Pith. sign in

REVIEW 4 major objections 4 minor 65 references

A View-consistent Sampling Method for Regularized Training of Neural Radiance Fields

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

Pith's one-line read This paper claims that importance-sampling NeRF rays according to a multi-view feature-agreement distribution, plus a depth-pushing loss, regularizes training and improves novel-view synthesis without any depth supervision.

desk verdict The view-consistency sampling idea is new and sensible, and the reported gains are consistent enough to take seriously; the paper just needs error bars, a corrected ablation sentence, and a fairer depth baseline. read the letter →

arxiv 2507.04408 v1 pith:YXYGEGWY submitted 2025-07-06 cs.CV

classification cs.CV
keywords neuralradiancefieldsview-consistentsamplingimplicitregularizationnovelviewsynthesisfeaturedistillationDINOv2featuresdepth-freesparse-viewreconstruction
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 tries to establish that a NeRF can be regularized for real-world, unbounded scenes by sampling along rays according to a view-consistency distribution rather than relying on predicted depth. The distribution is built from low-level color and distilled high-level DINOv2 features at projected 2D locations, and importance sampling from it concentrates training samples near points that look the same across views, which are assumed to be surface points. If true, this gives a depth-free regularizer that improves novel-view synthesis, especially with few views, and avoids the generalization failures of monocular depth estimators. On two benchmarks the method improves over the Nerfacto base and matches or beats depth-based regularizers, with 21.40 versus 19.05 PSNR on Mip-NeRF360 and 19.45 versus 18.29 on Tanks&Temples.

What carries the argument

The view-consistency metric $s_i$ of Eq. (4): the average over views $j \in \mathcal{V}_i$ of an indicator that both the normalized color similarity and the normalized distilled-feature similarity exceed a threshold $\delta = 0.4$. It is computed at $M$ uniform pre-samples along each ray and then fed to a PDF importance sampler from Nerfstudio, which produces the $S$ samples used in volume rendering. The distilled features come from a DINOv2 model projected from 384 dimensions to 32 by a ResNet bottleneck block trained on MegaDepth with a CLIP-style symmetric cross-entropy loss. A depth-pushing loss $\mathcal{L}_{\mathrm{depu}} = -\log(d(r) + \varepsilon)$, where $d(r)$ is expected depth along the ray, penalizes nearby density and prevents background collapse. Together these focus training samples near surfaces without any depth map.

What would settle it

Use a dataset with ground-truth depth and heavy occlusion, compute the view-consistency distribution along each ray, and record whether the argmax of $s_i$ lies within a small distance of the true surface depth for most rays; if the hit rate is close to the fraction of the ray interval occupied by the surface, the metric carries no surface signal and the regularization claim collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that NeRF training can be regularized without depth supervision by replacing naive ray sampling with importance sampling from a per-ray view-consistency distribution. The distribution scores each pre-sampled 3D point by the fraction of views in which its projected color feature and distilled DINOv2 feature both agree with the reference ray's feature beyond a threshold. Because geometrically consistent points are presumed to lie on true surfaces, concentrating samples there imposes an implicit geometric prior; a supplementary depth-pushing loss discourages background collapse. On Mip-NeRF360 and Tanks&Temples, with 10 to 110 training views, VS-NeRF outperforms Nerfacto and both monocular and multi-view depth regularizers, e.g., 21.40 versus 19.05 PSNR on Mip-NeRF360 and 19.45 versus 18.29 on Tanks&Temples.

Load-bearing premise

The load-bearing premise is that points on the true surface, despite occlusions and feature noise, score higher on the view-consistency metric than other points along the ray; if occlusions or repeated textures break that ordering, the sampler concentrates training away from real surfaces.

Editorial extensions

If this is right

  • NeRF training no longer requires a pretrained depth estimator or depth supervision; the prior comes from the input images' own multi-view feature agreement.
  • The method improves performance most when views are sparse, and Table 1 and Fig. 5 show the gap over Nerfacto and over multi-view depth grows as training views decrease.
  • Adding view-consistent sampling to the slower Mip-NeRF360 also helps (22.54 to 23.37 PSNR), so the regularizer is not tied to the Nerfacto implementation.
  • The depth-pushing loss, used alone, already helps (18.88 to 19.10 PSNR in the ablation), and combining it with view-consistent sampling gives the best result, so the two regularizers address complementary failure modes.
  • The paper suggests the same regularization could be ported to 3D Gaussian Splatting by replacing point sampling with Gaussian placement, though it does not demonstrate this.

