The load-bearing premise of the paper is that the per-pixel ray-deflection model Θ = S·ft + S⊙D is identifiable from photometric loss alone. D is a per-scene H×W×3 tensor, initialized to zero, bounded only by Tanh, shared across views, and optimized solely by the photometric-plus-anisotropy objective of Eq. (19); there is no smoothness prior, low-rank constraint, or independent calibration signal. In the no-pose-prior experiments, poses start at the origin, so the optimizer could reduce training loss by absorbing pose misalignment and geometry errors into D rather than by recovering true lens distortion. Because Table 7 evaluates pose errors only on the training views used for optimization, it does not rule out such absorption. I recommend a controlled identifiability experiment: inject a known, non-radial distortion pattern into synthetic OmniBlender images, run the method with clean and with perturbed poses, and report the error of the recovered D against the ground-truth distortion, not just rendering quality. In addition, evaluating D on held-out views, or using the learned D to undistort images before a standard SfM pipeline, would demonstrate that the camera model transfers rather than overfits the training views.
[Table 7; §5.3] The abstract states that the method can recover radiance fields 'even no pose prior', but the paper's own Limitation paragraph concedes that from-scratch training fails on the real multi-room 360Roam scenes, and the from-scratch success in Table 1 is limited to three synthetic OmniBlender scenes with ideal cameras and ground-truth or rendered depth initialization. Since synthetic OmniBlender images contain no real stitch distortion, the from-scratch experiments do not exercise the joint identification of pose and distortion. The claim in the abstract and conclusion should be narrowed to synthetic object-centric scenes, or real-world from-scratch experiments need to be added. As written, the no-pose-prior claim is materially overstated.
[Tables 1–7; §5.2] Camera pose accuracy is reported only for the training views that were used in optimization. This is a necessary but insufficient check for calibration quality: a per-pixel warp D of this capacity can memorize per-view misregistration without producing a physically meaningful camera model. The evaluation should also report pose errors on held-out test views after training, or otherwise demonstrate that the learned D and poses generalize. The test-view rendering metrics in Table 2 are suggestive but do not isolate whether the gain comes from true calibration or from flexible warping that happens to interpolate well on test views.
[§5.1, Eq. (17)] All quantitative metrics are single runs with no error bars or repeated-seed statistics. Given that the method includes stochastic initialization of 3D Gaussians, random point-cloud initialization, and joint optimization of high-capacity parameters, the reported margins—especially the smaller gains in the real-world ablation study (Table 3)—should be accompanied by means and standard deviations over at least a few seeds. Without this, the robustness claims in Sec. 5.4 are not statistically supported.

minor comments (7)

[Eq. (17)] In the second paragraph of Sec. 5.1, 'we evaluated SG-OmniGS' should read 'SC-OmniGS'.
[Algorithm 1] The notation ΔI = wI in Eq. (17) should be clarified as an elementwise product, and w(u) is defined in terms of the row coordinate v but written as a function of u; please make the coordinate dependence explicit.
[Eq. (15)] The line 'S ← φ′(u)' should specify the inverse equirectangular projection and the exact normalization, since φ′ is not defined in the main text and this step is central to the camera model.
[Appendix B.3] Since ft is fixed to 1 throughout the experiments, the term S·ft in Eq. (15) is redundant with the spherical grid S; please state why this term is kept, or fold it into the definition of S.
[Table 5] The perturbation description uses α drawn from [−1, 1) for translation and a normal distribution over an angle range for rotation; please state the exact distribution for β and the units of Rscale so that the perturbation magnitude is reproducible.
[Appendix C.2] The ablation in Table 5 reports only one scene (Classroom) and one seed; adding per-scene results would make the benefit of Lwsp and re-initialization more convincing.
[Table 6] The caption of Table 6 lists “SC-OmniGS” twice in the header; please disambiguate the random-initialization and SfM-initialization columns.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pose-gradient derivation is a self-contained chain-rule calculation, and the camera model's identifiability is an underdetermination concern, not a definitional reduction.

