Pith. sign in

REVIEW 2 major objections 4 minor 33 references

Posterior Variance Is a Constraint Map, Not an Error Map: Closed-Form Uncertainty for Radiative Gaussian Splatting in Sparse-View CT

T0 review · 2 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Posterior variance in sparse-view CT is a constraint map, not an error map.

desk verdict A rare calibration study that actually diagnoses the instrument: the full-volume/foreground reversal is real, and the closed form is a useful tool. read the letter →

arxiv 2607.13682 v2 pith:DWYYB6HM submitted 2026-07-15 cs.CV cs.LGeess.IV

Chulin Zhao , Yiran Xu , Shu Liu This is my paper
classification cs.CVcs.LGeess.IV
keywords posteriorvarianceradiativeGaussiansplattingsparse-viewCTuncertaintycalibrationmaskedbiasdecompositionclosed-formPoissondoseresponse
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 establish that per-voxel posterior variance in radiative Gaussian splatting CT is a data-constraint map, not an error map. On the official 15-scene benchmark, the uncertainty ranks true error over the full volume on 14 of 15 scenes, but restricted to foreground tissue—the region a clinician reads—the ranking collapses to a median Spearman of 0.11 with 0 of 15 scenes passing, identically for the paper's rectified closed form, a deep ensemble, and a strictly positive log-normal posterior. The mechanism is structural: roughly 90% of in-object error reproduces across retrainings and is invisible to disagreement, while 73–81% of the full-volume correlation comes from object/surround contrast. The paper also derives an exact one-pass closed-form variance and shows that the error scale, unlike location, is repairable by reparameterizing the posterior, making one scene-agnostic temperature transferable and tracking photon count at the Poisson-predicted −1/2 power. If correct, the result changes how uncertainty maps should be read: alarms are trustworthy, all-clears are not.

What carries the argument

The load-bearing identity is the strict linearity of X-ray rendering in the per-Gaussian densities: both projection and volume are weighted sums of densities, unlike RGB splatting. Squaring a Gaussian kernel halves its covariance, so the unchanged rasterizer/voxelizer evaluates the predictive variance exactly in one extra forward pass, with a clamp-aware form for rectified moments. Two analytic results carry the diagnosis: the law of total covariance split into within-group and between-group terms (showing global correlation can be pure group contrast), and an exact-posterior Kendall reference that computes the best possible error-ranking score for a given sigma-spread.

What would settle it

Measure the seed-reproducible bias fraction with seeds that vary architecture, optimizer, and initialization; if the bias fraction drops far below ~90% or a disagreement-based construction reaches foreground Spearman ≈0.6 on the 15-scene masked protocol, the structural claim collapses. Alternatively, compute the exact-posterior reference under a positively correlated error model; if the reference rises enough to cover the observed values, the claimed deficit disappears.

Watch

Extended reading notes

Core claim

The central claim is that posterior variance measures where the data constrain the reconstruction, not where the reconstruction errs. Over the whole volume the two nearly coincide because the dominant error is the unconstrained surround and boundary, but inside a well-exposed organ they decouple: the map cannot rank the one realized error field, even though it is exactly calibrated in the repeated-sampling sense. The paper demonstrates this with three uncertainty constructions from two estimator families, traces the collapse to a seed decomposition in which ~90% of foreground error is bias that reproduces across retrainings, and quantifies the illusion with a covariance decomposition showing

Load-bearing premise

The quantified in-object deficit assumes realized errors are draws e_j = sigma_j z_j with independent standard-normal z_j; real residuals are spatially correlated and can be correlated with sigma, so the "four- to five-fold" shortfall is only as strong as that independence idealization.

