pith. sign in

arxiv: 2604.17831 · v1 · submitted 2026-04-20 · 💻 cs.CV · cs.GR

PCM-NeRF: Probabilistic Camera Modeling for Neural Radiance Fields under Pose Uncertainty

Shravan Venkatraman , Rakesh Raj Madavan , Pavan Kumar Sathya Venkatesh This is my paper

Pith reviewed 2026-05-10 05:20 UTC · model grok-4.3

classification 💻 cs.CV cs.GR
keywords neural radiance fieldspose uncertaintysurface reconstructionneural renderingcamera modelingSfM
0
0 comments X

The pith

Modeling each camera pose as a distribution with learnable uncertainty enables robust neural surface reconstruction despite SfM pose errors.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces PCM-NeRF to address the problem of inaccurate camera poses in neural surface reconstruction. It augments the SG-NeRF method by representing each camera pose as a probabilistic distribution with a learnable mean and variance. The variance is initialized from SfM quality and used to modulate learning rates, damping updates from uncertain views. This prevents bad poses from corrupting the 3D model, leading to better performance on challenging scenes with outliers. A sympathetic reader would care because real-world pose estimation is rarely perfect, and this offers a lightweight fix.

Core claim

PCM-NeRF augments neural surface reconstruction with per-camera learnable uncertainty. Each pose is represented as a distribution with learnable mean and variance, initialized from SfM correspondence quality. An uncertainty regularization loss couples the learned variance to view confidence, and the uncertainty directly modulates the effective pose learning rate, giving uncertain cameras damped gradient updates. This lightweight mechanism requires no changes to the rendering pipeline and adds negligible overhead.

What carries the argument

Per-camera uncertainty initialized from SfM correspondence quality that modulates pose learning rate via an uncertainty regularization loss

If this is right

  • Reconstructions achieve lower Chamfer Distance and higher F-Score on scenes with severe pose outliers.
  • The method works without requiring foreground masks.
  • Performance gains are largest for geometrically complex structures.
  • The addition adds negligible overhead and requires no changes to the rendering pipeline.

Where Pith is reading between the lines

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

  • The per-camera uncertainty idea could transfer to other neural rendering pipelines that optimize poses jointly with geometry.
  • It might reduce the need for expensive pose refinement steps in large-scale capture pipelines.
  • Synthetic experiments with controlled Gaussian noise on camera rotations and translations would isolate how much error the damping tolerates.

Load-bearing premise

Initializing per-camera variance from SfM correspondence quality and coupling it via uncertainty regularization will reliably identify and dampen corrupting views without introducing new optimization instabilities or biases.

What would settle it

Compare Chamfer Distance and F-Score on a scene with deliberately added pose outliers when the uncertainty modulation is enabled versus when it is replaced by uniform learning rates.

Figures

Figures reproduced from arXiv: 2604.17831 by Pavan Kumar Sathya Venkatesh, Rakesh Raj Madavan, Shravan Venkatraman.

