Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

BayesSDF: Surface-Based Laplacian Uncertainty Estimation for 3D Geometry with Neural Signed Distance Fields

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read BayesSDF claims that local surface uncertainty in a neural SDF reconstruction can be read from a Hessian-style tensor of color gradients with respect to a learned deformation field, and reports that this tensor ranks depth errors as well…

desk verdict A cheap sensitivity heuristic that is over-sold as a Laplace approximation; the empirical evidence does not back the headline claim. read the letter →

arxiv 2507.06269 v3 pith:WPFCU5XG submitted 2025-07-08 cs.CV cs.AI

Rushil Desai This is my paper · ORCID
classification cs.CVcs.AI
keywords uncertaintyquantificationneuralsigneddistancefieldsLaplaceapproximationdeformationfieldhashencodingsparsificationcurvedeptherrorcorrelation3Dreconstruction
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

BayesSDF aims to give neural signed distance function reconstructions a cheap, surface-aligned uncertainty estimate. The paper's mechanism is to treat the sensitivity of rendered RGB color to small perturbations of a learned hash-encoded deformation field as a proxy for local geometric instability, accumulating squared color gradients into a 3D tensor and rendering that tensor as a per-pixel uncertainty map. The central empirical claim is that this map ranks pixels by depth error: removing high-uncertainty pixels lowers mean absolute error in a sparsification curve, with scores comparable to an ensemble baseline. If true, the method would provide interpretable geometry-aware uncertainty for SDF-based 3D perception without training multiple models.

What carries the argument

The central object is the aggregated squared color-gradient tensor $H$, stored on the hash-encoded deformation grid used to shape the SDF. For each voxel corner, the paper accumulates per-channel derivatives of rendered color with respect to the deformation-vector components $d_1,d_2,d_3$, squares them, and combines them into a scalar that is meant to represent local instability in color space. The uncertainty function $\sigma(x)=\sqrt{H(\text{grid index}(x))}$ maps that tensor back to every rendered pixel, and the paper treats this map as the Hessian-based Laplace uncertainty of the surface. The whole pipeline is designed so that uncertainty can be read out in one backward pass per color channel, with no stochastic perturbations or multiple model copies.

What would settle it

Render a flat, well-observed plane painted with a high-contrast checkerboard pattern. BayesSDF's color-gradient tensor will be large at every texture edge and predict high uncertainty there; if the measured depth error in those pixels is near zero, the claimed correlation between uncertainty and geometric error fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that local geometric instability in a neural SDF reconstruction is encoded in the sensitivity of rendered color to small perturbations of a deformation field. Concretely, BayesSDF backpropagates each RGB channel through the renderer, squares the partial derivatives with respect to the three components of the deformation vector at each of the eight hash-grid corners, and accumulates them into a scalar field $H$; the per-pixel uncertainty is $\sigma(x)=\sqrt{H(\text{grid index}(x))}$. The paper frames this quantity as a Laplace approximation over SDF weights and shows, through sparsification curves and Area Under the Sparsification Error, that it correlates with mean absolute depth error on the Basket, Africa, Statue, and Torch scenes. The result, if it holds, is a single-pass uncertainty map that is geometrically aligned rather than ray-based, at much lower cost than ensembles or stochastic sampling.

Load-bearing premise

The load-bearing premise is that how much rendered color changes when a separately learned deformation field is nudged tells you how uncertain the reconstructed surface is, even though the paper never derives a mathematical link between that color sensitivity and the statistical uncertainty of the SDF geometry.

Editorial extensions

If this is right

  • SDF-based reconstruction systems gain a per-pixel and per-voxel uncertainty map without an ensemble, making active view selection and next-best-view planning possible.
  • Uncertainty is aligned with surface geometry rather than ray variance, so downstream collision checks or physics simulations can trust low-uncertainty surface regions preferentially.
  • The hash-encoded deformation grid keeps the uncertainty footprint small, so the method scales to higher-resolution scenes than ray-space variance approaches.
  • On synthetic and real benchmarks, the map ranks pixels by depth error comparably to ensemble variance, suggesting it could replace ensembles where compute is limited.

