Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

3D Gaussian Splatting can train directly on raw fisheye images, without undistortion, when Gaussians are jointly optimized across overlapping 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 ↗

T0 review · grok-4.5

2026-07-13 14:56 UTC pith:QC77JIN7

load-bearing objection Solid engineering fix for native fisheye 3DGS plus a transferable multi-view regularizer; residual floaters are real, CVO is plausible but under-isolated in the extract. the 3 major comments →

arxiv 2604.00648 v2 pith:QC77JIN7 submitted 2026-04-01 cs.CV

DirectFisheye-GS: Enabling Native Fisheye Input in Gaussian Splatting with Cross-View Joint Optimization

Zhengxian Yang , Fei Xie , Xutao Xue , Rui Zhang , Taicheng Huang , Yang Liu , Mengqi Ji , Tao Yu This is my paper
classification cs.CV
keywords 3D Gaussian Splattingfisheye camerasnative fisheye inputcross-view joint optimizationnovel view synthesisundistortion artifactsreal-time rendering
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.

Fisheye cameras capture a much wider field of view than ordinary lenses, so they should let 3D Gaussian Splatting rebuild a scene from fewer photos. In practice most pipelines first “undistort” the fisheye frames into pinhole images; that step throws away edge pixels, invents black borders, and stretches detail into low-frequency patches that the optimizer overfits, producing blur and floating artifacts. This paper embeds the fisheye projection model inside the original rasterizer so training can use the raw fisheye photos unchanged. Even with correct projection, random single-view updates still let peripheral Gaussians grow into extreme shapes because distortion rises toward the image edge and cross-view consistency is ignored. A feature-overlap-driven joint optimization step then forces the same Gaussians to satisfy geometric and photometric constraints from several views at once. On public benchmarks the method matches or exceeds prior results while preserving more peripheral detail, and the joint step also helps ordinary pinhole pipelines.

Core claim

Integrating a fisheye camera model into the original 3D Gaussian Splatting rasterizer enables native fisheye training without any undistortion preprocessing; residual edge floaters are then suppressed by a feature-overlap-driven cross-view joint optimization that imposes consistent geometric and photometric constraints across views, yielding reconstruction quality that matches or surpasses state-of-the-art methods on public datasets.

What carries the argument

Feature-overlap-driven cross-view joint optimization: Gaussians that share feature overlap across multiple views are selected and updated together so their shapes and colors cannot diverge into oversized or elongated forms at the distorted periphery.

Load-bearing premise

The method treats a feature-overlap score as a reliable proxy for true multi-view geometric correspondence; if that score is noisy or biased, the joint update can still leave extreme Gaussian shapes.

What would settle it

On the same fisheye sequences, disable only the feature-overlap joint step while keeping native fisheye projection; if peripheral floaters and edge-region metrics then fail to degrade relative to the full method, the claim that joint optimization is what removes residual floaters is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on 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

3 major / 4 minor

Summary. The paper proposes DirectFisheye-GS, which integrates a fisheye camera model directly into the 3D Gaussian Splatting rasterization pipeline so that training can use native fisheye images without undistortion preprocessing. The authors argue that undistortion wastes FOV (black borders) and dilutes high-frequency detail via stretch-and-interpolate resampling, causing blur and floaters. After correct fisheye projection they still observe peripheral floaters, which they attribute to 3DGS’s per-iteration single-view random optimization producing extreme Gaussian shapes under strong peripheral distortion. They introduce a feature-overlap–driven cross-view joint optimization (CVO) that selects Gaussians for multi-view geometric and photometric updates, claiming this regularizer is also useful for ordinary pinhole pipelines. Qualitative results and ablations (Figs. 1, 9, 10) and claims of matching or surpassing SOTA on ScanNet++ and Den-SOFT are presented.

Significance. Native fisheye support for 3DGS is a practically useful engineering contribution for wide-FOV capture in VR/AR and robotics, where undistortion is known to discard information and create low-frequency artifacts. Framing residual floaters as a cross-view consistency problem and proposing a general CVO regularizer is a reasonable and potentially transferable idea beyond fisheye. If the quantitative gains hold under controlled ablations and the overlap selection is shown to be robust, the work would be a solid incremental advance for the 3DGS community. The public project page and multi-dataset evaluation are positive reproducibility signals.

