Pith. sign in

REVIEW 3 major objections 5 minor 66 references

A Probability-guided Sampler for Neural Implicit Surface Rendering

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A surface renderer's own signed distance field can be recycled into an occlusion-aware per-camera ray-sampling distribution, and with new surface losses it improves 3D reconstruction for NeuS and Neuralangelo with no extra input.

desk verdict The new sampler is clever and mostly well-engineered, but the paper never isolates it from its surface loss, so the central claim remains unsupported. read the letter →

arxiv 2506.08619 v1 pith:7UE46XDI submitted 2025-06-10 cs.CV

Gonçalo Dias Pais , Valter Piedade , Moitreya Chatterjee , Marcus Greiff , Pedro Miraldo This is my paper
classification cs.CV
keywords neuralimplicitsurfacesnon-uniformsamplingprobabilitydensityfunctionsigneddistancefunctionsrayradiancefields3Dreconstructionocclusion-aware
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 argues that neural implicit surface renderers waste a fixed training budget by sampling image rays uniformly, and that the signed distance field (SDF) the renderer is already learning can be reused as a guide for where to sample. The method converts the SDF into a probability density over a per-camera three-dimensional image space whose axes are pixel coordinates plus depth, weights that density by a transmittance estimate so occluded regions are down-weighted, and draws training rays from the result via inverse transform sampling. The depth attached to each sampled ray then anchors surface reconstruction losses with near-surface, empty-space, and background terms. Combined with NeuS and Neuralangelo without changing the backbones or the ray budget, the paper reports finer foreground meshes with fewer holes on object-centric benchmarks and modestly better masked-region image quality, at roughly ten percent more training time. A sympathetic reader would care because the scheme turns the network's own geometry estimate into a sampling policy that needs no masks, depth maps, or point-cloud priors, and because the ablations make the role of each component explicit.

What carries the argument

The carrying object is the three-dimensional image space $U$ of each camera, with coordinates $u = [u, v, \lambda]$ (normalized pixel coordinates plus depth), discretized as per-camera grids $G_U$. The machinery has four linked pieces: the SDF-derived density $p(x) = \phi_s(S(x))$, a logistic density peaked at the surface's zero-level set; the change-of-variables interpolation that maps it onto $G_U$ as a Riemann sum of $\eta^{-2}p(x)$ over transformed scene-grid cells, so each camera cell carries a probability; the view-dependency step $\tilde{p}(u_i) = p(u_i)e^{-\sum_{k=0}^{i}p(u_k)}$, which treats $p(u)$ as a volume density to weight by transmittance and account for occlusions; and conditional sampling with inverse transform sampling over marginals and conditionals to draw $\tilde{u}$, $\tilde{v}$, $\tilde{\lambda}$ from the grid. The sampled depth $\tilde{\lambda}$ anchors the surface reconstruction losses $L_{\mathrm{Near}} = \sum |S(x_i)|w_i$, $L_{\mathrm{Empty}} = \sum [(S(x_j)-\epsilon)w_j]^2$, and a background term, which together regularize the SDF during training. The machinery carries the argument by turning the model's own geometry estimate into a view-dependent sampling policy that runs without masks, depth maps, or structure-from-motion points.

What would settle it

Take a fixed DTU subset and run a two-by-two comparison: probability-guided sampling on or off, crossed with the surface losses on or off, keeping ray budgets identical. The paper's ablations already show the losses dominate the Chamfer improvement (0.57 mm with losses versus 0.82 mm without, view dependency on), so the decisive cell is the sampler with no losses: if its Chamfer distance does not beat uniform sampling with no losses on the same subset, the claim that the sampling strategy itself improves 3D reconstruction is falsified. A complementary check is to compare the PDF's sampled depth columns against DTU's structured-light ground-truth depth to quantify the depth bias the authors concede.