Editorial extensions

If this is right

  • Read uncertainty as one-sided: high-sigma flags low-constraint voxels and can drive re-scan triage, but low-sigma inside an organ certifies nothing.
  • Evaluation practice should report masked calibration on the clinically read region; full-volume numbers can be inflated by object/surround contrast.
  • A deep ensemble does not repair localization: any disagreement-based estimator inherits the same blindness to seed-reproducible bias.
  • The error scale is an engineering problem: a log-normal posterior with a fixed prior and explicit likelihood makes a single temperature transferable across scenes and physically responsive to photon count.
  • Exact one-pass variance is available at roughly 8× lower cost than the Monte-Carlo estimator and can replace it without changing conclusions.

Reading between the lines

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

  • By implication, any linear inverse problem with a large null space—not just sparse-view CT—may show the same one-sided validity for variance-based uncertainty, since the dominant unmeasured component is reproducible bias.
  • A testable extension: train a bias regressor on synthetic ground truth that consumes the exact one-pass variance as an input feature; the paper suggests this route, and its closed form isolates the variance term such a model must remove.
  • On real clinical scans, voxel truth is unavailable, so the diagnosis would need phantom-based masked calibration or co-registered prior studies to verify whether the foreground collapse persists outside synthetic benchmarks.
  • The independence assumption in the exact-posterior reference (errors as exact draws from the predictive) is the main place the quantified deficit could be overstated; positive error–sigma correlation would raise the reference and shrink the gap.
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

2 major / 4 minor

Summary. The paper studies per-voxel uncertainty in radiative Gaussian splatting for sparse-view CT, proposing a closed-form propagation of a factorized variational posterior over per-Gaussian densities. Evaluating on the official 15-scene benchmark, it reports that full-volume rank correlation with true error is high (14/15 scenes above 0.6), but restricted to foreground tissue the ranking collapses (median Spearman ~0.11, 0/15 pass) for three different uncertainty constructions. Using a covariance decomposition and an exact-posterior reference, it attributes the collapse to foreground error being dominated by seed-reproducible bias and to contrast inflation. It then shows that reparameterizing the posterior with a log-normal distribution repairs the error-scale temperature transfer and dose-response behavior while leaving the foreground ranking collapse unchanged. The paper concludes that posterior variance is a data-constraint map with one-sided validity, and proposes evaluation practices including masked calibration.

Significance. If the diagnosis is correct, this is an important negative result for uncertainty quantification in Gaussian-splatting CT and for variance-based UQ in null-space-dominated inverse problems more broadly. The paper's strengths include an exact closed-form uncertainty readout with a single forward pass, a strict toy control that validates both theoretical results, and a machine-readable release of per-run evidence, seeds, and protocols. The empirical core is unusually thorough: 75 official runs, 3 seeds for the variational column, threshold robustness, and per-scene tables. The finding that disagreement-based estimators (deep ensemble, log-normal) also fail in the foreground, despite being better calibrated than the proposed method on full-volume metrics, supports the structural nature of the diagnosis. The proposed evaluation practices (masked calibration, bias decomposition, exact-posterior reference) are a useful contribution to the field.

major comments (2)
  1. [Sec. 5.3] The seed-decomposition bias fraction is load-bearing but undefined. The text states that 'training K=5 independent seeds and splitting each voxel's error into a seed-varying component and a seed-reproducible component' yields a median bias fraction of 0.906, but no equation or algorithmic description is given for how this fraction is computed. Please specify the per-voxel decomposition (e.g., bias^2 = mean_k(e_k)^2, var = mean_k((e_k - mean_k(e_k))^2), and how the fraction is aggregated over voxels), how the 5 seeds are initialized (shared backbone and initialization scheme?), and report sensitivity to K and seed set. Without this, the central numerical claim cannot be independently audited.
  2. [Sec. 5.2 and Abstract] The 'constraint map' reading is partly built into the method: the KL prior in Eq. (3) deliberately anchors s_i around sigma0 in the null space, so 'unobserved = high variance' is a design choice. The empirical discovery is the one-sided validity (alarms trustworthy, all-clears not), and the full-volume/foreground reversal. The paper should state this distinction explicitly to avoid the impression that the constraint-map property is empirically discovered rather than in part constructed. Consider softening the title claim or adding a sentence in Sec. 5.3 that separates the prior-induced behavior from the empirical reversal.
minor comments (4)
  1. [Sec. 3.3] The claim of 'zero CUDA changes' is qualified by two input-side safeguards (Appendix D.2–D.3: float64 moment evaluation and variance-channel pre-scaling). Please mention these safeguards in the main text for reproducibility.
  2. [Table A3] The 'gate' row would be clearer if it indicated per scene whether the failure is due to Spearman or ECE (e.g., LN fails on jaw for Spearman and on beetle for ECE). This would help readers understand the 13/15 vs 9/15 counts.
  3. [Abstract] The phrase 'identically for a deep ensemble and for a strictly positive log-normal posterior' overstates the agreement; median foreground Spearmans are 0.108, 0.199, and 0.106. Suggest 'similarly' or 'for all three constructions'.
  4. [Appendix B.4] The ensemble comparison rests on a single K=5 run per scene while the variational column is a 3-seed mean; this asymmetry is stated but should be remembered in any table caption where the two are compared directly.