Figure 1
Figure 1. Figure 1: Left: Qualitative reconstruction showing PCM-NeRF’s ability to capture fine surface detail. Right: PCM-NeRF outperforms all compared methods across both Chamfer Distance (lower is better) and F-Score (higher is better) Abstract Neural surface reconstruction methods typically treat cam￾era poses as fixed values, assuming perfect accuracy from Structure-from-Motion (SfM) systems. This assumption breaks down … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our PCM-NeRF pipeline. From left to right: (1) Multi-view images are processed to establish feature correspon [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Volumetric Distribution Alignment (VDA) for geo￾metric consistency. Before alignment (left), initial camera pose uncertainty (σR, σt) causes intersecting rays from views Ci and Cj to produce spatially separated density blobs along the ray dis￾tance t. Our VDA optimization (right) utilizes the volumetric con￾sistency loss to align top-k density peaks. By representing density as a Mixture of Gaussians (MoG),… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the Uncertainty Feedback Loop. The framework establishes a dynamic link between view reliability and pose precision. (1) Initialization: Initial reliability is computed from feature correspondences Mi,j to provide a starting prior for camera uncertainty. (2) Dynamic Performance: A feedback buffer Bi tracks rolling PSNR scores to compute the dynamic confidence γ (t) i = (1 − α)γ (0) i + αγˆ (t) … view at source ↗
Figure 5
Figure 5. Figure 5: Camera Viewpoint Distribution. A 3D visualization of the camera positions for a representative scene (e.g., ’Farmer’) after outlier pose removal. Each blue point represents a camera origin Oi in SE(3), illustrating the dense hemispherical coverage required for capturing complex surface geometry in our inward￾facing dataset. fixed covariance of 0.1, fused via normalized weighted sum. The volumetric IoU loss… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of weights on Chamfer Distance and F-Score. Lower IoU weights combined with moderate uncertainty weights consistently yield better reconstruction performance [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparisons across Clock, Deaf, and Farmer scenes. PCM-NeRF recovers sharper surface detail and more complete [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of reconstructions under different weight configurations. Higher IoU weights (0.8) produce over [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
read the original abstract

Neural surface reconstruction methods typically treat camera poses as fixed values, assuming perfect accuracy from Structure-from-Motion (SfM) systems. This assumption breaks down with imperfect pose estimates, leading to distorted or incomplete reconstructions. We present PCM-NeRF, a probabilistic framework that augments neural surface reconstruction with per-camera learnable uncertainty, built on top of SG-NeRF. Rather than treating all cameras equally throughout optimization, we represent each pose as a distribution with a learnable mean and variance, initialized from SfM correspondence quality. An uncertainty regularization loss couples the learned variance to view confidence, and the resulting uncertainty directly modulates the effective pose learning rate: uncertain cameras receive damped gradient updates, preventing poorly initialized views from corrupting the reconstruction. This lightweight mechanism requires no changes to the rendering pipeline and adds negligible overhead. Experiments on challenging scenes with severe pose outliers demonstrate that PCM-NeRF consistently outperforms state-of-the-art methods in both Chamfer Distance and F-Score, particularly for geometrically complex structures, without requiring foreground masks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. PCM-NeRF augments SG-NeRF with per-camera learnable pose uncertainty, where each pose is modeled as a distribution with mean and variance initialized from SfM correspondence quality. An uncertainty regularization loss links the variance to view confidence, modulating the pose learning rate so that uncertain cameras receive damped updates. This is claimed to prevent corruption from poor pose estimates in neural surface reconstruction, with experiments showing superior Chamfer Distance and F-Score on challenging scenes with severe pose outliers, without foreground masks.

Significance. Should the central claim hold, the method offers a practical, low-overhead solution to a common issue in NeRF applications where SfM poses are inaccurate. By avoiding the need for foreground masks and integrating seamlessly with existing pipelines, it could facilitate more reliable reconstructions in uncontrolled environments. The probabilistic modeling of poses is a natural extension that merits further exploration if validated.

major comments (1)
  1. The initialization of per-camera variance from SfM correspondence quality is central to the framework's ability to identify corrupting views. In the presence of severe and potentially global pose outliers, this initialization may fail to provide differentiated signals, as SfM correspondences could be degraded across the board, leaving the uncertainty regularization loss without a clear mechanism to selectively dampen updates.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment point by point below.

read point-by-point responses

  1. Referee: The initialization of per-camera variance from SfM correspondence quality is central to the framework's ability to identify corrupting views. In the presence of severe and potentially global pose outliers, this initialization may fail to provide differentiated signals, as SfM correspondences could be degraded across the board, leaving the uncertainty regularization loss without a clear mechanism to selectively dampen updates.