full rationale

SC-OmniGS's central mathematical contribution is the pose-gradient chain rule in Sec. 4.2 (Eqs. 13-14), which follows from differentiating the previously specified omnidirectional splatting equations (Eqs. 9-12) and does not assume the target pose estimates. The camera model of Eq. 15 (Theta = S*ft + S⊙D) is optimized jointly with poses and Gaussians by the photometric loss of Eq. 19, so the separation between learned distortion and pose error is an identifiability issue rather than a circular one: the paper never defines D in terms of the pose residuals, and its test-view PSNR plus the ablation in Table 3 provide independent (if incomplete) evidence that the warp does not simply memorize training views. The only overlapping-author citations (OmniGS for the rasterizer, 360Roam for the dataset) are used as baselines/data and are not invoked as a uniqueness theorem or as the justification for the camera model; the projection equations are stated in full and the dataset is externally collected. The paper's own Limitation paragraph states that no-pose-prior training fails on real multi-room scenes, which narrows but does not circularly justify the abstract claim. Therefore no load-bearing step reduces to its own inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities; the camera model is a mathematical parameterization, not a new force, particle, or dimension. The main load-bearing additions are the fitted distortion field D and per-view poses, plus a few hand-set hyperparameters.

free parameters (4)

Per-scene angle distortion field D = learned per scene, initialized to zeros, activated by tanh
Central to the omnidirectional camera model (Eq. 15); optimized with learning rate 1e-4 on photometric loss, with no external calibration target.
Per-view camera poses [q|t] = optimized per view; initialization varies (SfM, perturbed, or origin)
The object of self-calibration; poses are optimized jointly with the scene and camera model. In the from-scratch setting all cameras start at the origin.
Photometric loss weight lambda = 0.2
Hand-set balance between MAE and SSIM in Eq. 16; not critical to the central claim but is a chosen constant.
Anisotropy ratio threshold gamma = 10
Hand-set threshold in Eq. 18 to penalize filamentous Gaussians; a standard regularizer strength.

assumptions (5)

domain assumption 3D Gaussian splatting with the local affine approximation accurately renders omnidirectional radiance fields (Eqs. 3, 7, 12).
The method relies on 3D-GS and its Jacobian approximation as a faithful rendering model; a failure here would break all photometric optimization.
domain assumption The equirectangular projection in Eq. 9 with focal scale ft fixed to 1 is a valid base model for 360-degree cameras.
The whole pipeline treats 360 images as ideal equirectangular projections plus a learned per-pixel correction; inaccuracies in the base model must be absorbed by D.
ad hoc to paper The per-pixel ray-deflection model Θ = S·ft + S⊙D in Eq. 15 can represent real stitched-camera distortion and is identifiable from photometric loss alone.
This is the central new modeling assumption; it is plausible but unverified, since D has millions of parameters and no calibration target.
domain assumption A single camera model shared across all views of a scene is valid, i.e., distortion is static across the capture.
The method uses one D per scene; if the camera's stitching or lens state varies across views, the model cannot represent it.
domain assumption SfM poses in the 360Roam dataset are accurate enough to serve as evaluation ground truth for real-world pose metrics.
Table 7 compares recovered poses against SfM poses; if SfM is wrong, pose error numbers are unreliable. This is common practice but an assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting." pith.science (2026). https://pith.science/paper/WVC273EB

@misc{pith2026250204734,
  author       = {Pith},
  title        = {Pith review of: SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WVC273EB}},
  note         = {Machine review of arXiv:2502.04734}
}

read the original abstract

360-degree cameras streamline data collection for radiance field 3D reconstruction by capturing comprehensive scene data. However, traditional radiance field methods do not address the specific challenges inherent to 360-degree images. We present SC-OmniGS, a novel self-calibrating omnidirectional Gaussian splatting system for fast and accurate omnidirectional radiance field reconstruction using 360-degree images. Rather than converting 360-degree images to cube maps and performing perspective image calibration, we treat 360-degree images as a whole sphere and derive a mathematical framework that enables direct omnidirectional camera pose calibration accompanied by 3D Gaussians optimization. Furthermore, we introduce a differentiable omnidirectional camera model in order to rectify the distortion of real-world data for performance enhancement. Overall, the omnidirectional camera intrinsic model, extrinsic poses, and 3D Gaussians are jointly optimized by minimizing weighted spherical photometric loss. Extensive experiments have demonstrated that our proposed SC-OmniGS is able to recover a high-quality radiance field from noisy camera poses or even no pose prior in challenging scenarios characterized by wide baselines and non-object-centric configurations. The noticeable performance gain in the real-world dataset captured by consumer-grade omnidirectional cameras verifies the effectiveness of our general omnidirectional camera model in reducing the distortion of 360-degree images.