Circularity Check

2 steps flagged · score 3.0 of 10

Mild partial circularity: the constraint-map positive half is built into the KL prior and the Poisson dose exponent is a same-model consistency check; the central foreground-collapse diagnosis is independent.

  1. self definitional [Sec. 3.2, Eq. (3); operationalized in Sec. 5.3 restatement]
    "The prior is centred at the current mean, so the KL regularizes only the scale: for Gaussians no ray constrains — the null space of sparse-view tomography — the data term supplies no gradient to s_i and the prior alone anchors s_i≈σ_0. This is how 'uncertain where unobserved' arises"

    The high-σ side of the headline claim ('posterior variance is a constraint map') is an input to the training objective: Eq. (3) deliberately anchors s_i≈σ0 wherever no ray constrains the density, and Eq. (4) defines σ^2(x)=Σ g_i^2 s_i^2. Therefore 'high σ marks low-constraint voxels' follows from the chosen prior by construction rather than from an independent empirical finding. The paper is transparent about this, and the foreground-collapse (error-map) half is separately measured, so this is partial circularity, not a full reduction.

  2. other [Sec. 6 vs. Appendix G (dose-response experiment)]
    "Replacing the data term with an explicit Poisson transmission likelihood — counts c_p∼Poisson(I0 e^{−ℓ_p}) synthesized at three doses spanning two decades — and retraining the log-normal posterior at each dose, the absolute posterior scale follows a power law in I0 with exponents −0.470 to −0.481 against the Fisher-predicted −1/2. ... [Appendix G:] the likelihood used in training is the exact generative model of the synthesis — so the −0.48 exponent certifies internal physical consistency, not real-detector behavior."

    The predicted −1/2 exponent is an analytic property of the Poisson likelihood (Fisher information scaling), and the experimental data are synthesized from that same likelihood. Measuring −0.48 therefore validates the inference machinery against the model's own assumption rather than testing an external physical prediction; calling it 'a measured physical response, not a fitted constant' overstates the self-consistency check. The appendix discloses this, so the circularity is mild and not central to the uncertainty diagnosis.

full rationale

The central negative claim — that variance-based uncertainty cannot rank error inside foreground tissue — is established independently of the two flagged design choices: the foreground collapse (median Spearman 0.11, 0/15 gate passes) is measured against synthetic ground truth and reproduces across three constructions and two estimator families (closed form, deep ensemble, log-normal); the exact toy control and the closed-form variance identity are mathematical derivations, not fitted outputs; and there is no load-bearing self-citation chain. The two mild circularities are: (1) the positive 'constraint map' half is engineered through the KL prior that anchors s_i in the null space, so 'high σ = unconstrained' is partly built into the objective; and (2) the Poisson dose-response exponent is a check of the posterior against data generated from the same Poisson likelihood that supplies the predicted scaling. Neither reduces the main error-map diagnosis, so a low-moderate score is appropriate.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