Watch

Extended reading notes

Core claim

The paper's central claim is that the SDF output of a neural implicit surface renderer is itself enough to define a good training-ray distribution. Writing $p(x) = \phi_s(S(x))$ with $\phi_s$ the logistic density, the method transports this density into a per-camera orthographic image space $U$ with coordinates $u = [u, v, \lambda]$ (normalized image coordinates and depth), using the change of variables $p(f^{-1}(v)) = \eta^{-2} p(x)$ discretized over scene and camera grids and accumulated as a Riemann sum. Because this density ignores occlusions, the paper treats $p(u)$ as a volume density and computes transmittance along each depth column, defining the view-dependent PDF $\tilde{p}(u_i) = p(u_i)\exp(-\sum_{k=0}^{i} p(u_k))$, and it flags that identification as a naive approximation that biases depth estimates, as the cited literature notes. Conditional inverse-transform sampling from $\tilde{p}$ yields rays concentrated on surfaces as seen from each camera, mixed with a growing fraction of uniform rays for the background, and the sampled depth $\tilde{\lambda}$ builds surface losses that pull the SDF to zero near the surface, push empty space toward a small constant, and fade out on background rays. Integrated into NeuS and Neuralangelo with fixed backbones and ray budgets, the scheme reports improved DTU reconstructions, better masked-region rendering, qualitatively improved BMVS meshes, and the best mean Chamfer distance the paper's tables report when applied to Neuralangelo, alongside a small reported PSNR decrease on large-scale Tanks and Temples sequences.

Load-bearing premise

The load-bearing premise is that the SDF-derived sampling probability $p(u)$ can stand in for a physical volume density when the method computes transmittance for its occlusion-aware weighting; the paper itself notes this identification biases depth estimates, and if it fails the view-dependent PDF loses its justification, leaving the added surface losses as the likely source of the reported gains.

Editorial extensions

If this is right

  • On the DTU benchmark the combined method improves mean PSNR over the backbones (0.04 dB for NeuS and 0.57 dB for Neuralangelo unmasked, more in masked regions) and lowers Neuralangelo's mean Chamfer distance to 0.60 mm, the best reconstruction average the paper's tables report.
  • The sampler and losses transfer across backbones without changing the ray budget or the backbones' default settings, adding roughly ten percent training time; evaluation overhead is reported as negligible.
  • On the large-scale Tanks and Temples sequences the paper reports a small F1 gain (0.58 to 0.59) alongside a small PSNR drop (25.99 to 25.65 dB), so the benefit concentrates in object-centric scenes.
  • The ablations attribute the reconstruction gains mainly to the surface losses and the view-dependency term: removing the losses raises Chamfer distance from 0.57 to 0.82 mm on the tested DTU subset, and removing view dependency raises it to 1.04 mm.
  • Unfiltered SDF meshes show fewer spurious boundary surfaces and filled holes when the sampler is active, because rays aimed at surfaces keep the model from inventing geometry in empty space.

Reading between the lines

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

  • Because the ablations never run the sampler-off, losses-off cell, a clean two-by-two attribution of the gains between the sampler and the losses remains untested; the paper's own numbers suggest the losses carry most of the Chamfer improvement, while view dependency carries most of the sampler's own contribution.
  • The reported PSNR drop on large-scale Tanks and Temples sequences suggests the SDF-peaked density may starve background content that wide scenes need; making the uniform-ray fraction a learned or scene-adaptive schedule instead of a fixed 20-to-80 percent ramp would test this directly.
  • The interpolation step is effectively a fixed normalizing flow from scene space to each camera's image space, a connection the paper itself notes; learning the partition factor or the logistic scale during training could let the sampler tighten as the surface sharpens, since the ablation shows a fixed small scale produces high-variance samples.
  • Because the sampled depth comes free from the same SDF, the scheme pairs naturally with depth-prior pipelines: feeding the sampled depth estimate as a soft prior into a depth-guided renderer would test whether the conceded depth bias acts as a useful regularizer rather than a defect.
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 / 5 minor

Summary. The paper proposes a training-time, backbone-agnostic sampling and loss framework for neural implicit surface rendering. It models a probability density over a 3D orthographic image space from the SDF of the scene, samples image coordinates and depths conditionally, and adds surface reconstruction losses (near-surface, empty-space, background) that use the sampled depth. The method is tested as a plug-in for NeuS and Neuralangelo on DTU, BMVS, and Tanks and Temples, reporting improved PSNR, Chamfer distance, and F1 over the backbones, plus an ablation study and computational-overhead measurements. The central claim is that combining the proposed sampler with the proposed loss yields more accurate 3D reconstructions and improved rendering, especially in regions of interest.