Reading between the lines

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

  • The same accumulated tensor could be reused during training, for example by up-weighting the reconstruction loss in low-uncertainty regions, turning the estimate from a post-hoc diagnostic into a regularizer.
  • Extending the deformation field with a time axis would yield per-frame uncertainty maps for dynamic scenes, linking the estimator to tracking or SLAM failures.
  • Because the tensor is a scalar per voxel, it can be projected directly onto extracted meshes, giving uncertainty-colored surfaces for surgical or simulation planning.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes BayesSDF, a method for estimating per-pixel uncertainty in NeuS-style neural signed distance field reconstructions. The proposal is to train a single-level hash-encoded deformation field d(x)=f_hash(x), backpropagate rendered RGB color to the deformation parameters at voxel corners, accumulate squared first derivatives of color with respect to those parameters into a 3D grid, and use the square root of the accumulated grid as a per-pixel uncertainty map. The method is evaluated with sparsification curves and AUSE against an ensemble baseline on four scenes. The abstract and conclusion claim that this constitutes a Laplace approximation over SDF weights yielding Hessian-based metrics that correlate strongly with reconstruction error.

Significance. If the core claim were established, a cheap, single-model uncertainty map for SDF reconstructions would be valuable for scientific and robotic applications. The paper deserves credit for evaluating uncertainty with an external sparsification benchmark against ground-truth depth error rather than by construction, and for addressing a real gap in neural implicit surfaces. However, the manuscript does not deliver the advertised Bayesian or Hessian object, and the quantitative evidence in Table 2 contradicts the abstract's strength claim. The significance of the paper therefore depends on an unsupported identification of color-gradient sensitivity with geometric posterior covariance.

major comments (3)
  1. [Section 4.3, displayed equation] The quantity called 'hessian' is not a Hessian, and no Laplace approximation is constructed anywhere in the paper. A Laplace approximation requires a posterior over the SDF weights and the Hessian of its negative log-likelihood; the paper defines no posterior, no likelihood, and no second derivatives. The displayed equation in Section 4.3 accumulates products of squared first derivatives of color with respect to deformation parameters, which is neither a Hessian matrix nor a curvature of any log-posterior. This leaves the central methodological claim in the abstract and in Section 1 unsupported. In addition, the prose says the channel terms are squared and summed, but the displayed formula multiplies the three channel terms; this internal inconsistency must be resolved.
  2. [Sections 4.1 and 4.5] The proxy assumption that color sensitivity to a separately learned deformation field equals geometric surface uncertainty is stated without justification. Color gradients can be large on textured or specular surfaces even when the underlying SDF geometry is tightly constrained, and small on textureless regions even when geometry is poorly constrained. The paper provides no derivation connecting partial derivatives of color with respect to deformation parameters to the posterior covariance of the SDF surface, and no experiment that would separate geometric instability from appearance sensitivity (for example, constant-albedo or textureless scenes, or perturbations of the SDF weights themselves). Without this bridge, the uncertainty map has no demonstrated geometric meaning.
  3. [Table 2] The empirical results do not support the abstract's claim of 'strongly correlate' or the conclusion's claim of 'very high calibration.' On three of four scenes BayesSDF's AUSE is worse than the Ensemble baseline (Basket 0.218 versus 0.206, Africa 0.268 versus 0.201, Torch 0.182 versus 0.164), and only on Statue is it marginally better (0.148 versus 0.150). No error bars, repeated runs, or significance tests are reported, and no metric for correlation strength (for example, rank correlation between predicted uncertainty and per-pixel depth error) is given. As presented, the table is evidence against the central claim rather than for it.
minor comments (4)
  1. [Section 1 and References] The Introduction contains an unresolved citation 'CamP [?]', and reference [60] is explicitly incomplete with the note that 'detailed author list and page numbers can be added when available'; these should be completed before resubmission.
  2. [Section 5.3] The text contains the typo 'BBy relating' instead of 'By relating', and the phrase 'Figure ∆MAE vs. Pixel Sparsification' refers to a figure without a number; the sparsification plots in Figures 3-6 should be referenced consistently.
  3. [Section 4.2] The sentence 'dc are are the corner deformation vectors' contains a duplicated word, and the relationship between the hash table size 2^{3ℓ+1} and the grid resolution (2^ℓ+1)^3 is not explained, making the deformation field specification difficult to reproduce.
  4. [Section 5.2] The captions for Figures 5 and 6 describe 'Comparison of the Ensemble method and BayesSDF in terms of AUSE', but the captions do not define the red line or the axes for the reader; the text mentions a red line only in the body.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the uncertainty map H is an operational sensitivity heuristic validated against external ground-truth depth error; the unsupported 'Laplace/Hessian' label is a correctness gap, not a circular reduction.