No new physical or model entities are postulated; the 'constraint map' is an interpretive relabeling of the existing posterior-variance field, not an invented quantity. The central derivation uses only the linearity of the forward model, Gaussian moment identities, and the law of total covariance. The free parameters are calibration hyperparameters, a per-scene temperature, and a fitted dose-response slope; none of these is a new physical constant or entity.

free parameters (6)
  • KL prior scale sigma0 = 0.5
    Hand-set anchor for null-space variance (Sec. 3.2); determines the absolute scale of the resulting variance map.
  • KL weight lambda_KL = 1e-3
    Weight of KL term; ablation (Table A9) shows monotone calibration improvement from 0 to 1e-2, so 1e-3 is a chosen operating point on a fitted curve.
  • Temperature tau = per-scene 1.59-99.50; global 13.21; log-normal LOSO 9.9-10.9
    Fitted per scene or globally to minimize interval-coverage ECE; rank-invariant but central to magnitude-calibration claims.
  • Log-normal prior rho0, sigma0 = rho0=0.05, sigma0=0.5
    Fixed by hand for the control posterior; part of the repair that makes temperature transferable.
  • Dose-response power-law exponent = -0.470 to -0.481
    Slope fitted to three dose points in Table A12; compared against the Fisher-predicted -1/2; a validation fit, not a prediction free of fitting.
  • Foreground mask threshold = 5% of peak; robustness checked 1-10%
    Defines the clinically read tissue; the 0/15 collapse holds across thresholds, so it is a reported choice rather than a tuned parameter.
assumptions (7)
  • domain assumption Rendering and voxelization are strictly linear in per-Gaussian densities (Eq. 2).
    Inherited from R2-Gaussian [32]; without it the closed-form variance propagation fails.
  • domain assumption Per-Gaussian densities are independent after training.
    Factorized variational posterior; discards along-ray correlations; acknowledged in Sec. 7 and App. G.
  • standard math Rectified moments of clamped Gaussians (Eq. 5) are the exact moments of the deployed predictive distribution.
    Derived via Gaussian integrals; matches the training-time clamp.
  • ad hoc to paper Perfect-posterior reference uses e_j = sigma_j z_j with i.i.d. z_j.
    Idealized error model for Result 2 (Sec. 5.2); flagged by the authors as a yardstick, not an upper bound.
  • ad hoc to paper Poisson transmission likelihood used in training is the exact generative model of the synthesized counts.
    Makes the -1/2 dose exponent an internal consistency check; stated in App. G.
  • domain assumption Synthetic ground truth permits voxel-wise calibration.
    Real scans lack exact references; clinical transfer untested.
  • standard math Law of total covariance and standard Gaussian integral identities.
    Used for Eq. 7 and Eq. 8; no extra assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Posterior Variance Is a Constraint Map, Not an Error Map: Closed-Form Uncertainty for Radiative Gaussian Splatting in Sparse-View CT." pith.science (2026). https://pith.science/paper/DWYYB6HM

@misc{pith2026260713682,
  author       = {Pith},
  title        = {Pith review of: Posterior Variance Is a Constraint Map, Not an Error Map: Closed-Form Uncertainty for Radiative Gaussian Splatting in Sparse-View CT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DWYYB6HM}},
  note         = {Machine review of arXiv:2607.13682}
}
read the original abstract