Figures

Figures reproduced from arXiv: 2502.04734 by the authors.

**Figure 2.** A schematic overview of SC-OmniGS optimization flow. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** For model initialization, we create a spherical grid [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: Qualitative comparisons of 360-degree novel views among calibration methods. Our [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Performance with different camera perturbations (PSNR [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Ablation study of weighted spherical photometric loss [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

**Figure 7.** Figure 7: Qualitative comparisons of 360-degree novel views among calibration methods equipped [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

**Figure 8.** Figure 8: Depth visualization of 360-degree views rendered by calibration methods equipped with omnidirectional sampling. Our results outperform in geometry accuracy and details. † indicates training from scratch, ◦ indicates baselines modified via omnidirectional sampling. 19 …

**Figure 9.** Figure 9: Novel views on synthetic scene Flat among baselines trained from scratch. Ground-truth BARF L2G-NeRF CamP Ours [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

**Figure 10.** Figure 10: Novel views on real scene Cafe among baselines trained with camera perturbation. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

**Figure 11.** Figure 11: Novel views on real scene Bar among baselines trained with camera perturbation. Ground-truth BARF L2G-NeRF CamP Ours [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]

**Figure 12.** Figure 12: Novel views on real scene Base among baselines trained with camera perturbation. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Self-Calibrating Gaussian Splatting for Large Field of View Reconstruction
cs.CV 2025-02 conditional novelty 6.0 of 10

A hybrid invertible-network distortion field plus cubemap rendering lets Gaussian Splatting reconstruct scenes from uncalibrated fisheye photos, outperforming prior fisheye methods and reducing the number of captures needed.

Reference graph

Works this paper leans on

13 extracted references · 8 canonical work pages · cited by 1 Pith paper

[1]

Framework for Easily Invertible Architectures (FrEIA), 2018-2022

Lynton Ardizzone, Till Bungert, Felix Draxler, Ullrich K ¨othe, Jakob Kruse, Robert Schmier, and Peter Sorrenson. Framework for Easily Invertible Architectures (FrEIA), 2018-2022. URL https://github.com/vislearn/FrEIA. Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout- guided panoramic gaussian splatting for indoor ro...

arXiv 2018
[4]

Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras

11 Published as a conference paper at ICLR 2025 Huajian Huang, Longwei Li, Cheng Hui, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024a. Huajian Huang, Changkun Liu, Yipeng Zhu, H...

work page 2025
[5]

360fusionnerf: Panoramic neural radiance fields with joint guidance

Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 7202–7209. IEEE,

work page 2023
[6]

Omnigs: Omnidirectional gaus- sian splatting for fast radiance field reconstruction using omnidirectional images

Longwei Li, Huajian Huang, Sai-Kit Yeung, and Hui Cheng. Omnigs: Omnidirectional gaus- sian splatting for fast radiance field reconstruction using omnidirectional images. arXiv preprint arXiv:2404.03202,

arXiv
[7]

Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T Barron, and Ricardo Martin-Brualla

doi: 10.1145/3528223.3530127. Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T Barron, and Ricardo Martin-Brualla. Camp: Camera preconditioning for neural radiance fields.ACM Transactions on Graphics (TOG), 42(6):1–11,

arXiv
[12]

It provides ground-truth camera parameters, and we additionally rendered a ground-truth depth map of each scene to initialize a sparse point cloud for 3D-GS based methods

contains multi-view 360-degree images rendered from Blender synthetic single indoor scenes under MIT License. It provides ground-truth camera parameters, and we additionally rendered a ground-truth depth map of each scene to initialize a sparse point cloud for 3D-GS based methods. The synthetic Blender scene Classroom is under CC0 license, Barbershop and ...

work page 2025
[13]

Perturb” indicates perturbed camera poses as inputs, “Point Init

Under different scenes and different point cloud initializations, SC-OmniGS outperforms other calibration baselines achieving robust camera calibration capability. Figures 7-8 supplement some qualitative rendering and depth comparisons among adapted calibra- tion baselines with omnidirectional sampling in the scenes same as Figure 4 in the main manuscript...

