Pith. sign in

REVIEW 4 major objections 3 minor 41 references

WarpRF: Multi-View Consistency for Training-Free Uncertainty Quantification and Applications in Radiance Fields

T0 review · 4 major / 3 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 WarpRF claims that warping rendered training views into an unseen viewpoint and measuring the L1 consistency of depths and colors is enough to quantify where a radiance field is uncertain, with no training and no access to the model's…

desk verdict A simple, useful training-free uncertainty wrapper for radiance fields whose empirical claims mostly hold, but the abstract overreaches and the source-view fidelity assumption needs a diagnostic. read the letter →

arxiv 2506.22433 v1 pith:B2DCMOC4 submitted 2025-06-27 cs.CV

Sadra Safadoust , Fabio Tosi , Fatma Güney , Matteo Poggi This is my paper · ORCID
classification cs.CV
keywords uncertaintyquantificationradiancefieldsmulti-viewconsistencybackwardwarpingactiveviewselectionmapping3DGaussianSplattingNeural
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

The paper sets out to show that multi-view consistency — a cue long used in stereo matching and monocular depth self-supervision — is enough to quantify the uncertainty of any trained radiance field. WarpRF renders depth and optionally color at a candidate unseen viewpoint, warps the renderings of the training views into that viewpoint using known poses, and measures the per-pixel absolute difference as the uncertainty. Because it needs only renderable depth, images, and poses, it applies to NeRF, 3D Gaussian Splatting, and newer frameworks like SVRaster with no retraining and no access to the model's internal structure. If correct, this gives every radiance-field model a free uncertainty signal, and the paper reports that this simple signal outperforms framework-specific methods on depth-error correlation, active view selection, and active mapping.

What carries the argument

Backward warping across viewpoints, with rendered depth acting as the geometric proxy for the scene the radiance field has learned. The warp takes each source view's rendering into the target viewpoint, and the resulting reprojection residual — averaged for depth (Eq. 8), min-reduced for color (Eq. 9) — is the uncertainty estimate; this single operation is what allows the method to be implementation-agnostic.

What would settle it

Render a target view that is known to be accurate, but corrupt a single training view (blur its image or perturb its pose) before computing WarpRF; because the corrupted source view will warp inconsistently, WarpRF will flag accurate target regions as uncertain, and the AUSE-style correlation with true depth error will degrade — a direct test of the reliability assumption.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that multi-view consistency is a sufficient uncertainty signal for radiance fields. Given a trained model and a candidate viewpoint, WarpRF renders the depth and image at that viewpoint, renders depth and images at the training source views, warps those source renderings into the target viewpoint by backward warping using the relative camera poses and the target depth, and reads disagreement as uncertainty: per-pixel uncertainty is the average absolute depth difference (Eq. 8), while image-level uncertainty is the summed per-pixel minimum color difference across sources (Eq. 9). The paper further claims this training-free, model-agnostic procedure outperforms framework-specific uncertainty methods on depth-error correlation measured by AUSE on ScanNet++ and ETH3D, and delivers better active view selection and active mapping for NeRF, 3DGS, and SVRaster.

Load-bearing premise

The load-bearing premise is that the radiance field renders accurate images and depths at the training viewpoints, so any inconsistency after warping can be attributed to the unseen target view; if training-view renderings are themselves poor, the reference signals carry the error and the uncertainty estimate is no longer specific to the target.

Editorial extensions

If this is right

  • Any radiance field that can render depth and images and has known training poses inherits per-view and per-pixel uncertainty at no training cost.
  • Uncertainty can drive active view selection: picking the candidate view with the largest warping inconsistency improves novel-view synthesis over random, farthest-view, and FisherRF-based selection in the reported settings.
  • The same signal supports active mapping, where adding the most inconsistent view improves reconstructed meshes for both NeRF-based and 3DGS-based surface reconstruction.
  • Frameworks with no existing uncertainty method, such as SVRaster, can immediately adopt WarpRF, enabling the first reported active view selection with that renderer.