Radiative Gaussian splatting reconstructs sparse-view CT fast and accurately, and recent work attaches per-Gaussian posteriors to yield per-voxel uncertainty maps. We ask what such a map actually measures: posterior variance is a data-constraint map, not an error map -- its alarms are trustworthy, its all-clears are not. Exploiting the strict linearity of X-ray rendering in the per-Gaussian densities, we derive a clamp-aware closed form that the unchanged rasterizer evaluates exactly in one forward pass, in volume and projection space: the infinite-sample limit of the sampling estimator of concurrent work, at ~8x lower cost. On the official 15-scene benchmark this uncertainty ranks true error on 14 of 15 scenes. Restricted to the object interior -- the tissue a clinician reads -- the ranking collapses (median Spearman 0.11, 0/15 pass), identically for a deep ensemble and for a strictly positive log-normal posterior: three constructions, two estimator families, no survivors. The mechanism is structural: about 90% of in-object error is bias that reproduces across retrainings, invisible to model disagreement; 73-81% of the full-volume correlation is carried by object/surround contrast; and an exactly solvable control puts the observed in-object ranking 4-5x below what a perfectly calibrated posterior with the same sigma-spread would score. The error scale, by contrast, is an engineering problem, and we solve it: reparameterizing the posterior contracts the cross-scene temperature spread from 19.3x to 2.6x, one scene-agnostic temperature transfers to unseen scenes (10/15 leave-one-scene-out), and the repaired scale tracks photon count at the Poisson-predicted -1/2 power. We distill evaluation practice that would have caught the illusion -- masked calibration, seed-wise bias decomposition, an exact-posterior reference -- and release all protocols, seeds and per-run evidence.

Figures

Figures reproduced from arXiv: 2607.13682 by the authors.

Figure 1
Figure 1. Training and uncertainty pipeline. We extend R2 -Gaussian [12] with a variational density posterior ρi ∼ N (mi , s2 i ) per Gaussian. Training renders reparameterized density samples through the unchanged rasterizer and applies a KL prior that anchors si where no ray constrains the density (Sec. 3.2). At evaluation, mean densities mi define the reconstruction V¯ , and the √ 2-scaling identity turns the same rasteriz… view at source ↗
Figure 1
Figure 1. Where the ranking lives. Per-scene Spearman ρ(σ, e) over the full volume (left) and inside foreground tissue (right) for three uncertainty constructions. The full-volume signal is carried by object/surround contrast; inside the object none of the three reaches the gate line on any scene. The ensemble is better calibrated; ours is exact and ∼5× cheaper. On the benchmark the ensemble attains lower scaled ECE on 15/15 … view at source ↗
Figure 2
Figure 2. Coverage-gated calibrated acquisition (CGCA). Phase 1 (model-free): while the maximum uncovered arc exceeds the gate threshold γ, acquire the midpoint of the largest gap — degenerate clustering, the only catastrophic failure mode, is prevented geometrically. Phase 2 (calibrated): once coverage is secured, retrain to convergence on the acquired set and pick the candidate maximizing the one-pass closed-form projection… view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Three-way calibration visualization (chest, 25 views). Left: sparsification curves (lower AUSE is better); the deep ensemble and our variational posterior track the oracle lower bound, while the perturbation curve’s residual error climbs back after ≈65% of voxels are r…
Figure 4
Figure 4. Figure 4: View-level oracle spectra (three regimes). Bars: true one-step gain; markers: argmax of each score. Symmetric self-test. On uniform chest, neither our closed-form score nor the competitor’s shows a detectable positive rank correlation with true oracle gains (point esti…
Figure 5
Figure 5. Figure 5: Four-arm closed-loop selections (6 views, 12k iter). Shaded arc: largest uncovered angular gap; PSNR labels from final eval [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Metric confound analysis (chest, 6 training views). Pixel variance chases path￾length thickness; SSIM variance is less confounded. Findings. On abdomen, the coverage shell is decisive: random collapses to 25.61 dB while geo recovers 26.43 dB (+0.82 dB). With the shell …

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 4 linked inside Pith

  1. [1]

    Deep bayesian inversion.arXiv preprint arXiv:1811.05910, 2018

    Jonas Adler and Ozan Öktem. Deep bayesian inversion.arXiv preprint arXiv:1811.05910, 2018

    arXiv 2018

  2. [2]