major comments (3)
  1. The load-bearing claim that feature-overlap CVO is what removes residual peripheral floaters (after correct fisheye projection) rests on an unvalidated proxy. The abstract and introduction state that single-view optimization produces extreme shapes and that CVO establishes consistent geometric/photometric constraints via feature overlap. Figure 10 shows qualitative ablation benefit, but the manuscript extract supplies no quantitative isolation of the selection criterion (e.g., PSNR/SSIM/LPIPS with vs. without CVO under identical densification, and with random multi-view pairing as a control), no sensitivity analysis of the overlap threshold/schedule, and no verification that selected Gaussians share consistent 3D support across views under strong fisheye distortion. Without these, quality gains may be driven mainly by native projection rather than the claimed regularizer.
  2. Quantitative support for “matches or surpasses state-of-the-art on public datasets” is incomplete in the provided manuscript body. Claims reference ScanNet++ and Den-SOFT and cite baselines (Fisheye-GS, 3DGUT, MVGS, etc.), yet full comparison tables with standard metrics, number of views, training iterations, and Gaussian counts are not present in the extract; only qualitative Figs. 9–10 appear. For a journal claim of SOTA parity/superiority, complete tables (including error bars or multiple runs if variance is material) and an explicit statement of which baselines used undistorted vs. native fisheye inputs are required so the contribution of native modeling vs. CVO can be assessed.
  3. The free parameters of CVO (feature-overlap selection threshold/schedule and interaction with standard 3DGS densification/pruning) are not characterized. Because the method invents a new selection step that decides which Gaussians receive joint multi-view updates, the paper should report how sensitive final quality and floater rates are to that threshold, and whether the same schedule transfers from fisheye to pinhole (as claimed). Absent this, the “equally applicable to pinhole pipelines” claim remains an untested assertion rather than a demonstrated result.
minor comments (4)
  1. The supplied manuscript text jumps from the start of the introduction to Acknowledgments/References and late figures; methods equations, algorithm boxes, and full experimental protocol are missing from the extract. Ensure the camera-model integration (Kannala–Brandt or equivalent) and the exact CVO loss/selection formula appear with numbered equations and a clear algorithm listing.
  2. Figure 10 caption states “ablation studies on cross-view joint optimization (CVO) with fisheye or pinhole camera inputs” but does not define the exact variants (native only / CVO only / both) or report corresponding metrics next to the images; add a small quantitative inset or companion table.
  3. Related-work placement of concurrent fisheye/distorted-camera GS works (Fisheye-GS [24], 3DGUT [39], Self-calibrating GS [5]) should more explicitly contrast rasterization-native vs. ray-based or undistort-then-train pipelines so the novelty boundary is crisp.
  4. Minor prose: “random-selecting-view optimization” and “feature-overlap–driven” are slightly awkward; standardize terminology (e.g., “single-view random sampling” vs. “cross-view joint optimization”) throughout.

Circularity Check

0 steps flagged

No circular derivation: native fisheye projection plus CVO is an empirical method evaluated on external public datasets, not a result forced by its own definitions or self-citation.

full rationale

DirectFisheye-GS is a systems/methods paper. Its load-bearing claims are (1) integrating a fisheye camera model into the 3DGS rasterizer so training can use native fisheye images without undistortion, and (2) a feature-overlap-driven cross-view joint optimization (CVO) that regularizes Gaussians across views to reduce peripheral floaters. Neither claim is obtained by defining a quantity in terms of the quantity being predicted, nor by fitting a free parameter and then reporting a closely related statistic as a prediction. Performance is measured against external public datasets (e.g., ScanNet++, Den-SOFT) and published baselines (Fisheye-GS, 3DGUT, etc.); the reported PSNR/SSIM/LPIPS-style gains are not algebraically forced by the training objective. Self-citations (e.g., Den-SOFT) supply evaluation data or ordinary background and are not used as uniqueness theorems that forbid alternatives. The skeptic concern that feature-overlap is an imperfect proxy for multi-view correspondence is a validity/assumption risk, not circularity: the method does not redefine success as the overlap measure itself. No step reduces Eq. X to Eq. Y by construction. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

The paper rests on the standard 3DGS representation and optimization loop, a classical fisheye projection model (Kannala-Brandt style), and the empirical claim that feature-overlap is a useful proxy for multi-view consistency. No new physical entities are postulated. Free parameters are the usual 3DGS hyper-parameters plus whatever thresholds define “feature overlap” and the joint-update schedule; none are given numeric values in the extract.