Reading between the lines

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

  • WarpRF's uncertainty conflates model error with scene dynamics: moving objects, reflections, or illumination changes would produce high inconsistency even for a perfect model, so the score could double as an unsupervised change or distractor detector.
  • Because Eq. 9 takes the minimum over source views, a target region that is consistent with even one training view is deemed certain; this suits view selection but may underestimate uncertainty for regions seen from only a single training direction.
  • The same consistency residual could be turned into a training signal — regularizing a radiance field to minimize WarpRF inconsistency at novel views — turning the free uncertainty estimator into a self-supervised loss.
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

4 major / 3 minor

Summary. WarpRF proposes a training-free, model-agnostic uncertainty quantification for radiance fields. Given a trained NeRF, 3DGS, or SVRaster model, it renders depth maps and images at the training views, warps them into a novel target viewpoint using the rendered target depth, and defines uncertainty as the L1 difference between these warped reference signals and the target view's own renderings (Eqs. 8–9). The per-pixel depth-based uncertainty (Eq. 8) is evaluated with AUSE on ScanNet++ and ETH3D, while the image-level color-based uncertainty (Eq. 9) is used for active view selection and active mapping. The paper reports consistent gains over framework-specific baselines (BayesRays, FisherRF, Manifold) and over random/farthest-view selection for SVRaster.

Significance. If the reported results hold, WarpRF would be a valuable practical tool: it requires no modifications to the radiance field training, no additional learned parameters, and works with any model that can render depth and color. The core idea—using multi-view consistency as an uncertainty signal—is simple and clearly specified, and the paper demonstrates it across several downstream tasks, including first-time active view selection with SVRaster. Strengths include the explicit statement of the underlying assumption (accurate rendering at training views), the use of standard protocols where they apply, and per-scene tables in the supplement that allow readers to see variation. The main limitations are the untested nature of that core assumption, the lack of statistical error bars or significance tests in a paper that makes strong comparative claims, and a dataset-switch in the AUSE evaluation that is not quantitatively justified.

major comments (4)
  1. [Sec. 3.2, Eqs. (8)–(9)]
  2. [Sec. 4.1] The decision to drop the Light Field dataset for AUSE evaluation is justified only qualitatively ('we found its ground truth depths to be highly inaccurate') without quantitative evidence. Since prior methods such as BayesRays and FisherRF were originally evaluated on Light Field, changing the protocol without reporting numerical comparisons on the same split (or at least a quantitative characterization of the depth inaccuracy) weakens the comparability of Table 1 with previously published numbers. Please report the specific depth-error statistics for Light Field and, if possible, also provide AUSE numbers on Light Field with a filtered or reprocessed ground truth, or clearly state that the results are not comparable to earlier published values.
  3. [Tables 2–7 and Supp. Tables 10–18] Many key comparisons are very close (e.g., Table 2: PSNR 20.715 vs. 20.654 for Manifold; Table 10, bicycle PSNR 17.850 vs. 18.345 for Manifold and 18.006 for FisherRF; Table 16, hotdog PSNR 27.971 vs. 29.456 for FisherRF). The paper reports averages of three runs but no variance or significance tests. The claim of 'consistently outperforming' all baselines is stronger than the evidence supports without error bars and paired tests. At minimum, report per-run standard deviations; ideally, add a paired significance test over scenes (e.g., Wilcoxon signed-rank) for the headline tables.
  4. [Sec. 4.2 and abstract] The claim that WarpRF is 'inexpensive' and applicable 'for free' is not substantiated by any runtime or computational analysis. Equation (9) requires rendering and warping all source views for every candidate target view; in the Mip-NeRF360 experiment the candidate pool is all training views (potentially hundreds). No wall-clock time, per-candidate rendering count, or FLOPs is reported. Since active view selection is a decision-time procedure, total overhead matters. Please report the cost of computing uncertainty per candidate and compare it against the cost of the baselines.
minor comments (3)
  1. [Sec. 4.2, paragraph 2] The text says 'adding one every 100 epochs until 20 views are selected' while training is described in iterations ('30 000 iterations in total'); please clarify the schedule (e.g., one view per 100 training epochs, but only 16 additions are needed to go from 4 to 20 views).
  2. [Table 5 caption] Typo: 'Follwoing' should be 'Following'.
  3. [Sec. 3.2.1] The description of the warping process ('applying the relative camera transformation between Sn and T, followed by a backward warping process performed according to DT') is imprecise about whether the source depth is forward-projected and then resampled using the target depth, or whether the target depth is used to find corresponding source pixels. A one-sentence formalization (e.g., with the standard p_t = K(R p_s + t)/z_s and a z-buffer or bilinear resampling) would help reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