full rationale

The derivation chain is self-contained as a heuristic. In §4.1–§4.5, H is accumulated from squared partial derivatives of rendered color with respect to the learned deformation field, and the uncertainty function is defined as σ(x)=√H(grid index(x)). No parameter is fitted to the ground-truth depth errors used in evaluation: the sparsification protocol of §5.2 sorts pixels by H and measures ∆MAE against ground-truth depth, an external benchmark, so the correlation claim is not forced by construction. The paper does assert a 'Laplace approximation over SDF weights' (abstract and §2.3/§4.1), but §4.3 never constructs a posterior over SDF weights and Δ(x) is a product of squared first derivatives, not a second-derivative Hessian; this is an omitted derivation/mislabeling, and the supporting citations ([9],[35]) concern Laplace methods generally rather than this deformation-field sensitivity. That is a substantive correctness/support problem, but it is not circular: the empirical ranking test remains independent of the definition. There are no load-bearing self-citations (the sole author cites only external work), and the manuscript's self-flagged incompleteness ('CamP [?]' in §1, '[7?]' in §7, reference [60]'s author-note, and the 'BBy' typo in §5.3) does not create a circular equivalence. Score 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central uncertainty quantity rests on five assumptions that are not derived in the paper: (1) NeuS volume rendering is an accurate surface model; (2) the hash-encoded deformation field is a meaningful learned component whose training is left unspecified; (3) color-channel gradient magnitude with respect to deformation parameters is a faithful proxy for geometric instability; (4) squared first derivatives can stand in for a Laplace or Hessian curvature; (5) benchmark ground-truth depth is reliable for ranking. The first is standard prior work, the last is standard evaluation, but the middle three are ad hoc to this paper and are the load-bearing bridge between the computed quantity and the claimed uncertainty.

free parameters (5)
  • Hash grid level ℓ = not reported (min res = max res = 2^ℓ)
    Sets the spatial resolution of the deformation field and the hash table size 2^(3ℓ+1); no ablation or chosen value is given, yet H and uncertainty maps depend on it.
  • Hash table size = 2^(3ℓ+1), with ℓ unspecified
    Controls memory and resolution of the deformation field; cited from instant-ngp but not adapted or reported for the experiments.
  • AABB normalization bounds = not reported
    normalize(x, aabb) in Section 4.2 determines grid indices and interpolation coefficients, but the scene bounds are not stated.
  • Uncertainty normalization and thresholding = not specified
    Section 4.5 says σ may be normalized or thresholded; the evaluation does not state whether this postprocessing was applied.
  • Deformation field corner vectors d_c = learned per scene, values undisclosed
    The uncertainty is defined as derivatives with respect to these fitted vectors; without training details the measure is not reproducible.
assumptions (5)
  • domain assumption NeuS SDF-to-opacity volume rendering accurately reconstructs the surface.
    Section 3.1 adopts NeuS as the base representation without independent verification; this is prior work, not proven in the paper.
  • ad hoc to paper The learned hash-encoded deformation field captures meaningful local geometric variation.
    Section 4.2 introduces f_hash but gives no training objective, regularization, or loss term, so the meaning of the fitted vectors is unspecified.
  • ad hoc to paper Sensitivity of rendered color to deformation parameters is a faithful proxy for geometric surface uncertainty.
    Section 4.1 asserts this interpretation; no derivation connects color gradients to SDF geometry or posterior covariance.
  • ad hoc to paper Squared first derivatives of color with respect to deformation parameters approximate a Hessian or Laplace curvature.
    Sections 4.3 through 4.5 label the accumulated gradient product as a Hessian and as Laplace uncertainty without computing second derivatives of a log-posterior.
  • domain assumption Ground-truth depth in the benchmark datasets is reliable for sparsification evaluation.
    Section 5.2 uses depth error as ground truth; the datasets are accepted as standard benchmarks.
invented entities (1)
  • Hash-encoded deformation field d(x) = f_hash(x)
    purpose: A learned per-scene displacement field whose color-gradient sensitivity is interpreted as geometric uncertainty.
    Introduced in Section 4.2 with no falsifiable prediction outside the paper; it is a modeling component, not an independently observed quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BayesSDF: Surface-Based Laplacian Uncertainty Estimation for 3D Geometry with Neural Signed Distance Fields." pith.science (2026). https://pith.science/paper/WPFCU5XG