free parameters (2)
  • feature-overlap selection threshold / schedule
    Controls which Gaussians enter the joint multi-view update; value is chosen by the authors and not derived from first principles.
  • standard 3DGS densification / pruning / learning-rate schedule
    Inherited from Kerbl et al. and possibly retuned for fisheye; affects final quality but is not the scientific claim.
axioms (3)
  • domain assumption 3D scene can be adequately represented by a set of anisotropic 3D Gaussians with spherical-harmonic appearance (Kerbl et al. 2023).
    Foundation of the entire pipeline; taken as given.
  • domain assumption A classical polynomial / equidistant fisheye projection model correctly maps 3D points to the observed image plane for the cameras used.
    Invoked when the authors “integrate fisheye camera model into the original 3DGS framework.”
  • ad hoc to paper Feature overlap across views is a reliable proxy for geometric correspondence of the same Gaussian.
    Core of the proposed cross-view joint optimization; not independently proven outside the paper’s ablations.
invented entities (1)
  • feature-overlap-driven cross-view joint optimization (CVO) no independent evidence
    purpose: To impose consistent geometric and photometric constraints on Gaussians that appear in multiple views, preventing extreme shapes at high-distortion periphery.
    New algorithmic construct introduced by the paper; independent evidence is limited to the paper’s own ablations and qualitative figures.

pith-pipeline@v1.1.0-grok45 · 11309 in / 2726 out tokens · 20089 ms · 2026-07-13T14:56:16.961778+00:00 · methodology

0 comments
read the original abstract

3D Gaussian Splatting (3DGS) has enabled efficient 3D scene reconstruction from everyday images with real-time, high-fidelity rendering, greatly advancing VR/AR applications. Fisheye cameras, with their wider field of view (FOV), promise high-quality reconstructions from fewer inputs and have recently attracted much attention. However, since 3DGS relies on rasterization, most subsequent works involving fisheye camera inputs first undistort images before training, which introduces two problems: 1) Black borders at image edges cause information loss and negate the fisheye's large FOV advantage; 2) Undistortion's stretch-and-interpolate resampling spreads each pixel's value over a larger area, diluting detail density -- causes 3DGS overfitting these low-frequency zones, producing blur and floating artifacts. In this work, we integrate fisheye camera model into the original 3DGS framework, enabling native fisheye image input for training without preprocessing. Despite correct modeling, we observed that the reconstructed scenes still exhibit floaters at image edges: Distortion increases toward the periphery, and 3DGS's original per-iteration random-selecting-view optimization ignores the cross-view correlations of a Gaussian, leading to extreme shapes (e.g., oversized or elongated) that degrade reconstruction quality. To address this, we introduce a feature-overlap-driven cross-view joint optimization strategy that establishes consistent geometric and photometric constraints across views-a technique equally applicable to existing pinhole-camera-based pipelines. Our DirectFisheye-GS matches or surpasses state-of-the-art performance on public datasets. Project Page: https://yzxqh.github.io/DirectFisheye-GS/ .

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. UniSHARP: Universal Sharp Monocular View Synthesis

    cs.CV 2026-06 unverdicted novelty 6.0

    UniSHARP performs universal sharp monocular view synthesis by implicit alignment of diverse camera images in a unified omnidirectional latent space using ray-arranged Gaussian primitives and UniK3D-inspired feature decoding.

Reference graph

Works this paper leans on

50 extracted references · 10 linked inside Pith · cited by 1 Pith paper

  1. [1]

    360-gs: Layout-guided panoramic gaussian splatting for indoor roaming.arXiv preprint arXiv:2402.00763, 2024

    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. 3

    Pith/arXiv arXiv 2024

  2. [2]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130–141, 2023. 2

    2023

  3. [3]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean con- ference on computer vision, pages 333–350. Springer, 2022. 2

    2022

  4. [4]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. InEuropean Conference on Computer Vision, pages 370–386. Springer, 2024. 2

    2024

  5. [5]

    Self-calibrating gaussian splatting for large field of view reconstruction.arXiv preprint arXiv:2502.09563,

    Youming Deng, Wenqi Xian, Guandao Yang, Leonidas Guibas, Gordon Wetzstein, Steve Marschner, and Paul De- bevec. Self-calibrating gaussian splatting for large field of view reconstruction.arXiv preprint arXiv:2502.09563,

    Pith/arXiv arXiv

  6. [6]

    Mvgs: Multi-view- regulated gaussian splatting for novel view synthesis.arXiv preprint arXiv:2410.02103, 2024

    Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view- regulated gaussian splatting for novel view synthesis.arXiv preprint arXiv:2410.02103, 2024. 3, 4, 7, 8

    Pith/arXiv arXiv 2024

  7. [7]

    Distwar: Fast differentiable rendering on raster-based ren- dering pipelines.arXiv preprint arXiv:2401.05345, 2023

    Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, and Nandita Vijaykumar. Distwar: Fast differentiable rendering on raster-based ren- dering pipelines.arXiv preprint arXiv:2401.05345, 2023. 3