Significance. Should the attribution gap be closed, the contribution would be practically valuable: it is a training-time add-on that needs no extra supervision or pretrained models, works with two strong backbone implementations, and adds only modest overhead (about 10% training time and less than 1% evaluation time in the reported configuration). The empirical scope is a strength: three datasets, quantitative and qualitative comparisons, and an explicit ablation table. The main reservation is that the current experiments do not measure the sampler's marginal contribution, and the supplementary material explicitly concedes that the sampling scheme alone does not substantially improve 3D reconstruction; as a result, Contribution 1 and the title's emphasis on the sampler are not yet supported.

major comments (3)
  1. [Sec. 5.2, Table 3, App. B] The experiments do not isolate the probability-guided sampler from the proposed surface losses. The only conditions are A1 (sampler on, L=0) and A2 (sampler on, L=500), so the Chamfer improvement from 0.82 to 0.57 is attributed to LSurf, not to the sampler; there is no condition with LSurf enabled and the default (uniform) pixel and ray sampling. The supplementary text (Appendix B, paragraph 'Different Weights for Surface Reconstruction Loss') explicitly states that 'Modifying only the sampling scheme, i.e., guided sampling, without considering additional 3D information during training, does not substantially improve the 3D reconstruction results.' This statement, together with Table 3, directly undermines Contribution 1 and the title's central claim. Please add ablations that turn the sampler on and off while keeping the loss fixed (including the full method), and separately report the marginal PSNR and Chamfer/F1 attributable to each component.
  2. [Sec. 4.2, Eq. (7)] Equation (7) is not algebraically consistent. If v=f(x), the change-of-variables formula reads p_V(v)=p_X(f^{-1}(v)) |det ∂f^{-1}(v)/∂v|, which the supplementary material derives as p_V(v)=η^{-2} p_X(f^{-1}(v)). The main text instead writes p(x)=p(f^{-1}(v)) |...| and then concludes p(f^{-1}(v))=η^{-2} p(x), placing p(x) on both sides of the initial equality. Please correct this typo and clarify which PDF is being transformed, since the interpolation and the sampling probabilities in Secs. 4.2-4.4 rest on this formula.
  3. [Sec. 4.3, Eq. (8)] The view-dependent probability ẽp is computed by setting the volume density per cell equal to the sampling density p(u) and using it in a Beer-Lambert transmittance. This is an acknowledged ad hoc identification—the authors note the depth bias from [50,65]—not a consequence of the model. Since the ablation A2 vs A3 shows that view dependency has a large effect (Chamfer 0.57 vs 1.04), the reader cannot tell whether the gain comes from a principled occlusion model or from this particular heuristic. Please either justify this choice through a derivation or validate it by comparing alternative occlusion weightings (e.g., using the backbone's actual volume density, or no weighting) while holding the sampler and the losses fixed.
minor comments (5)
  1. [Sec. 4.5, Eq. (14)] The empty-space loss minimizes (S(x_j)-epsilon)^2 w_j^2, which drives SDF values toward epsilon in the complement of the near-surface set; because w_j is itself transmittance-weighted, the interaction between this loss and the backbone's Eikonal regularization should be discussed, and the value of epsilon should be stated.
  2. [Table 1] In the Chamfer (mm) row for NeuS+Ours, several numerical entries appear garbled in the typeset table (e.g., '2.1', '0.6', '1.2'), which prevents the reader from verifying the reported mean of 1.2; please regenerate the table.
  3. [Sec. 4.2 and Sec. 4.4] The construction of p(u) is not explicitly normalized, yet inverse-transform sampling in Sec. 4.4 requires a normalized cumulative distribution; please state whether the grid weights are normalized before the marginals and conditionals are computed and how boundary cells are handled.
  4. [Sec. 5.1] The paper says 'we sample 32 points around the sampled depth' while also reporting 'the same number of rays and ray points as the baselines'; please clarify how the near-surface Gaussian samples coexist with the fixed per-ray point budget.
  5. [Tables 1 and 2] Quantitative results are reported as point estimates from a single run; for the small reported gains (e.g., DTU Chamfer 0.61 vs 0.60 and PSNR 33.84 vs 34.41), reporting variance or multiple seeds would substantially increase confidence.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the method is validated against external benchmarks with honest ablations; the two self-citations are not load-bearing.