@misc{pith2026250706269,
  author       = {Pith},
  title        = {Pith review of: BayesSDF: Surface-Based Laplacian Uncertainty Estimation for 3D Geometry with Neural Signed Distance Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPFCU5XG}},
  note         = {Machine review of arXiv:2507.06269}
}
read the original abstract

Accurate surface estimation is critical for downstream tasks in scientific simulation, and quantifying uncertainty in implicit neural 3D representations still remains a substantial challenge due to computational inefficiencies, scalability issues, and geometric inconsistencies. However, current neural implicit surface models do not offer a principled way to quantify uncertainty, limiting their reliability in real-world applications. Inspired by recent probabilistic rendering approaches, we introduce BayesSDF, a novel probabilistic framework for uncertainty estimation in neural implicit 3D representations. Unlike radiance-based models such as Neural Radiance Fields (NeRF) or 3D Gaussian Splatting, Signed Distance Functions (SDFs) provide continuous, differentiable surface representations, making them especially well-suited for uncertainty-aware modeling. BayesSDF applies a Laplace approximation over SDF weights and derives Hessian-based metrics to estimate local geometric instability. We empirically demonstrate that these uncertainty estimates correlate strongly with surface reconstruction error across both synthetic and real-world benchmarks. By enabling surface-aware uncertainty quantification, BayesSDF lays the groundwork for more robust, interpretable, and actionable 3D perception systems.

Figures

Figures reproduced from arXiv: 2507.06269 by the authors.

Figure 1
Figure 1. Africa Dataset: BayesSDF-derived view with uncer [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Basket Dataset: BayesSDF-generated view with uncer [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Basket Dataset: Spar￾sification curve showing the change in mean absolute error (∆MAE) with increasing levels of sparsification [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Africa Dataset: Comparison of the Ensemble method and BayesSDF in terms of AUSE [PITH_FULL_IMAGE:figures/full_fig_p006_5.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

  1. CASA-SDF: Curriculum-Aware Spatial Adaptation with Curvature-Guided Density for Neural Implicit Surface Reconstruction

    cs.CV 2026-07 conditional novelty 6.0 of 10

    By combining uncertainty-annealed prior supervision with curvature-adaptive SDF sharpness, CASA-SDF improves completeness and recall in indoor neural surface reconstruction.

Reference graph

Works this paper leans on

61 extracted references · 58 canonical work pages · cited by 1 Pith paper

  1. [60]

    Mvsnerf++: Fast generalizable radiance field reconstruction from multi-view stereo

    Yingqi Zhang et al. Mvsnerf++: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) , 2023. Accepted paper, detailed author list and page numbers can be added when available. 3, 7

    work page 2023

  2. [1]

    A review of uncertainty quantification in deep learning: Techniques, applica- tions and challenges

    Moloud Abdar, Farhad Pourpanah, Sadiq Hus- sain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul Fieguth, Abbas Khosravi, U Ra- jendra Acharya, Vladimir Makarenkov, Saeid Naha- vandi, and Xiaochun Cao. A review of uncertainty quantification in deep learning: Techniques, applica- tions and challenges. Information Fusion , 76:243– 297, 2021. 1, 2, 4, 6, 7

    work page 2021

  3. [2]

    Estimating and exploiting the aleatoric uncertainty in surface normal estimation

    Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 1852–1861, 2021. 2, 4, 7

    work page 2021

  4. [3]

    Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P

    Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 1, 2

    work page 2021

  5. [4]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5473– 5482, 2022. 1, 2, 7

    work page 2022

  6. [5]