    Pith/arXiv arXiv 2023

  8. [8]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, Zhangyang Wang, et al. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. Advances in neural information processing systems, 37: 140138–140158, 2024. 3

    2024

  9. [9]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2

    2022

  10. [10]

    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. In2022 IEEE International Conference on Multi- media and Expo (ICME), pages 1–6. IEEE, 2022. 3

    2022

  11. [11]

    Instruct-nerf2nerf: Edit- ing 3d scenes with instructions

    Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Edit- ing 3d scenes with instructions. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 2

    2023

  12. [12]

    Self-calibrating neural radiance fields

    Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 5846– 5854, 2021. 3, 6, 7

    2021

  13. [13]

    Identifying unnecessary 3d gaussians using clustering for fast rendering of 3d gaussian splatting.arXiv preprint arXiv:2402.13827,

    Joongho Jo, Hyeongwon Kim, and Jongsun Park. Identifying unnecessary 3d gaussians using clustering for fast rendering of 3d gaussian splatting.arXiv preprint arXiv:2402.13827,

    Pith/arXiv arXiv

  14. [14]

    Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses.IEEE transactions on pattern analysis and machine intelligence, 28(8):1335–1340, 2006. 2, 3, 4

    2006

  15. [15]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    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,

  16. [16]

    A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024

    Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024. 2

    2024

  17. [17]

    3d gaussian splat- ting as markov chain monte carlo.Advances in Neural Infor- mation Processing Systems, 37:80965–80986, 2024

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splat- ting as markov chain monte carlo.Advances in Neural Infor- mation Processing Systems, 37:80965–80986, 2024. 2

    2024

  18. [18]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics, 36(4), 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics, 36(4), 2017. 3

    2017

  19. [19]

    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 2023 ieee. InRSJ International Conference on In- telligent Robots and Systems, pages 7202–7209. 3

    2023

  20. [20]

    Compact 3d gaussian representation for radiance field

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719– 21728, 2024. 2, 3

    2024

  21. [21]

    Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion

    Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 20775–20785,

  22. [22]

    Omnigs: Fast radiance field reconstruction using omnidirec- tional gaussian splatting

    Longwei Li, Huajian Huang, Sai-Kit Yeung, and Hui Cheng. Omnigs: Fast radiance field reconstruction using omnidirec- tional gaussian splatting. In2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV), pages 2260–2268. IEEE, 2025. 3

    2025

  23. [23]

    Uhdnerf: ultra-high-definition neural radiance fields

    Quewei Li, Feichao Li, Jie Guo, and Yanwen Guo. Uhdnerf: ultra-high-definition neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 23097–23108, 2023. 2

    2023

  24. [24]

    Fisheye-gs: Lightweight and extensible gaus- sian splatting module for fisheye cameras.arXiv preprint arXiv:2409.04751, 2024

    Zimu Liao, Siyan Chen, Rong Fu, Yi Wang, Zhongling Su, Hao Luo, Li Ma, Linning Xu, Bo Dai, Hengjie Li, et al. Fisheye-gs: Lightweight and extensible gaus- sian splatting module for fisheye cameras.arXiv preprint arXiv:2409.04751, 2024. 2, 3, 6

    Pith/arXiv arXiv 2024

  25. [25]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, You- liang Yan, et al. Vastgaussian: Vast 3d gaussians for large scene reconstruction. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5166–5175, 2024. 2, 3

    2024

  26. [26]

    Neural sparse voxel fields.Advances in Neural Information Processing Systems, 33:15651–15663,

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields.Advances in Neural Information Processing Systems, 33:15651–15663,

  27. [27]

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians

    Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Jun- ran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision, pages 265–282. Springer, 2024. 2, 3

    2024

  28. [28]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 2, 3

    2024

  29. [29]

    Ever: Exact volumet- ric ellipsoid rendering for real-time view synthesis.arXiv preprint arXiv:2410.01804, 2024

    Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T Barron, and Yinda Zhang. Ever: Exact volumet- ric ellipsoid rendering for real-time view synthesis.arXiv preprint arXiv:2410.01804, 2024. 3

    Pith/arXiv arXiv 2024

  30. [30]

    Taming 3dgs: High-quality radiance fields with limited resources

    Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fer- nando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. InSIGGRAPH Asia 2024 Con- ference Papers, pages 1–11, 2024. 2

    2024

  31. [31]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    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

    2021

  32. [32]

    3d gaussian ray trac- ing: Fast tracing of particle scenes.ACM Transactions on Graphics (TOG), 43(6):1–19, 2024

    Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ric- cardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray trac- ing: Fast tracing of particle scenes.ACM Transactions on Graphics (TOG), 43(6):1–19, 2024. 3