full rationale

The paper's central comparison is external: it plugs the proposed probability-guided sampler and surface losses into NeuS and Neuralangelo and measures PSNR and Chamfer distance against fixed baselines on standard datasets (DTU, TNT, BMVS), so no fitted parameter is renamed as a prediction and no reported number is an identity consequence of the method's definitions. The two apparent self-references are not circular reductions: (i) the sampling distribution p(x) = phi_s(S(x)) (Eq. 6) is built from the model's own SDF, and Eq. 8 treats p(u) as volume density for occlusion weighting, but this is an explicitly flagged heuristic assumption, not an equivalence imposed by construction, and the evaluation metric (Chamfer to ground-truth scans) is independent of it; (ii) the near-surface loss LNear (Eq. 13) weights SDF values by volume densities, which is the standard SDF self-supervision pattern common to NeuS-style renderers, and its success is judged against ground truth, not against its own definition. The self-citations ([11], [24]) are related-work mentions that carry no load-bearing argument. Per the reviewing rule, the limitation passages were weighed explicitly: Appendix B states 'Modifying only the sampling scheme, i.e., guided sampling, without considering additional 3D information during training, does not substantially improve the 3D reconstruction results,' and Table 3 (A1 vs A2) shows the Chamfer gain is driven by the surface loss rather than the sampler alone. That is an honest attribution gap relevant to the paper's headline emphasis, and it weakens the claim that the sampler itself improves 3D reconstruction, but it is not circularity, because there is no condition in which a result is forced by its own definition or by a self-citation chain, and no external benchmark is constructed from the method's own outputs. Verdict: no significant circularity; score reflects only the minor, non-load-bearing self-citations.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on an adaptive sampling loop driven by the model's own SDF and several design assumptions. The most load-bearing is treating the SDF-derived probability as volume density for occlusion weighting; the paper acknowledges this causes depth bias. Loss weights and grid resolutions are hand-tuned.

free parameters (7)
  • Lambda 1 and Lambda 2 (surface loss weights) = 0.5 each
    Set in Sec. 5.1; control balance of near-surface vs empty/background losses.
  • LSurf weight in total training loss = 500
    Set in Sec. 5.1; ablation shows Chamfer improves from 0.82 to 0.57 when weight is raised from 0 to 500.
  • Partition factor F = 2
    Chosen in supp. B as accuracy/compute tradeoff; F=4 looks similar but costs 8x update time.
  • Uniform ray sampling fraction schedule = 20%, 40%, 60%, 80%
    Sec. 5.1; fraction of rays sampled uniformly increases over training.
  • Number of near-surface samples per ray = 32
    Sec. 5.1; additional Gaussian samples around sampled depth.
  • Grid resolutions (GX and GU) = 128^3 and 64x64x128
    Sec. 5.1; resolution of scene and image-space grids.
  • Epsilon and beta in losses = unspecified
    Eqs. 14 and 15 use epsilon and beta but values are not given in main text or supplement.
assumptions (5)
  • standard math The camera projection f_c is bijective with Jacobian determinant eta^2.
    Used in Eq. 7 for change of variables; proved in supp. C.1.
  • domain assumption The SDF value at the current training state is a reliable indicator of surface location for guiding sampling.
    Entire sampler builds p(x) from phi_s(S(x)); early in training SDF is noisy.
  • ad hoc to paper Volume density per cell can be set equal to the sampling probability p(u) for transmittance computation.
    Sec. 4.3 Eq. 8; authors note this naive density causes depth bias but keeps occlusion awareness.
  • standard math Riemann sum over partitioned cells approximates the continuous change of variables for p(u).
    Sec. 4.2 and supp. C.2; requires F to be large enough, yet F=2 is used.
  • domain assumption The logistic distribution conversion phi_s(S(x)) is a suitable proposal density for sampling.
    Sec. 4.1 Eq. 6; concentrates samples near the zero-level set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Probability-guided Sampler for Neural Implicit Surface Rendering." pith.science (2026). https://pith.science/paper/7UE46XDI