Low circularity: WarpRF's uncertainty is an external proxy evaluated against held-out depth/NVS/mesh benchmarks, and the derivation does not reduce to its inputs; two minor self-citations are not load-bearing.

full rationale

No circular step is exhibited. WarpRF defines uncertainty as the L1 residual between rendered target depth/image and source-view renderings warped into the target (Eqs. 8-9). The warped source signals are not the evaluation targets: AUSE compares against ground-truth depth, and active view selection/mapping are scored on held-out test views, PSNR/SSIM/LPIPS and mesh F1/accuracy/completion, none of which are used to construct the uncertainty. The AUSE correlation is an empirical claim about a proxy, not a tautology, since source-warped depths are not equal to ground truth and depend on the stated assumption that training-view renderings are accurate (Sec. 3.2). That assumption is a limitation (no diagnostic is provided for failures), not a circularity. Self-citations [22] and [26] provide contextual support for multi-view consistency as a confidence cue and for the rendering-geometry link, but they are joined by independent references [3,4,6,25] and do not carry the central claim; hence they do not make the derivation circular. Overall score 2 reflects only the minor non-load-bearing self-citations and the untested premise, not a reduction by construction.

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

WarpRF introduces no new physical entities and no fitted numerical constants. It rests on four domain assumptions about the reliability of source-view renderings and about reprojection consistency as a proxy for model uncertainty. The empirical claims are therefore supported by a clean, parameter-free design, but the design is only as good as those assumptions.

assumptions (5)
  • domain assumption Photometric and geometric consistency should hold among images rendered by an accurate radiance field; inconsistency indicates uncertainty.
    Central premise of Section 3.2, first paragraph: 'Built upon the assumption that photometric and geometric consistency should hold among images rendered by an accurate model.' This is assumed, not derived.
  • domain assumption Depth and images rendered at training viewpoints are accurate and reliable.
    Section 3.2, final paragraph: 'we assume that both images and depth maps are accurately rendered in correspondence of the training viewpoints.' If source renderings are wrong, the inconsistency signal conflates source error with target uncertainty.
  • domain assumption The depth map rendered by the radiance field is a valid proxy for scene geometry for the purpose of warping.
    Section 3.2 states 'the accuracy of the depth maps rendered by the model serves as a vital proxy for the quality of the underlying scene structure', and backward warping uses the target depth to project source cues into the target view.
  • domain assumption Camera poses for training and target views are known.
    Used throughout Section 3.2.1 to apply relative camera transformations between source and target views; this is standard in radiance field pipelines but is a required external input for WarpRF.
  • domain assumption Per-pixel minimum reprojection error over source views robustly handles occlusions and disocclusions.
    Equation 9 uses the minimum over source views instead of the average, citing [6], but the paper provides no analysis of when this min-based choice fails for radiance field renderings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WarpRF: Multi-View Consistency for Training-Free Uncertainty Quantification and Applications in Radiance Fields." pith.science (2026). https://pith.science/paper/B2DCMOC4

@misc{pith2026250622433,
  author       = {Pith},
  title        = {Pith review of: WarpRF: Multi-View Consistency for Training-Free Uncertainty Quantification and Applications in Radiance Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2DCMOC4}},
  note         = {Machine review of arXiv:2506.22433}
}
read the original abstract

We introduce WarpRF, a training-free general-purpose framework for quantifying the uncertainty of radiance fields. Built upon the assumption that photometric and geometric consistency should hold among images rendered by an accurate model, WarpRF quantifies its underlying uncertainty from an unseen point of view by leveraging backward warping across viewpoints, projecting reliable renderings to the unseen viewpoint and measuring the consistency with images rendered there. WarpRF is simple and inexpensive, does not require any training, and can be applied to any radiance field implementation for free. WarpRF excels at both uncertainty quantification and downstream tasks, e.g., active view selection and active mapping, outperforming any existing method tailored to specific frameworks.