    Authors: We appreciate this concern regarding potential limitations in initialization. The per-camera variances are not fixed after initialization but are learnable parameters jointly optimized with the scene representation. The uncertainty regularization loss explicitly couples variance to view confidence derived from reconstruction consistency during optimization, creating a dynamic adjustment mechanism: views that align well with the emerging geometry receive lower variance (and thus higher effective learning rates), while inconsistent views are assigned higher variance to dampen their pose updates. This feedback occurs regardless of whether initial SfM signals are uniformly degraded, as the loss evaluates contribution on-the-fly. In our experiments on scenes with severe pose outliers, this enabled selective damping without foreground masks. We acknowledge that in a hypothetical case of perfectly uniform global degradation across all views, differentiation would be limited, but such uniformity is rare in practice and our results demonstrate robustness on challenging data. revision: no

Circularity Check

0 steps flagged

No significant circularity; framework adds independent parameters and loss on SG-NeRF base

full rationale

The derivation introduces per-camera learnable variance initialized from external SfM correspondence quality, an uncertainty regularization loss, and modulation of pose learning rate. These are presented as additive mechanisms whose grounding is external to the core reconstruction optimization. No equations reduce a claimed prediction or result to the fitted inputs by construction, no self-citation chain is load-bearing for the central claim, and the experimental outperformance is not asserted as a mathematical identity. The paper remains self-contained against the stated benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on a small number of domain assumptions about SfM initialization and the effectiveness of uncertainty-based damping; no new physical entities are postulated.

free parameters (1)
  • per-camera pose variance
    Learnable variance parameter for each camera's pose distribution, initialized from SfM correspondence quality.
axioms (1)
  • domain assumption SfM correspondence quality provides a reliable starting point for per-camera uncertainty
    Used to initialize the variance of each pose distribution.

pith-pipeline@v0.9.0 · 5486 in / 1195 out tokens · 48984 ms · 2026-05-10T05:20:26.100111+00:00 · methodology

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

53 extracted references · 53 canonical work pages · 1 internal anchor

  1. [1]

    Nerfels: Renderable neural codes for improved camera pose estima- tion, 2022

    Gil Avraham, Julian Straub, Tianwei Shen, Tsun-Yi Yang, Hugo Germain, Chris Sweeney, Vasileios Balntas, David Novotny, Daniel DeTone, and Richard Newcombe. Nerfels: Renderable neural codes for improved camera pose estima- tion, 2022. 2

    work page 2022

  2. [2]

    Porf: Pose residual field for accurate neural sur- face reconstruction, 2024

    Jia-Wang Bian, Wenjing Bian, Victor Adrian Prisacariu, and Philip Torr. Porf: Pose residual field for accurate neural sur- face reconstruction, 2024. 7, 8

    work page 2024

  3. [3]

    W. Bian, Z. Wang, K. Li, J. W. Bian, and V . A. Prisacariu. Nope-NeRF: Optimising Neural Radiance Field with No Pose Prior. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4160–4169, 2023. 2

    work page 2023

  4. [4]

    Recovering fine details for neural implicit surface reconstruction, 2022

    Decai Chen, Peng Zhang, Ingo Feldmann, Oliver Schreer, and Peter Eisert. Recovering fine details for neural implicit surface reconstruction, 2022. 2

    work page 2022

  5. [5]

    Y . Chen, X. Chen, X. Wang, Q. Zhang, Y . Guo, Y . Shan, and F. Wang. Local-to-Global Registration for Bundle-Adjusting Neural Radiance Fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8264–8273, 2023. 2, 7, 8

    work page 2023

  6. [6]

    Sg-nerf: Neural surface recon- struction with scene graph optimization, 2024

    Yiyang Chen, Siyan Dong, Xulong Wang, Lulu Cai, Youyi Zheng, and Yanchao Yang. Sg-nerf: Neural surface recon- struction with scene graph optimization, 2024. 2, 4, 5, 6, 7, 8

    work page 2024

  7. [7]

    Improv- ing robustness for joint optimization of camera poses and decomposed low-rank tensorial radiance fields, 2024

    Bo-Yu Cheng, Wei-Chen Chiu, and Yu-Lun Liu. Improv- ing robustness for joint optimization of camera poses and decomposed low-rank tensorial radiance fields, 2024. 2, 7, 8

    work page 2024

  8. [8]

    Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs,

    Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs,

    work page

  9. [9]