    2024

  33. [33]

    Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022. 2

    2022

  34. [34]

    Structure-from-motion revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-motion revisited. InConference on Com- puter Vision and Pattern Recognition (CVPR), 2016. 5, 6

    2016

  35. [35]

    V olumetric rendering with baked quadrature fields

    Gopal Sharma, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. V olumetric rendering with baked quadrature fields. InEuropean Conference on Computer Vision, pages 275–292. Springer, 2024. 2

    2024

  36. [36]

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

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459– 5469, 2022. 2

    2022

  37. [37]

    Hybridnerf: Efficient neu- ral rendering via adaptive volumetric surfaces

    Haithem Turki, Vasu Agrawal, Samuel Rota Bul `o, Lorenzo Porzi, Peter Kontschieder, Deva Ramanan, Michael Zollh¨ofer, and Christian Richardt. Hybridnerf: Efficient neu- ral rendering via adaptive volumetric surfaces. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19647–19656, 2024. 2

    2024

  38. [38]

    Perf: Panoramic neural radiance field from a single panorama.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024

    Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024. 3

    2024

  39. [39]

    3dgut: Enabling distorted cameras and secondary rays in gaussian splatting.arXiv preprint arXiv:2412.12507, 2024

    Qi Wu, Janick Martinez Esturo, Ashkan Mirzaei, Nicolas Moenne-Loccoz, and Zan Gojcic. 3dgut: Enabling distorted cameras and secondary rays in gaussian splatting.arXiv preprint arXiv:2412.12507, 2024. 2, 3, 6

    Pith/arXiv arXiv 2024

  40. [40]

    Fbinerf: Feature-based integrated recurrent network for pin- hole and fisheye neural radiance fields.arXiv preprint arXiv:2408.01878, 2024

    Yifan Wu, Tianyi Cheng, Peixu Xin, and Janusz Konrad. Fbinerf: Feature-based integrated recurrent network for pin- hole and fisheye neural radiance fields.arXiv preprint arXiv:2408.01878, 2024. 3

    Pith/arXiv arXiv 2024

  41. [41]

    Neural lens modeling

    Wenqi Xian, Alja ˇz Bo ˇziˇc, Noah Snavely, and Christoph Lassner. Neural lens modeling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8435–8445, 2023. 3

    2023

  42. [42]

    Point- nerf: Point-based neural radiance fields

    Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point- nerf: Point-based neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438–5448, 2022. 2

    2022

  43. [43]

    Msi-nerf: Linking omni-depth with view synthesis through multi-sphere image aided generalizable neural radi- ance field

    Dongyu Yan, Guanyu Huang, Fengyu Quan, and Haoyao Chen. Msi-nerf: Linking omni-depth with view synthesis through multi-sphere image aided generalizable neural radi- ance field. In2025 IEEE/CVF Winter Conference on Ap- plications of Computer Vision (WACV), pages 2517–2526. IEEE, 2025. 3

    2025

  44. [44]

    Multi-scale 3d gaussian splatting for anti-aliased rendering

    Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 3

    2024

  45. [45]

    Bakedsdf: Meshing neural sdfs for real- time view synthesis

    Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real- time view synthesis. InACM SIGGRAPH 2023 Conference Proceedings, pages 1–9, 2023. 2

    2023

  46. [46]

    Scannet++: A high-fidelity dataset of 3d in- door scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d in- door scenes. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12–22, 2023. 6, 8

    2023

  47. [47]

    Den-soft: Dense space-oriented light field dataset for 6-dof immersive experience.arXiv preprint arXiv:2403.09973, 2024

    Xiaohang Yu, Zhengxian Yang, Shi Pan, Yuqi Han, Haox- iang Wang, Jun Zhang, Shi Yan, Borong Lin, Lei Yang, Tao Yu, et al. Den-soft: Dense space-oriented light field dataset for 6-dof immersive experience.arXiv preprint arXiv:2403.09973, 2024. 6, 8, 4

    Pith/arXiv arXiv 2024

  48. [48]

    Mip-splatting: Alias-free 3d gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 19447–19456,

  49. [49]

    Cor-gs: sparse-view 3d gaussian splatting via co-regularization

    Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatting via co-regularization. InEuropean Conference on Computer Vision, pages 335–352. Springer, 2024. 2

    2024

  50. [50]

    Fsgs: Real-time few-shot view synthesis using gaussian splatting

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. InEuropean conference on computer vision, pages 145–163. Springer, 2024. 2 Figure 9. Qualitative comparison on Den-SOFT [47] dataset. Our method achieves the best results in both indoor and outdoor large-scale scenes. Figure 10. Qualit...

    2024