@misc{pith2026250608619,
  author       = {Pith},
  title        = {Pith review of: A Probability-guided Sampler for Neural Implicit Surface Rendering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7UE46XDI}},
  note         = {Machine review of arXiv:2506.08619}
}
read the original abstract

Several variants of Neural Radiance Fields (NeRFs) have significantly improved the accuracy of synthesized images and surface reconstruction of 3D scenes/objects. In all of these methods, a key characteristic is that none can train the neural network with every possible input data, specifically, every pixel and potential 3D point along the projection rays due to scalability issues. While vanilla NeRFs uniformly sample both the image pixels and 3D points along the projection rays, some variants focus only on guiding the sampling of the 3D points along the projection rays. In this paper, we leverage the implicit surface representation of the foreground scene and model a probability density function in a 3D image projection space to achieve a more targeted sampling of the rays toward regions of interest, resulting in improved rendering. Additionally, a new surface reconstruction loss is proposed for improved performance. This new loss fully explores the proposed 3D image projection space model and incorporates near-to-surface and empty space components. By integrating our novel sampling strategy and novel loss into current state-of-the-art neural implicit surface renderers, we achieve more accurate and detailed 3D reconstructions and improved image rendering, especially for the regions of interest in any given scene.

Figures

Figures reproduced from arXiv: 2506.08619 by the authors.

Figure 1
Figure 1. Schematic of our pipeline and an example: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Proposed guided-sampling: The scene is represented as a 3D grid GX , and characterized by a PDF p(x) computed from the SDF network and modeled by a logistic distribution of the SDF values ϕs(S(x)). We propose to use a 3D image space that includes depth, represented as GU , where one can define p(u) based on p(x) as described in Interpolation – Sec. 4.2. Then, we consider the camera viewpoint of the scene (such as oc… view at source ↗
Figure 3
Figure 3. Bird’s-eye view illustration of the Rie [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: DTU scan 37: When our sampling and surface reconstruction losses are in￾cluded in NeuS and Neuralangelo backbones, we get a sharper 3D reconstruction of the foreground objects. Our approach also removes the hole obtained by Neuralangelo. (a) Neuralangelo (b) Neuralange…
Figure 5
Figure 5. Figure 5: DTU scans 118 and 114: We can extract more detailed meshes with our sampler and losses. The synthesized normal images are less noisy, with sharper edges in both scans, leading to better reconstruction. Image quality remains similar. and NeuS, respectively. For a fair c…
Figure 6
Figure 6. Figure 6: BMVS sequence "Sphere": We observe that Neuralangelo + Ours get a sig￾nificantly more complete 3D surface reconstruction. Also, our sampling better captures small and intricate regions, such as the small hole highlighted in the object foreground. (a) NeuS (b) NeuS + Ou…
Figure 7
Figure 7. Figure 7: BMVS sequence "Bandstand": Our sampling and surface reconstruction losses have clear advantages. While NeuS fails to capture the foreground surfaces, the proposed sampling obtains a reasonable 3D structure and high-quality images. with object masks, when available, to …
Figure 8
Figure 8. Figure 8: Reconstruction Error and Variance over training. Our sampling converges faster to a better 3D representation (solid line), while significantly reducing the variance of the logistic output density (dashed) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 65 canonical work pages

  1. [1]

    In: IEEE/CVF Conf

    Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O’Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 16610–16620 (2023)

    work page 2023

  2. [2]

    In: IEEE/CVF Conf

    Azinović, D., Martin-Brualla, R., Goldman, D.B., Nießner, M., Thies, J.: Neural rgb-d surface reconstruction. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 6290–6301 (2022)

    work page 2022

  3. [3]

    In: IEEE/CVF Int’l Conf

    Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srini- vasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 5855–5864 (2021)

    work page 2021

  4. [4]