    S. F. Chng, S. Ramasinghe, J. Sherrah, and S. Lucey. Gaus- sian Activated Neural Radiance Fields for High Fidelity Re- construction and Pose Estimation. InEuropean Confer- ence on Computer Vision (ECCV), pages 264–280. Springer,

    work page

  10. [10]

    Light3r-sfm: Towards feed-forward structure- from-motion, 2025

    Sven Elflein, Qunjie Zhou, S ´ergio Agostinho, and Laura Leal-Taix´e. Light3r-sfm: Towards feed-forward structure- from-motion, 2025. 2

    work page 2025

  11. [11]

    Neural reprojection error: Merging feature learning and cam- era pose estimation, 2021

    Hugo Germain, Vincent Lepetit, and Guillaume Bourmaud. Neural reprojection error: Merging feature learning and cam- era pose estimation, 2021. 2

    work page 2021

  12. [12]

    Bayes’ rays: Uncertainty quantifica- tion for neural radiance fields, 2023

    Lily Goli, Cody Reading, Silvia Sell ´an, Alec Jacobson, and Andrea Tagliasacchi. Bayes’ rays: Uncertainty quantifica- tion for neural radiance fields, 2023. 2

    work page 2023

  13. [13]

    Uc-nerf: Uncertainty-aware conditional neural radiance fields from endoscopic sparse views, 2024

    Jiaxin Guo, Jiangliu Wang, Ruofeng Wei, Di Kang, Qi Dou, and Yun hui Liu. Uc-nerf: Uncertainty-aware conditional neural radiance fields from endoscopic sparse views, 2024. 2

    work page 2024

  14. [14]

    Cavalheiro, and Sertac Karaman

    Juyeop Han, Lukas Lao Beyer, Guilherme V . Cavalheiro, and Sertac Karaman. Nvins: Robust visual inertial navigation fused with nerf-augmented camera pose regressor and un- certainty quantification, 2024. 3

    work page 2024

  15. [15]

    Kim, and Jin-Hwa Kim

    Hwan Heo, Taekyung Kim, Jiyoung Lee, Jaewon Lee, Soohyun Kim, Hyunwoo J. Kim, and Jin-Hwa Kim. Robust camera pose refinement for multi-resolution hash encoding,

    work page

  16. [16]

    Self- calibrating neural radiance fields, 2021

    Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Ani- mashree Anandkumar, Minsu Cho, and Jaesik Park. Self- calibrating neural radiance fields, 2021. 2, 7, 8

    work page 2021

  17. [17]

    Dickerson

    A. Jignasu, E. Herron, Z. Jiang, S. Sarkar, C. Hegde, B. Ganapathysubramanian, A. Balu, and A. Krishnamurthy. STITCH: Surface reconstrucTion using Implicit neural rep- resentations with Topology Constraints and persistent Ho- mology.https://doi.org/10.48550/arxiv. 2412.18696, 2024. arXiv:2412.18696. 1

    work page internal anchor Pith review doi:10.48550/arxiv 2024

  18. [18]

    Modelling uncertainty in deep learning for camera relocalization, 2016

    Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization, 2016. 3

    work page 2016

  19. [19]

    Sources of uncertainty in 3d scene reconstruction,

    Marcus Klasson, Riccardo Mereu, Juho Kannala, and Arno Solin. Sources of uncertainty in 3d scene reconstruction,

    work page

  20. [20]

    Dense-sfm: Structure from motion with dense consistent matching, 2025

    JongMin Lee and Sungjoo Yoo. Dense-sfm: Structure from motion with dense consistent matching, 2025. 2

    work page 2025

  21. [21]

    Bayesian nerf: Quantifying uncertainty with volume density for neural implicit fields.IEEE Robotics and Au- tomation Letters, 10(3):2144–2151, 2025

    Sibaek Lee, Kyeongsu Kang, Seongbo Ha, and Hyeonwoo Yu. Bayesian nerf: Quantifying uncertainty with volume density for neural implicit fields.IEEE Robotics and Au- tomation Letters, 10(3):2144–2151, 2025. 2

    work page 2025

  22. [22]

    Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin

    Zhaoshuo Li, Thomas M ¨uller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neu- ralangelo: High-fidelity neural surface reconstruction, 2023. 2, 7, 8

    work page 2023

  23. [23]

    Noksr: Kernel-free neural surface reconstruction via point cloud serialization, 2025

    Zhen Li, Weiwei Sun, Shrisudhan Govindarajan, Shaobo Xia, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasac- chi. Noksr: Kernel-free neural surface reconstruction via point cloud serialization, 2025. 2

    work page 2025

  24. [24]

    Barf: Bundle-adjusting neural radiance fields,

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Si- mon Lucey. Barf: Bundle-adjusting neural radiance fields,

    work page

  25. [25]

    Vela, and Stan Birchfield

    Yunzhi Lin, Thomas M ¨uller, Jonathan Tremblay, Bowen Wen, Stephen Tyree, Alex Evans, Patricio A. Vela, and Stan Birchfield. Parallel inversion of neural radiance fields for robust pose estimation, 2023. 2

    work page 2023

  26. [26]

    Shadowneus: Neural sdf reconstruction by shadow ray supervision, 2023

    Jingwang Ling, Zhibo Wang, and Feng Xu. Shadowneus: Neural sdf reconstruction by shadow ray supervision, 2023. 2

    work page 2023

  27. [27]

    Neuraludf: Learning unsigned distance fields for multi-view reconstruction of surfaces with arbitrary topolo- gies, 2022

    Xiaoxiao Long, Cheng Lin, Lingjie Liu, Yuan Liu, Peng Wang, Christian Theobalt, Taku Komura, and Wenping Wang. Neuraludf: Learning unsigned distance fields for multi-view reconstruction of surfaces with arbitrary topolo- gies, 2022. 2

    work page 2022

  28. [28]

    Ganesh: Generalizable nerf for lensless imaging

    Rakesh Raj Madhavan, Akshat Kaimal, Badhrinarayanan K.V , Vinayak Gupta, Rohit Choudhary, Chandrakala Shan- muganathan, and Kaushik Mitra. Ganesh: Generalizable nerf for lensless imaging. InProceedings of the Winter Con- ference on Applications of Computer Vision (WACV), pages 9481–9490, 2025. 2

    work page 2025

  29. [29]

    Verf: Runtime monitoring of pose estimation with neural radiance fields, 2023

    Dominic Maggio, Courtney Mario, and Luca Carlone. Verf: Runtime monitoring of pose estimation with neural radiance fields, 2023. 3

    work page 2023

  30. [30]

    Neat: Learning neural implicit surfaces with arbitrary topologies from multi- view images, 2023

    Xiaoxu Meng, Weikai Chen, and Bo Yang. Neat: Learning neural implicit surfaces with arbitrary topologies from multi- view images, 2023. 2 9

    work page 2023

  31. [31]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthe- sis.Commun. ACM, 65(1):99–106, 2021. 1

    work page 2021

  32. [32]

    Differentiable volumetric rendering: Learn- ing implicit 3d representations without 3d supervision, 2020

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learn- ing implicit 3d representations without 3d supervision, 2020. 1

    work page 2020

  33. [33]

    Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction, 2021

    Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction, 2021. 2

    work page 2021

  34. [34]

    Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild, 2024

    Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild, 2024. 2

    work page 2024

  35. [35]

    Permutosdf: Fast multi-view reconstruction with implicit surfaces using per- mutohedral lattices, 2023

    Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using per- mutohedral lattices, 2023. 2

    work page 2023

  36. [36]

    Savani, M

    Y . Savani, M. Finzi, and J. Z. Kolter. Diffusing Differentiable Representations.https : / / doi . org / 10 . 48550 / arxiv.2412.06981, 2024. arXiv:2412.06981. 1

    work page arXiv 2024

  37. [37]