Figures

Figures reproduced from arXiv: 2506.22433 by the authors.

Figure 1
Figure 1. Active mapping through WarpRF uncertainty quantification. Given a radiance field-based surface reconstruction framework trained on an initial set of images – e.g. the 3D Gaussian Splatting 3DGSt0 (left) trained on the blue viewpoints – WarpRF estimates the best next view (in orange in the figure) by quantifying the rendering uncertainty U0 associated to it through warping. This is added to the training set and used … view at source ↗
Figure 2
Figure 2. Uncertainty quantification with WarpRF. Given an initial set of source viewpoints S = {S0, S1, S2, . . . , SS} available for training, a radiance field is optimized over images {IS0 , IS1 , IS2 , . . . , ISS }. Then, upon selecting a novel target viewpoint T , uncertainty can be quantified by rendering depth DT and use it to project either images ISn or depth maps DSn – in this figure, we show an example of the latt… view at source ↗
Figure 3
Figure 3. Uncertainty Visualization. From left to right: (a) rendered image, (b) rendered depth, (c) FisherRF uncertainty, (d) WarpRF uncertainty, (e) depth error. and then use the rendered depth DS to project it into the target pose T by means of backward warping, denoted as ISn→T . The image-level uncertainty UT is then computed as: UT = X i,j min n [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative results of active camera selection on NeRF Synthetic dataset with 10 views. From top to bottom, we collect rendered results by 3DGS trained with to active view selection being performed by Manifold, FIsherRF and WarpRF, followed by the real images. and util…
Figure 5
Figure 5. Figure 5: Qualitative results of active camera selection on MipNeRF360 dataset. From top to bottom, we collect rendered results by 3DGS trained with to active view selection being performed by Manifold, FIsherRF and WarpRF, followed by the real images. In this experiment, we eva…
Figure 6
Figure 6. Figure 6: Reconstructed meshes using active mapping on Tanks and Temples dataset. We show meshes reconstructed by 3DGS trained through active mapping being performed by Fish￾erRF and WarpRF. ping, at the expense of more intrusive solutions tailored for the specific radiance fiel…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages

  1. [1]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470– 5479, 2022. 6

    work page 2022

  2. [2]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 3, 7

    arXiv 2024

  3. [3]

    Depth-supervised NeRF: Fewer views and faster training for free

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ra- manan. Depth-supervised NeRF: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,

    work page

  4. [4]

    Accurate, dense, and ro- bust multiview stereopsis

    Yasutaka Furukawa and Jean Ponce. Accurate, dense, and ro- bust multiview stereopsis. IEEE TPAMI, 32(8):1362–1376,

    work page

  5. [5]

    A survey of uncertainty in deep neural networks

    Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. Arti- ficial Intelligence Review, 56(Suppl 1):1513–1589, 2023. 1

    work page 2023

  6. [6]

    Cl ´ement Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J. Brostow. Digging into self-supervised monocular depth prediction. In ICCV, 2019. 2, 5

    work page 2019

  7. [7]

    Bayes’ Rays: Uncertainty quantifica- tion in neural radiance fields

    Lily Goli, Cody Reading, Silvia Sell ´an, Alec Jacobson, and Andrea Tagliasacchi. Bayes’ Rays: Uncertainty quantifica- tion in neural radiance fields. In CVPR, 2024. 2, 5

    work page 2024

  8. [8]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. In CVPR, 2024. 1, 3

    work page 2024

Show all 41 references
  1. [9]

    Scone: Surface coverage optimization in unknown environ- ments by volumetric integration

    Antoine Gu ´edon, Pascal Monasse, and Vincent Lepetit. Scone: Surface coverage optimization in unknown environ- ments by volumetric integration. In NeurIPS, pages 20731– 20743, 2022. 2

    2022

  2. [10]