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360:Unboundedanti-aliasedneuralradiancefields.In:IEEE/CVFConf.Computer Vision and Pattern Recognition (CVPR). pp. 5470–5479 (2022)

    work page 2022

  5. [5]

    IEEE/CVF Conf

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR) (2023)

    work page 2023

  6. [6]

    IEEE Trans

    Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J.M., Tardós, J.D.: Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam. IEEE Trans. Robotics (T-RO)37(6), 1874–1890 (2021)

    work page 2021

  7. [7]

    In: IEEE/CVF Conf

    Darmon, F., Bascle, B., Devaux, J.C., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 6260–6269 (2022)

    work page 2022

  8. [8]

    In: IEEE Int’l Conf

    Davison: Real-time simultaneous localisation and mapping with a single camera. In: IEEE Int’l Conf. Computer Vision (ICCV). pp. 1403–1410 (2003)

    work page 2003

Show all 66 references
  1. [9]

    In: IEEE/CVF Int’l Conf

    Deng, J., Wu, Q., Chen, X., Xia, S., Sun, Z., Liu, G., Yu, W., Pei, L.: Nerf- loam: Neural implicit representation for large-scale incremental lidar odometry and mapping. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 8218–8227 (2023)

    2023

  2. [10]

    Advances in Neural Information Processing Systems (NeurIPS)35, 3403–3416 (2022)

    Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems (NeurIPS)35, 3403–3416 (2022)

    2022

  3. [11]

    In: Int’l Conf

    Gaur, A., Pais, G.D., Miraldo, P.: Oriented-grid encoder for 3d implicit represen- tations. In: Int’l Conf. 3D Vision (3DV) (2024)

    2024

  4. [12]

    In: European Conf

    Geppert, M., Larsson, V., Speciale, P., Schönberger, J.L., Pollefeys, M.: Privacy preserving structure-from-motion. In: European Conf. Computer Vision (ECCV). pp. 333–350 (2020)

    2020

  5. [13]

    Cam- bridge University Press, 2 edn

    Hartley, R.I., Zisserman, A.: Multiple View Geometry in Computer Vision. Cam- bridge University Press, 2 edn. (2004)

    2004

  6. [14]

    In: IEEE/CVF Int’l Conf

    Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 19774–19783 (2023) 16 G. D. Pais, V. Piedade, M. Chatterjee, M. Greiff, and ...

    2023

  7. [15]

    In: IEEE Conf

    Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanæs, H.: Large scale multi-view stereopsis evaluation. In: IEEE Conf. Computer Vision and Pattern Recognition (CVPR). pp. 406–413 (2014)

    2014

  8. [16]

    In: IEEE/CVF Conf

    Johnson, E., Habermann, M., Shimada, S., Golyanik, V., Theobalt, C.: Unbiased 4d: Monocular 4d reconstruction with a neural deformation model. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 6597–6606 (2023)

    2023

  9. [17]

    ACM Transactions on Graphics (TOG)36(4) (2017)

    Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmark- ing large-scale scene reconstruction. ACM Transactions on Graphics (TOG)36(4) (2017)

    2017

  10. [18]

    In: IEEE/CVF Conf

    Kong, X., Liu, S., Taher, M., Davison, A.J.: vmap: Vectorised object mapping for neural field slam. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 952–961 (2023)

    2023

  11. [19]

    Kurz, A., Neff, T., Lv, Z., Zollhöfer, M., Steinberger, M.: Adanerf: Adaptive sam- plingforreal-timerenderingofneuralradiancefields.In:EuropeanConf.Computer Vision (ECCV). pp. 254–270 (2022)

    2022

  12. [20]

    In: IEEE/CVF Conf

    Li, T., Slavcheva, M., Zollhöfer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3d video synthesis from multi-view video. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 5521–5531 (2022)

    2022

  13. [21]

    In: IEEE/CVF Conf

    Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 8456–8465 (2023)

    2023

  14. [22]

    In: IEEE/CVF Conf

    Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: IEEE/CVF Conf. Computer Vision and Pat- tern Recognition (CVPR). pp. 6498–6508 (2021)

    2021

  15. [23]

    Advances in Neural Information Processing Systems (NeurIPS)33, 15651–15663 (2020)

    Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems (NeurIPS)33, 15651–15663 (2020)

    2020

  16. [24]

    In: IEEE/CVF Conf

    Liu, X., Tai, Y.w., Tang, C.K., Miraldo, P., Lohit, S., Chatterjee, M.: Gear-nerf: Free-viewpoint rendering and tracking with motion-aware spatio-temporal sam- pling. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR) (2024)

    2024

  17. [25]

    In: IEEE/CVF Conf

    Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy net- works: Learning 3d reconstruction in function space. In: IEEE/CVF Conf. Com- puter Vision and Pattern Recognition (CVPR). pp. 4460–4470 (2019)

    2019

  18. [26]

    In: Eu- ropean Conf

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Eu- ropean Conf. Computer Vision (ECCV) (2020)

    2020

  19. [27]

    Journal of Geophysical Research: Solid Earth100(B7), 12431–12447 (1995)

    Mosegaard, K., Tarantola, A.: Monte carlo sampling of solutions to inverse prob- lems. Journal of Geophysical Research: Solid Earth100(B7), 12431–12447 (1995)

    1995

  20. [28]

    ACM Transactions on Graphics (TOG)41(4), 1– 15 (2022)

    Müller,T.,Evans,A.,Schied,C.,Keller,A.:Instantneuralgraphicsprimitiveswith a multiresolution hash encoding. ACM Transactions on Graphics (TOG)41(4), 1– 15 (2022)

    2022

  21. [29]

    In: IEEE/CVF Conf

    Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumet- ric rendering: Learning implicit 3d representations without 3d supervision. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR) (2020)

    2020

  22. [30]

    In: IEEE/CVF Int’l Conf

    Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 5589–5599 (2021) A Probability-guided Sampler for Neural Implicit Surface Rendering 17

    2021

  23. [31]

    In: IEEE/CVF Int’l Conf

    Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin- Brualla, R.: Nerfies: Deformable neural radiance fields. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 5865–5874 (2021)

    2021

  24. [32]

    In: European Conf

    Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: European Conf. Computer Vision (ECCV). pp. 523–540 (2020)

    2020

  25. [33]

    MIT Press (2023)

    Pharr, M., Jakob, W., Humphreys, G.: Physically based rendering: From theory to implementation. MIT Press (2023)

    2023

  26. [34]

    In: IEEE/CVF Conf

    Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radi- ance fields for dynamic scenes. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 10318–10327 (2021)

    2021

  27. [35]

    In: IEEE/CVF Conf

    Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 84–93 (2020)

    2020

  28. [36]

    In: IEEE/CVF Int’l Conf

    Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 18433– 18444 (2023)

    2023

  29. [37]

    In: IEEE Conf

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: IEEE Conf. Computer Vision and Pattern Recognition (CVPR) (2016)

    2016

  30. [38]

    In: European Conf

    Schönberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: European Conf. Computer Vision (ECCV). pp. 501–518 (2016)

    2016

  31. [39]

    In: European Conf

    Shen, J., Agudo, A., Moreno-Noguer, F., Ruiz, A.: Conditional-flow nerf: Accurate 3d modelling with reliable uncertainty quantification. In: European Conf. Com- puter Vision (ECCV). pp. 540–557 (2022)

    2022

  32. [40]

    Statistics & Probability Letters11(1), 69–70 (1991)

    Stefanski, L.A.: A normal scale mixture representation of the logistic distribution. Statistics & Probability Letters11(1), 69–70 (1991)

    1991

  33. [41]

    In: IEEE/CVF Int’l Conf

    Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and positioning in real-time. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 6229–6238 (2021)

    2021

  34. [42]

    In: IEEE/CVF Conf

    Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast conver- gence for radiance fields reconstruction. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 5459–5469 (2022)

    2022

  35. [43]

    In: ACM SIGGRAPH (2022)

    Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3d reconstruction in the wild. In: ACM SIGGRAPH (2022)

    2022

  36. [44]

    In: IEEE/CVF Conf

    Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 15598–15607 (2021)

    2021

  37. [45]

    Computers & Graphics 118, 48–59 (2024)

    Sun, S., Liu, M., Fan, Z., Jiao, Q., Liu, Y., Dong, L., Kong, L.: Efficient ray sampling for radiance fields reconstruction. Computers & Graphics 118, 48–59 (2024)

    2024

  38. [46]

    In: IEEE/CVF Conf

    Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view syn- thesis. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 8248–8258 (2022)

    2022

  39. [47]

    In: IEEE/CVF Int’l Conf

    Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dy- namic scene from monocular video. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 12959–12970 (2021) 18 G. ...