    Sch ¨onberger and Jan-Michael Frahm

    Johannes L. Sch ¨onberger and Jan-Michael Frahm. Structure- from-motion revisited. In2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 4104– 4113, 2016. 2

    work page 2016

  38. [38]

    Stochastic neural radiance fields: Quanti- fying uncertainty in implicit 3d representations

    Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quanti- fying uncertainty in implicit 3d representations. In2021 In- ternational Conference on 3D Vision (3DV), pages 972–981,

    work page

  39. [39]

    Truong, M

    P. Truong, M. J. Rakotosaona, F. Manhardt, and F. Tombari. Sparf: Neural Radiance Fields from Sparse and Noisy Poses. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4190–4200,

    work page

  40. [40]

    S. Umeyama. Least-squares estimation of transformation pa- rameters between two point patterns.IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(4):376–380,

    work page

  41. [41]

    Uncertainty-aware camera pose estimation from points and lines, 2021

    Alexander Vakhitov, Luis Ferraz Colomina, Antonio Agudo, and Francesc Moreno-Noguer. Uncertainty-aware camera pose estimation from points and lines, 2021. 3

    work page 2021

  42. [42]

    Visual geometry grounded deep structure from motion, 2023

    Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Visual geometry grounded deep structure from motion, 2023. 2

    work page 2023

  43. [43]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,

    work page

  44. [44]

    Hf-neus: Improved surface reconstruction using high-frequency de- tails, 2022

    Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency de- tails, 2022. 2

    work page 2022

  45. [45]

    Pet-neus: Positional encoding tri-planes for neural surfaces, 2023

    Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Pet-neus: Positional encoding tri-planes for neural surfaces, 2023. 2

    work page 2023

  46. [47]

    arXiv:2102.07064. 2, 3

    work page arXiv

  47. [48]

    V oxurf: V oxel-based efficient and accurate neural surface reconstruction, 2023

    Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. V oxurf: V oxel-based efficient and accurate neural surface reconstruction, 2023. 2

    work page 2023

  48. [49]

    Q. Yan, Q. Wang, K. Zhao, J. Chen, B. Li, X. Chu, and F. Deng. CF-NeRF: Camera Parameter Free Neural Radi- ance Fields with Incremental Learning.https://arxiv. org/abs/2312.08760, 2023. arXiv:2312.08760. 2

    work page arXiv 2023

  49. [50]

    Blendedmvs: A large-scale dataset for generalized multi-view stereo net- works, 2020

    Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo net- works, 2020. 6

    work page 2020

  50. [51]

    Multiview neu- ral surface reconstruction by disentangling geometry and ap- pearance.Advances in Neural Information Processing Sys- tems, 33, 2020

    Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neu- ral surface reconstruction by disentangling geometry and ap- pearance.Advances in Neural Information Processing Sys- tems, 33, 2020. 1, 2

    work page 2020

  51. [52]

    V ol- ume rendering of neural implicit surfaces, 2021

    Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. V ol- ume rendering of neural implicit surfaces, 2021. 2 10 PCM-NeRF: Probabilistic Camera Modeling for Neural Radiance Fields under Pose Uncertainty Supplementary Material Supplementary Material 5.1. Probabilistic Pose Representation: Mean-Pose Approximation Each camera poseP i is modelled as a distri...

    work page 2021

  52. [53]

    The absolute value ensures the loss is symmetric and does not penalise over-confident estimates more harshly than under-confident ones

    to high target uncertainty. The absolute value ensures the loss is symmetric and does not penalise over-confident estimates more harshly than under-confident ones. Gradient flow.The confidence scoreγ i is derived from external geometric evidence (SfM match density and ren- dering PSNR) and isnotback-propagated through the neu- ral rendering. The variance ...

    work page

  53. [54]

    immediately afterloss.backward()and before optimizer pose.step(), so that the Adam update for cameraisees a gradient reduced by(1 + ¯σ iκ)−1. Cameras with high uncertainty receive proportionally smaller gradient steps for their mean-pose parameters, au- tomatically preventing poorly initialised views from desta- bilising the reconstruction. Cameras with l...

    work page