    Macarons: Mapping and coverage anticipation with rgb online self-supervision

    Antoine Gu ´edon, Tom Monnier, Pascal Monasse, and Vin- cent Lepetit. Macarons: Mapping and coverage anticipation with rgb online self-supervision. In CVPR, pages 940–951,

  3. [11]

    Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field

    Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. In ECCV, pages 93–

  4. [12]

    Fisherrf: Ac- tive view selection and mapping with radiance fields using fisher information

    Wen Jiang, Boshu Lei, and Kostas Daniilidis. Fisherrf: Ac- tive view selection and mapping with radiance fields using fisher information. In ECCV, pages 422–440, 2024. 2, 5, 6, 7, 8

    2024

  5. [13]

    What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS,

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS,

  6. [14]

    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, 42 (4), 2023. 1, 2, 3, 5, 6

    2023

  7. [15]

    4d gaus- sian splatting in the wild with uncertainty-aware regulariza- tion

    Mijeong Kim, Jongwoo Lim, and Bohyung Han. 4d gaus- sian splatting in the wild with uncertainty-aware regulariza- tion. Advances in Neural Information Processing Systems , 37:129209–129226, 2025. 2

    2025

  8. [16]

    Sources of uncertainty in 3d scene reconstruction

    Marcus Klasson, Riccardo Mereu, Juho Kannala, and Arno Solin. Sources of uncertainty in 3d scene reconstruction. In ECCV. Springer, 2024. 2

    2024

  9. [17]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) , 36 (4):1–13, 2017. 7

    2017

  10. [18]

    Uncertainty guided pol- icy for active robotic 3d reconstruction using neural radiance fields

    Soomin Lee, Le Chen, Jiahao Wang, Alexander Liniger, Suryansh Kumar, and Fisher Yu. Uncertainty guided pol- icy for active robotic 3d reconstruction using neural radiance fields. IEEE Robotics and Automation Letters, 7(4):12070– 12077, 2022. 2, 7, 8

    2022

  11. [19]

    Manifold sampling for differentiable uncer- tainty in radiance fields

    Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollh ¨ofer, Thomas Leimk ¨uehler, and Chris- tian Theobalt. Manifold sampling for differentiable uncer- tainty in radiance fields. In SIGGRAPH Asia Conference Proceedings, 2024. 2, 5, 6, 7

    2024

  12. [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 syn- thesis. In ECCV, 2020. 1, 2, 5, 6

    2020

  13. [21]

    Ac- tivenerf: Learning where to see with uncertainty estimation

    Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. Ac- tivenerf: Learning where to see with uncertainty estimation. In ECCV, pages 230–246. Springer, 2022. 2, 6

    2022

  14. [22]

    On the confidence of stereo matching in a deep- learning era: a quantitative evaluation

    M Poggi, S Kim, F Tosi, F Aleotti, D Min, K Sohn, and S Mattoccia. On the confidence of stereo matching in a deep- learning era: a quantitative evaluation. IEEE TPAMI, 44(9): 5293–5313, 2022. 2

    2022

  15. [23]

    Neurar: Neural uncertainty for autonomous 3d reconstruction with implicit neural representations

    Yunlong Ran, Jing Zeng, Shibo He, Jiming Chen, Lincheng Li, Yingfeng Chen, Gimhee Lee, and Qi Ye. Neurar: Neural uncertainty for autonomous 3d reconstruction with implicit neural representations. IEEE Robotics and Automation Let- ters, 8(2):1125–1132, 2023. 2

    2023

  16. [24]

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

    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. In CVPR, pages 8931–8940, 2024. 2

    2024

  17. [25]

    Barron, Ben Mildenhall, Pratul P

    Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth pri- ors for neural radiance fields from sparse input views. In 9 Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2022. 4

    2022

  18. [26]

    Self-evolving depth-supervised 3d gaussian splatting from rendered stereo pairs

    Sadra Safadoust, Fabio Tosi, Fatma G ¨uney, and Matteo Poggi. Self-evolving depth-supervised 3d gaussian splatting from rendered stereo pairs. In BMVC, 2024. 4, 5

    2024

  19. [27]

    A multi-view stereo benchmark with high- resolution images and multi-camera videos

    Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and An- dreas Geiger. A multi-view stereo benchmark with high- resolution images and multi-camera videos. In CVPR, pages 3260–3269, 2017. 5

    2017

  20. [28]

    Stochastic neural radiance fields: Quan- tifying uncertainty in implicit 3d representations

    Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quan- tifying uncertainty in implicit 3d representations. CoRR, abs/2109.02123, 2021. 2

    2021 arXiv

  21. [29]

    Conditional-flow nerf: Accurate 3d mod- elling with reliable uncertainty quantification

    Jianxiong Shen, Antonio Agudo, Francesc Moreno-Noguer, and Adria Ruiz. Conditional-flow nerf: Accurate 3d mod- elling with reliable uncertainty quantification. In ECCV,

  22. [30]

    Estimating 3d uncertainty field: Quantify- ing uncertainty for neural radiance fields

    Jianxiong Shen, Ruijie Ren, Adria Ruiz, and Francesc Moreno-Noguer. Estimating 3d uncertainty field: Quantify- ing uncertainty for neural radiance fields. In 2024 IEEE In- ternational Conference on Robotics and Automation (ICRA), pages 2375–2381. IEEE, 2024. 2

    2024

  23. [31]

    imap: Implicit mapping and positioning in real-time

    Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davi- son. imap: Implicit mapping and positioning in real-time. In ICCV, pages 6229–6238, 2021. 7

    2021

  24. [32]

    Sparse voxels rasterization: Real-time high-fidelity radiance field rendering

    Cheng Sun, Jaesung Choe, Charles Loop, Wei-Chiu Ma, and Yu-Chiang Frank Wang. Sparse voxels rasterization: Real-time high-fidelity radiance field rendering. ArXiv, abs/2412.04459, 2024. 2, 4, 5, 6

    2024 arXiv

  25. [33]

    Density-aware nerf ensembles: Quantifying predictive un- certainty in neural radiance fields

    Niko S ¨underhauf, Jad Abou-Chakra, and Dimity Miller. Density-aware nerf ensembles: Quantifying predictive un- certainty in neural radiance fields. In ICRA, 2023. 2

    2023

  26. [34]

    Dn-splatter: Depth and normal priors for gaussian splatting and meshing

    Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splatting and meshing. In WACV, 2025. 5

    2025

  27. [35]

    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 NeurIPS, 2021. 1, 3

    2021

  28. [36]

    Neural visibility field for uncertainty-driven active mapping

    Shangjie Xue, Jesse Dill, Pranay Mathur, Frank Dellaert, Panagiotis Tsiotras, and Danfei Xu. Neural visibility field for uncertainty-driven active mapping. In CVPR, pages 18122– 18132, 2024. 2, 7, 8

    2024

  29. [37]

    Active implicit object reconstruction us- ing uncertainty-guided next-best-view optimization

    Dongyu Yan, Jianheng Liu, Fengyu Quan, Haoyao Chen, and Mengmeng Fu. Active implicit object reconstruction us- ing uncertainty-guided next-best-view optimization. IEEE Robotics and Automation Letters, pages 1–8, 2023. 2

    2023

  30. [38]

    Active neural mapping

    Zike Yan, Haoxiang Yang, and Hongbin Zha. Active neural mapping. In ICCV, 2023. 2

    2023

  31. [39]

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

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, pages 12–22, 2023. 5

    2023

  32. [40]

    Efficient 3d object segmentation from densely sampled light fields with applications to 3d re- construction

    Kaan Y ¨ucer, Alexander Sorkine-Hornung, Oliver Wang, and Olga Sorkine-Hornung. Efficient 3d object segmentation from densely sampled light fields with applications to 3d re- construction. ACM Transactions on Graphics (TOG), 35(3): 1–15, 2016. 5

    2016

  33. [41]

    WarpRF: Multi-View Consistency for Training-Free Uncertainty Quantification and Applications in Radiance Fields

    Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. Activermap: Radiance field for active mapping and planning, 2022. 2 10 WarpRF: Multi-View Consistency for Training-Free Uncertainty Quantification and Applications in Radiance Fields Supplementary Material T...

    2022

Pith tools

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