    2021

  40. [48]

    In: IEEE/CVF Conf

    Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 12922–12931 (2022)

    2022

  41. [49]

    In: European Conf

    Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: European Conf. Computer Vision (ECCV). pp. 139–155 (2022)

    2022

  42. [50]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2021)

    Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)

    2021

  43. [51]

    In: IEEE/CVF Int’l Conf

    Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 3295–3306 (2023)

    2023

  44. [52]

    Advances in Neural Information Processing Systems (NeurIPS) 35, 1966–1978 (2022)

    Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems (NeurIPS) 35, 1966–1978 (2022)

    2022

  45. [53]

    In: European Conf

    Wei, X., Zhang, Y., Li, Z., Fu, Y., Xue, X.: Deepsfm: Structure from motion via deep bundle adjustment. In: European Conf. Computer Vision (ECCV). pp. 230– 247 (2020)

    2020

  46. [54]

    In: IEEE/CVF Conf

    Wen, B., Tremblay, J., Blukis, V., Tyree, S., Müller, T., Evans, A., Fox, D., Kautz, J., Birchfield, S.: Bundlesdf: Neural 6-dof tracking and 3d reconstruction of un- known objects. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 606–617 (2023)

    2023

  47. [55]

    In: ACM SIGGRAPH Asia

    Wu, J., Liu, L., Tan, Y., Jia, Q., Zhang, H., Zhang, X.: Actray: Online active ray sampling for radiance fields. In: ACM SIGGRAPH Asia. pp. 1–10 (2023)

    2023

  48. [56]

    In: IEEE/CVF Conf

    Yao, Y., Luo, Z., Li, S., Zhang, J., Ren, Y., Zhou, L., Fang, T., Quan, L.: Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 1790– 1799 (2020)

    2020

  49. [57]

    Advances in Neural Information Processing Systems (NeurIPS)34, 4805– 4815 (2021)

    Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems (NeurIPS)34, 4805– 4815 (2021)

    2021

  50. [58]

    In: ACM SIGGRAPH (2023)

    Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. In: ACM SIGGRAPH (2023)

    2023

  51. [59]

    Advances in Neural Information Processing Systems (NeurIPS)33, 2492– 2502 (2020)

    Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appear- ance. Advances in Neural Information Processing Systems (NeurIPS)33, 2492– 2502 (2020)

    2020

  52. [60]

    Advances in Neural Information Processing Systems (NeurIPS)35, 25018–25032 (2022)

    Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monoc- ular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS)35, 25018–25032 (2022)

    2022

  53. [61]

    In: IEEE/CVF Conf

    Zhang, J., Yao, Y., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Critical regularizations for neural surface reconstruction in the wild. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 6270–6279 (2022)

    2022

  54. [62]

    In: IEEE/CVF Int’l Conf

    Zhang, J., Yao, Y., Quan, L.: Learning signed distance field for multi-view surface reconstruction. In: IEEE/CVF Int’l Conf. Computer Vision (ICCV). pp. 6525–6534 (2021)

    2021

  55. [63]

    arXiv:2010.07492 (2020) A Probability-guided Sampler for Neural Implicit Surface Rendering 19

    Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020) A Probability-guided Sampler for Neural Implicit Surface Rendering 19

    2020 arXiv

  56. [64]

    In: IEEE/CVF Conf

    Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 5449–5458 (2022)

    2022

  57. [65]

    In: IEEE/CVF Conf

    Zhang, Y., Hu, Z., Wu, H., Zhao, M., Li, L., Zou, Z., Fan, C.: Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 4359–4368 (2023)

    2023

  58. [66]

    In: IEEE/CVF Conf

    Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). pp. 12786–12796 (2022) A Probability-guided Sampler for Neural Impli...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.