work page 2025
[2002]

It has broad potential impacts and applications in the real world

13 Published as a conference paper at ICLR 2025 APPENDIX A S OCIETAL IMPACTS This research explored the efficient and robust self-calibrating omnidirectional radiance field for large omnidirectional scenarios, experimenting with real-world data captured with the consumer- grade 360-degree camera and synthetic data. It has broad potential impacts and appli...

work page 2025

Show all 13 references

[2003]

The double sphere camera model

Vladyslav Usenko, Nikolaus Demmel, and Daniel Cremers. The double sphere camera model. In 2018 International Conference on 3D Vision (3DV) , pp. 552–560. IEEE,

2018
[2017]

A perspective on distortions

12 Published as a conference paper at ICLR 2025 Rahul Swaninathan, Michael D Grossberg, and Shree K Nayar. A perspective on distortions. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition,

2025
[2022]

360roam: Real-time indoor roaming using geometry-aware 360◦ radiance fields

Huajian Huang, Yingshu Chen, Tianjia Zhang, and Sai-Kit Yeung. 360roam: Real-time indoor roaming using geometry-aware 360◦ radiance fields. arXiv preprint arXiv:2208.02705,

arXiv
[2023]

Tensorf: Tensorial radiance fields

Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, pp. 333–350. Springer,

2022
[2024]

Nerf–: Neural radiance fields without known camera parameters

Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf–: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064,

arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Framework for Easily Invertible Architectures (FrEIA), 2018-2022

Lynton Ardizzone, Till Bungert, Felix Draxler, Ullrich K ¨othe, Jakob Kruse, Robert Schmier, and Peter Sorrenson. Framework for Easily Invertible Architectures (FrEIA), 2018-2022. URL https://github.com/vislearn/FrEIA. Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout- guided panoramic gaussian splatting for indoor ro...

arXiv 2018

[2] [4]

Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras

11 Published as a conference paper at ICLR 2025 Huajian Huang, Longwei Li, Cheng Hui, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024a. Huajian Huang, Changkun Liu, Yipeng Zhu, H...

work page 2025

[3] [5]

360fusionnerf: Panoramic neural radiance fields with joint guidance

Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 7202–7209. IEEE,

work page 2023

[4] [6]

Omnigs: Omnidirectional gaus- sian splatting for fast radiance field reconstruction using omnidirectional images

Longwei Li, Huajian Huang, Sai-Kit Yeung, and Hui Cheng. Omnigs: Omnidirectional gaus- sian splatting for fast radiance field reconstruction using omnidirectional images. arXiv preprint arXiv:2404.03202,

arXiv

[5] [7]

Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T Barron, and Ricardo Martin-Brualla

doi: 10.1145/3528223.3530127. Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T Barron, and Ricardo Martin-Brualla. Camp: Camera preconditioning for neural radiance fields.ACM Transactions on Graphics (TOG), 42(6):1–11,

arXiv

[6] [12]

It provides ground-truth camera parameters, and we additionally rendered a ground-truth depth map of each scene to initialize a sparse point cloud for 3D-GS based methods

contains multi-view 360-degree images rendered from Blender synthetic single indoor scenes under MIT License. It provides ground-truth camera parameters, and we additionally rendered a ground-truth depth map of each scene to initialize a sparse point cloud for 3D-GS based methods. The synthetic Blender scene Classroom is under CC0 license, Barbershop and ...

work page 2025

[7] [13]

Perturb” indicates perturbed camera poses as inputs, “Point Init

Under different scenes and different point cloud initializations, SC-OmniGS outperforms other calibration baselines achieving robust camera calibration capability. Figures 7-8 supplement some qualitative rendering and depth comparisons among adapted calibra- tion baselines with omnidirectional sampling in the scenes same as Figure 4 in the main manuscript...

work page 2025

[8] [2002]

It has broad potential impacts and applications in the real world

13 Published as a conference paper at ICLR 2025 APPENDIX A S OCIETAL IMPACTS This research explored the efficient and robust self-calibrating omnidirectional radiance field for large omnidirectional scenarios, experimenting with real-world data captured with the consumer- grade 360-degree camera and synthetic data. It has broad potential impacts and appli...

work page 2025