Reading between the lines

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

  • Because the view-consistency metric only sees views into which a point projects, scenes with heavy occlusion or few overlapping views will give small sets $\mathcal{V}_i$; the paper does not quantify how the metric degrades, so the method's advantage may shrink precisely where depth priors also fail.
  • Activating the sampling only in the first 5000 iterations suggests the main effect is shaping early geometry; one could test whether a curriculum that re-enables sampling after coarse geometry forms improves fine detail further.
  • The 32-dimension distilled feature is chosen by ablating 16 and 64, and the 64-dimension version actually scores slightly higher on PSNR in the ablation (21.65 versus 21.57) while taking longer; larger distilled features may be preferable when compute allows.
  • The view-consistency distribution can be read as a learned multi-view matching score, so swapping DINOv2 for other correspondence-friendly features, which the paper notes as an alternative, could yield different trade-offs for different scene types.
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 / 4 minor

Summary. The manuscript introduces VS-NeRF, a regularization scheme for NeRF training. Instead of using fixed depth estimates, it precomputes a view-consistency score for points along each camera ray from a combination of low-level color similarity and high-level features distilled from DINOv2, and then uses this score as an importance-sampling distribution for the volume rendering integral. A depth-pushing loss is added to discourage background collapse. Experiments on Mip-NeRF360 and Tanks&Temples show consistent PSNR/SSIM/LPIPS improvements over Nerfacto and other depth-based regularizers, with ablations attributing gains to both components.

Significance. The method is a practical, depth-free regularizer that does not require external depth supervision at inference time, which is a useful property for unbounded outdoor scenes. The work is grounded in a reproducible implementation on Nerfstudio, and the comparisons against monocular and multi-view depth baselines are valuable. The main claims depend on the view-consistency metric concentrating on true surfaces, and the paper would be strengthened by a direct validation of that premise. The reported gains are meaningful at the aggregate level, but the absence of variance estimates and an internal contradiction in the feature-dimension ablation currently prevent full confidence.

major comments (4)
  1. [Sec. 4.2, Table 2] The text states that increasing the distilled feature dimension to 64 degrades performance, but row G of Table 2 reports 21.65 PSNR with 64-dimensional features, which is higher than the complete model's 21.57 PSNR with 32-dimensional features. This direct contradiction calls the interpretation of the feature-dimension ablation into question and should be resolved or the text corrected before the paper can be accepted.
  2. [Sec. 3.3.1, Eq. (4)] The central premise that the view-consistency metric of Eq. (4) is concentrated at true surface points is asserted ('statistically the score is still prominent for surface points') but never quantified. The paper should provide a diagnostic experiment: for scenes with known or estimated depth, evaluate the metric at ground-truth surface points versus off-surface points, report precision/recall or the distribution overlap, and show how the metric behaves under occlusion (i.e., for points with small |Vi|). Without such evidence, the link between the proposed sampling distribution and surface-guided regularization remains unverified, and the improvement from VS alone in Table 2 could in principle arise from the distribution shape acting as a generic regularizer rather than from surface alignment.
  3. [Sec. 4.1, Table 1] The abstract and Sec. 4.1 claim 'significantly better' results, but all numbers in Tables 1 and 2 come from single runs with no standard deviation or significance test. On Tanks&Temples, the margin over +Multi-view Depth is only 0.17 dB (19.45 vs 19.28), so it is important to report variance across at least three seeds and/or a paired test across the eight scenes to support the significance claim.
  4. [Sec. 4, implementation details] The hyperparameters δ=0.4, λ_depu=0.0001, ε=0.01, and the choice to activate view-consistent sampling only in the first 5000 of 30000 iterations are fixed without any sensitivity analysis. Since these values control the shape of the sampling distribution and its interaction with the depth-pushing loss, the paper should include a sensitivity study over δ and the activation window (or at least over δ) to show that the reported improvements are not confined to a single hand-tuned configuration.
minor comments (4)
  1. [Sec. 4.1] There is an inconsistency in the naming of the depth baselines: the Baselines paragraph identifies ZoeDepth as the monocular method and MVSFormer++ as the multi-view method, but Sec. 4.1 says 'a monocular one using MVSformer++ and a multi-view one using ZoeDepth', which is reversed.
  2. [Sec. 3.3.2] There are typos: 'genratese' should be 'generates', and the notation (ti)pre / (ti) is not defined carefully; please define the pre-sample points and the final samples explicitly.
  3. [Sec. 4.1 and Sec. 4.2] There are several typographical errors: 'agains' should be 'against' in Sec. 4.1, 'ie last row' should be 'the last row' in Sec. 4.2, and 'B ONSAI' in the Fig. 3 caption has a stray space.
  4. [Sec. 4] The paper states that the Nerfacto proposal network sampling scheme 'is also left unchanged' while also claiming to replace Nerfacto's sampling scheme; clarify which stages are kept and which are replaced, since this is important for understanding what exactly is being ablated.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the view-consistency regularizer is an empirical proposal evaluated on held-out views, not a quantity defined by its own prediction.