    Angelopoulos, Amit Pal Kohli, Stephen Bates, Michael I

    Anastasios N. Angelopoulos, Amit Pal Kohli, Stephen Bates, Michael I. Jordan, Jitendra Malik, Thayer Alshaabi, Srigokul Upadhyayula, and Yaniv Romano. Image-to-image regression with distribution-free uncertainty quantification and applications in imaging. InInternational Conference on Machine Learning (ICML), 2022

    2022

  3. [3]

    Weight uncertainty in neural network

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. InInternational Conference on Machine Learning (ICML), 2015

    2015

  4. [4]

    Radiative gaussian splatting for efficient x-ray novel view synthesis

    Yuanhao Cai, Yixun Liang, Jiahao Wang, Angtian Wang, Yulun Zhang, Xiaokang Yang, Zongwei Zhou, and Alan Yuille. Radiative gaussian splatting for efficient x-ray novel view synthesis. InEuropean Conference on Computer Vision (ECCV), 2024

    2024

  5. [5]

    Radioactive 3d gaussian ray tracing for tomographic reconstruction.arXiv preprint arXiv:2602.01057, 2026

    Ling Chen and Bao Yang. Radioactive 3d gaussian ray tracing for tomographic reconstruction.arXiv preprint arXiv:2602.01057, 2026

    arXiv 2026

  6. [6]

    Uncertainty quantification for deep unrolling-based computational imaging.IEEE Transactions on Computational Imaging, 8, 2022

    Canberk Ekmekci and Müjdat Çetin. Uncertainty quantification for deep unrolling-based computational imaging.IEEE Transactions on Computational Imaging, 8, 2022

    2022

  7. [7]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. InInternational Conference on Machine Learning (ICML), 2016

    2016

  8. [8]

    Bayes’ rays: Uncertainty quantification for neural radiance fields

    Lily Goli, Cody Reading, Silvia Sellán, Alec Jacobson, and Andrea Tagliasacchi. Bayes’ rays: Uncertainty quantification for neural radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

Show all 33 references
  1. [9]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. InInternational Conference on Machine Learning (ICML), 2017

    2017

  2. [10]

    Uncertainty estimates and multi-hypotheses networks for optical flow

    Eddy Ilg, Özgün Çiçek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. Uncertainty estimates and multi-hypotheses networks for optical flow. InEuropean Conference on Computer Vision (ECCV), 2018

    2018

  3. [11]

    Rendering-aware bayesian 3d gaussian splatting with native uncertainty and adaptive complexity control.arXiv preprint arXiv:2607.05522, 2026

    Gaoxiang Jia, Vikram Appia, Junzhou Huang, and Xinlei Wang. Rendering-aware bayesian 3d gaussian splatting with native uncertainty and adaptive complexity control.arXiv preprint arXiv:2607.05522, 2026

    2026 arXiv

  4. [12]

    Fisherrf: Activeviewselectionanduncertaintyquantification for radiance fields using fisher information

    WenJiang, BoshuLei, andKostasDaniilidis. Fisherrf: Activeviewselectionanduncertaintyquantification for radiance fields using fisher information. InEuropean Conference on Computer Vision (ECCV), 2024

    2024

  5. [13]

    What uncertainties do we need in bayesian deep learning for computer vision? InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  6. [14]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), 2023

    2023

  7. [15]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. InInternational Conference on Learning Representations (ICLR), 2014

    2014

  8. [16]

    Accurate uncertainties for deep learning using calibrated regression

    Volodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated regression. InInternational Conference on Machine Learning (ICML), 2018

    2018

  9. [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. InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  10. [18]

    Evaluating and calibrating uncertainty prediction in regression tasks.Sensors, 22(15), 2022

    Dan Levi, Liran Gispan, Niv Giladi, and Ethan Fetaya. Evaluating and calibrating uncertainty prediction in regression tasks.Sensors, 22(15), 2022. 10

    2022

  11. [19]

    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. InEuropean Conference on Computer Vision (ECCV), 2020

    2020

  12. [20]