    Nerd: Neural reflectance decomposition from image collec- tions

    Martin Boss, Jonas Braun, Xiaoxiao Liu, Michael Zollhoefer, Matthias Nießner, and Angela Dai. Nerd: Neural reflectance decomposition from image collec- tions. In ICCV, 2021. 3

    work page 2021

  7. [6]

    T- nerf: Time-conditioned neural radiance fields for ef- ficient representation of dynamic scenes

    Xinhang Cao, Linjie Yang, Jingyi Yu, and Hao Su. T- nerf: Time-conditioned neural radiance fields for ef- ficient representation of dynamic scenes. In Proceed- ings of the European Conference on Computer Vision (ECCV), pages 507–523, 2022. 3

    work page 2022

  8. [7]

    Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2, 8

    work page 2021

Show all 61 references
  1. [8]

    Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures

    Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceed- ings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recogniti...

    2023

  2. [9]

    Laplace redux – effortless bayesian deep learning

    Erik Daxberger, Agustinus Kristiadi, Alexander Im- mer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace redux – effortless bayesian deep learning. In Advances in Neural Information Process- ing Systems (NeurIPS), pages 3184–3196, 2021. 1, 3, 4, 7

    2021

  3. [10]

    Bayes’ rays: Uncer- tainty quantification for neural radiance fields

    Lily Goli, Cody Reading, Silvia Sell ´an, Alec Jacob- son, and Andrea Tagliasacchi. Bayes’ rays: Uncer- tainty quantification for neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2024. 1, 2, 3, 5, 7

    2024

  4. [11]

    Hoffman, Tuan Anh Le, Pavel Sountsov, Christopher Suter, Ben Lee, Vikash K

    Matthew D. Hoffman, Tuan Anh Le, Pavel Sountsov, Christopher Suter, Ben Lee, Vikash K. Mansinghka, and Rif A. Saurous. Probnerf: Uncertainty-aware in- ference of 3d shapes from 2d images. In Proceed- ings of the 26th International Conference on Artificial Intelligence and Stat...

    2023

  5. [12]

    Gnt: Neural radiance fields with geometry-consistent transformer

    Yujing Jiang, Yinda Lin, Linjie Yang, Jingwan Yu, and Gang Yu. Gnt: Neural radiance fields with geometry-consistent transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 14976–14986,

  6. [13]

    Geonerf: Generalizing nerf with ge- ometry priors

    Mohammad Mahdi Johari, Yann Lepoittevin, and Franc ¸ois Fleuret. Geonerf: Generalizing nerf with ge- ometry priors. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 18365–18375, 2022. 2, 7

    2022

  7. [14]

    What uncertainties do we need in bayesian deep learning for computer vi- sion? In Advances in Neural Information Processing Systems (NeurIPS), 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vi- sion? In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 4, 6, 7

    2017

  8. [15]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) , 42(4):Article 1,

  9. [16]

    Sdfstudio methods

    Autonomous Vision Lab. Sdfstudio methods. https : / / github . com / autonomousvision / sdfstudio / blob / master / docs / sdfstudio - methods . md ,

  10. [17]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 4, 5, 6, 7

    2017

  11. [18]

    Bayesian nerf: Quantifying uncer- tainty with volume density for neural implicit fields

    Sibaek Lee, Kyeongsu Kang, Seongbo Ha, and Hyeonwoo Yu. Bayesian nerf: Quantifying uncer- tainty with volume density for neural implicit fields. In IEEE Robotics and Automation Letters (RA-L) , 2024. arXiv:2404.06727. 3, 5, 6, 7

    2024 arXiv

  12. [19]

    Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radi- ance fields for unconstrained photo collections. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), p...

    2021

  13. [20]

    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 synthesis. In European Conference on Computer Vision (ECCV), pages 405–421, 2020. 1, 2, 3, 7

    2020

  14. [21]

    de Cas- tro, Nick Pawlowski, Bernardo Marques, Konstanti- nos Kamnitsas, Mark van der Wilk, and Ben Glocker

    Miguel Monteiro, Lo ¨ıc le Folgoc, Daniel C. de Cas- tro, Nick Pawlowski, Bernardo Marques, Konstanti- nos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spa- tially correlated aleatoric uncertainty. In Advances in Neural Information...