    Cooper, and Milos Hauskrecht

    Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabil- ities using bayesian binning. InAAAI Conference on Artificial Intelligence, 2015

    2015

  13. [21]

    Activenerf: Learning where to see with uncertainty estimation

    Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. Activenerf: Learning where to see with uncertainty estimation. InEuropean Conference on Computer Vision (ECCV), 2022

    2022

  14. [22]

    Deep null space learning for inverse problems: Convergence analysis and rates.Inverse Problems, 35(2):025008, 2019

    Johannes Schwab, Stephan Antholzer, and Markus Haltmeier. Deep null space learning for inverse problems: Convergence analysis and rates.Inverse Problems, 35(2):025008, 2019

    2019

  15. [23]

    Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations

    Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. InInternational Conference on 3D Vision (3DV), 2021

    2021

  16. [24]

    Solving inverse problems in medical imaging with score-based generative models

    Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  17. [25]

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

    Niko Sünderhauf, Jad Abou-Chakra, and Dimity Miller. Density-aware nerf ensembles: Quantifying predictive uncertainty in neural radiance fields. InIEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  18. [26]

    Varsplat: Uncertainty-aware 3d gaussian splatting for robust rgb-d slam

    Anh Thuan Tran and Jana Kosecka. Varsplat: Uncertainty-aware 3d gaussian splatting for robust rgb-d slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    2026

  19. [27]

    Buckley, and Tim Verbelen

    Toon Van de Maele, Ozan Çatal, Alexander Tschantz, Christopher L. Buckley, and Tim Verbelen. Variational bayes gaussian splatting.arXiv preprint arXiv:2410.03592, 2024

    2024

  20. [28]

    Uncertainr: Uncertainty quantifica- tion of end-to-end implicit neural representations for computed tomography.Transactions on Machine Learning Research (TMLR), 2023

    Francisca Vasconcelos, Bobby He, Nalini Singh, and Yee Whye Teh. Uncertainr: Uncertainty quantifica- tion of end-to-end implicit neural representations for computed tomography.Transactions on Machine Learning Research (TMLR), 2023. arXiv:2202.10847

    2023 arXiv

  21. [29]

    Horseshoe splatting: Handling structural sparsity for uncertainty-aware gaussian-splatting radiance field rendering

    Feng Wu, Tsai Hor Chan, Yihang Chen, Lingting Zhu, Guosheng Yin, and Lequan Yu. Horseshoe splatting: Handling structural sparsity for uncertainty-aware gaussian-splatting radiance field rendering. InThe Fourteenth International Conference on Learning Representations (ICLR), 2026

    2026

  22. [30]

    Active view selection with perturbed gaussian ensemble for tomographic reconstruction.arXiv preprint arXiv:2603.06852, 2026

    Yulun Wu, Ruyi Zha, Wei Cao, Yingying Li, Yuanhao Cai, and Yaoyao Liu. Active view selection with perturbed gaussian ensemble for tomographic reconstruction.arXiv preprint arXiv:2603.06852, 2026. Accepted to ECCV 2026

    2026 arXiv

  23. [31]

    Exact-gs: Mathematically rigorous and accurate 3d gaussian splatting for 3d x-ray reconstruction

    Guangpu Yang, Steffen Kieß, Hanxiang Luo, Xingyu Liu, and Sven Simon. Exact-gs: Mathematically rigorous and accurate 3d gaussian splatting for 3d x-ray reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4902–4911, 2026

    2026

  24. [32]

    R2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction

    Ruyi Zha, Tao Jun Lin, Yuanhao Cai, Jiwen Cao, Yanhao Zhang, and Hongdong Li. R2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  25. [33]

    Share” is the between-group fraction of the global covarianceCov(σ,e ); “Ref

    Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross. Ewa volume splatting. In Proceedings of IEEE Visualization, pages 29–36, 2001. 11 Figure A1: Training and read-out pipeline.We extend R2-Gaussian [32] with a variational density posterior ρi∼N (mi,s 2 i )p...

    2001

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.