    2020

  15. [22]

    Morris, Ken ichi Kanatani, and Takeo Kanade

    Daniel D. Morris, Ken ichi Kanatani, and Takeo Kanade. Uncertainty modeling for optimal structure from motion. In Vision Algorithms: Theory and Prac- tice, pages 200–217. Springer, 1999. 7

    1999

  16. [23]

    Instant neural graphics primitives with a multiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (SIGGRAPH) , 2022. arXiv:2201.05989. 3, 4, 7

    2022 arXiv

  17. [24]

    Radford M. Neal. Bayesian Learning for Neural Net- works. Springer, 1995. 3, 5, 6, 7

    1995

  18. [25]

    Newcombe, Dieter Fox, and Steven M

    Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343–352, 2015. 3, 7, 8

    2015

  19. [26]

    Differentiable volumetric render- ing: Learning implicit 3d representations without 3d supervision

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric render- ing: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 3504–...

    2020

  20. [27]

    Barron, Ben Milden- hall, Mehdi S

    Michael Niemeyer, Jonathan T. Barron, Ben Milden- hall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 5...

    2022

  21. [28]

    Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruc- tion

    Michael Oechsle, Songyou Peng, Michael Niemeyer, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruc- tion. In ICCV, 2021. 2, 3, 7

    2021

  22. [29]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,

  23. [30]

    Barron, Sofien Bouaziz, Dan B

    Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5864– 5874, 2021. 1, 3, 4, 8

    2021

  24. [31]

    Barron, Sofien Bouaziz, Dan B

    Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypern- erf: A higher-dimensional representation for topolog- ically varying neural radiance fields. In IEEE/CVF Conference on Computer ...

    2022

  25. [32]

    D-nerf: Neural radi- ance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radi- ance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 10318–10327, 2020. 3, 4, 8

    2020

  26. [33]

    Srini- vasan, Jonathan T

    Konstantinos Rematas, Andrew Liu, Pratul P. Srini- vasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban ra- diance fields. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 12932–12942...

    2022

  27. [34]

    Nerf-tex: Neural reflectance field texture maps for high-fidelity facial avatars

    Gernot Riegler, Ayush Tewari, Justus Thies, Michael Zollh¨ofer, and Christian Theobalt. Nerf-tex: Neural reflectance field texture maps for high-fidelity facial avatars. ACM Transactions on Graphics (TOG) , 42 (4):1–13, 2023. 3

    2023

  28. [35]

    A scalable laplace approximation for neural networks

    Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable laplace approximation for neural networks. In International Conference on Learning Representa- tions (ICLR), 2018. 3, 4, 7

    2018

  29. [36]

    Non-linear sphere tracing for rendering deformed signed distance fields

    Dario Seyb, Alec Jacobson, Derek Nowrouzezahrai, and Wojciech Jarosz. Non-linear sphere tracing for rendering deformed signed distance fields. In Pro- ceedings of the 19th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, 2019. 2, 3, 7

    2019

  30. [37]

    Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d repre- sentations

    Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d repre- sentations. In 3DV, 2021. arXiv:2109.02123. 2, 3

    2021 arXiv

  31. [38]

    Conditional-flow nerf: Ac- curate 3d modelling with reliable uncertainty quantifi- cation

    Jianxiong Shen, Antonio Agudo, Francesc Moreno- Noguer, and Adria Ruiz. Conditional-flow nerf: Ac- curate 3d modelling with reliable uncertainty quantifi- cation. In ECCV, 2022. arXiv:2203.10192. 2

    2022 arXiv

  32. [39]

    Ralph C. Smith. Uncertainty Quantification: Theory, Implementation, and Applications. SIAM, 2013. 2, 3, 7, 8

    2013

  33. [40]

    Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T

    Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP...

    2021

  34. [41]

    imap: Implicit mapping and positioning in real-time

    Edgar Sucar, Shikun Sun, and Julien Valentin. imap: Implicit mapping and positioning in real-time. In ICCV, 2021. 2, 7

    2021

  35. [42]

    Warpcurves: A tool for explicit manipula- tion of implicit surfaces

    Masamichi Sugihara, Brian Wyvill, and Ryan Schmidt. Warpcurves: A tool for explicit manipula- tion of implicit surfaces. Computer Aided Geometric Design, 27(3):163–176, 2010. 7

    2010

  36. [43]

    V olume rendering digest (for nerf)

    Andrea Tagliasacchi and Ben Mildenhall. V olume rendering digest (for nerf). https://github. com / nerfstudio / nerfstudio / blob / main/docs/volumerendering_digest.md ,

  37. [44]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul P Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Sing- hal, Ravi Ramamoorthi, Jonathan T Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS,

  38. [45]

    Nerfstudio: A modular framework for neural radi- ance field development

    Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexan- der Kristoffersen, Jake Austin, Kamyar Salahi, Ab- hik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radi- ance field development. https...

    2023

  39. [46]

    Non-rigid neural radiance fields: Re- construction and novel view synthesis of a dynamic scene from monocular video

    Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollh ¨ofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Re- construction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International C...

    2021

  40. [47]

    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. In Advances in Neural Information Processing Systems (NeurIPS) , 2021. 1, 2, 3, 4, 7

    2021

  41. [48]

    Nerfbusters: Removing ghostly ar- tifacts from casually captured nerfs

    Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly ar- tifacts from casually captured nerfs. https : //github.com/nerfstudio/nerfstudio/ blob/main/examples/nerfbusters , 2023. 3, 7

    2023

  42. [49]

    Nerfingmvs: Guided optimiza- tion of neural radiance fields for indoor multi-view stereo

    Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Ji- wen Lu, and Jie Zhou. Nerfingmvs: Guided optimiza- tion of neural radiance fields for indoor multi-view stereo. In arXiv preprint, 2023. 7

    2023

  43. [50]

    Visualizing spec- tral bundle adjustment uncertainty

    Kyle Wilson and Scott Wehrwein. Visualizing spec- tral bundle adjustment uncertainty. Technical report, Washington College and Western Washington Univer- sity, 2023. 1, 7

    2023

  44. [51]

    Bayesian diffu- sion models for 3d shape reconstruction

    Haiyang Xu, Yu Lei, Zeyuan Chen, Xiang Zhang, Yue Zhao, Yilin Wang, and Zhuowen Tu. Bayesian diffu- sion models for 3d shape reconstruction. InIEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 10628–10638, 2024. 7, 8

    2024

  45. [52]

    Learning neu- ral implicit representations from photometric images

    Zexiang Xu, Lingjie Liu, Stephen Lombardi, Shub- ham Tulsiani, Ricardo Martin-Brualla, Tae-Hyun Kim, Keunhong Park, and Yael Pritch. Learning neu- ral implicit representations from photometric images. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern R...

    2022

  46. [53]

    Active neural mapping

    Zike Yan, Haoxiang Yang, and Hongbin Zha. Active neural mapping. In arXiv preprint, 2022. 3, 7

    2022

  47. [54]

    V ol- ume rendering of neural implicit surfaces

    Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Yaron Hasson, Ron Kimmel, and Yaron Lipman. V ol- ume rendering of neural implicit surfaces. InNeurIPS,

  48. [55]

    Plenoctrees for real-time rendering of neural radiance fields

    Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5752–5761, 2021. 3, 8

    2021

  49. [56]

    Mononerd: Nerf-like representations for monocular 3d object detection

    Zhixin Yu, Yichao Xu, Yifan Zhou, Tong Liu, Linjie Yang, Tao Kong, and Cewu Lu. Mononerd: Nerf-like representations for monocular 3d object detection. In NeurIPS, 2022. 4, 7

    2022

  50. [57]

    Neural ensemble search for uncertainty estimation and dataset shift

    Sheheryar Zaidi, Arber Zela, Thomas Elsken, Chris Holmes, Frank Hutter, and Yee Whye Teh. Neural ensemble search for uncertainty estimation and dataset shift. In Advances in Neural Information Processing Systems (NeurIPS), pages 2809–2820, 2022. 4, 6, 7

    2022

  51. [58]

    Activermap: Radiance field for active mapping and planning

    Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. Activermap: Radiance field for active mapping and planning. In arXiv preprint, 2022. 3, 7

    2022

  52. [59]

    Nerf++: Analyzing and improving neural radiance fields

    Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. In arXiv preprint, 2020. 2, 7

    2020

  53. [61]

    Oswald, and Marc Pollefeys

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 12776–...

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.