full rationale

The paper's central claim is an empirical improvement in held-out novel-view synthesis, measured on Mip-NeRF360 and Tanks&Temples against baselines (Table 1). Nothing in the derivation defines the reported PSNR in terms of the proposed regularization. The view-consistency metric in Eq. (4) is an input feature statistic (thresholded agreement of color and distilled DINOv2 features over visible views), and the adaptive sampler in Sec. 3.3.2 uses that statistic as a proposal distribution; the 'implicit regularization' is the sampling scheme itself, not a separately derived quantity, so there is no equation that reduces to its own inputs. The distilled-feature bottleneck is trained on MegaDepth with ground-truth correspondences, an external dataset, and the NeRF training itself uses no depth supervision. The assumption in Sec. 3.3.1 that 'statistically the score is still prominent for surface points' is an unverified empirical premise, and the threshold delta, depth-loss weight, and 5000-iteration schedule are hand-set partly with knowledge of the benchmark; these are validity and robustness concerns, not circularity. Self-citations ([46], [58]) appear only as contextual mentions of alternative pairwise matchers and do not carry the argument. Hence no circular step is identifiable by the reduction criterion.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The method rests on a learned feature prior (DINOv2 distilled on MegaDepth) transferred to new scenes, an unverified statistical claim about occlusions, an assumed normal distribution for similarity measures, and several hand-set hyperparameters. Beyond the standard volume rendering model, no heavy mathematical axioms are needed, but these domain assumptions are load-bearing.

free parameters (5)
  • similarity threshold δ in Eq. (4) = 0.4
    Set empirically; the paper assumes normalized similarity measures follow a normal distribution and picks δ without a sweep or validation split.
  • depth-pushing loss weight λ_depu = 0.0001
    Chosen by hand to balance L_depu and L_color; no sensitivity analysis is reported.
  • stabilizer ε in Eq. (5) = 0.01
    Small constant to avoid log(0); no sensitivity analysis is reported.
  • distilled feature dimension = 32
    Selected by ablation on Mip-NeRF360 with 50 views; the ablation does not clearly support the choice since 64-dim gives comparable or slightly better metrics.
  • iterations with view-consistent sampling active = 5000 out of 30000
    Empirically chosen; the paper activates the sampling only in the first 5000 of 30000 iterations without ablating this schedule.
assumptions (6)
  • domain assumption DINOv2 features encode geometric information transferable across views.
    Relied on in Sec. 3.2; support cited from [31,55], but the transfer to new scenes is assumed.
  • domain assumption Distilled features trained on MegaDepth generalize to target scenes without adaptation.
    The ResNet bottleneck is trained on MegaDepth pairs and then frozen when applied to Mip-NeRF360 and Tanks&Temples in Sec. 3.2.
  • domain assumption View-consistent points are more likely to lie on true surface points.
    This is the core premise of the sampling prior stated in Sec. 3.3; the paper provides only a visualization as evidence.
  • ad hoc to paper Occlusions do not break the view-consistency metric on average.
    Sec. 3.3.1 states 'although occlusions may hinder the effectiveness of this metric, statistically the score is still prominent for surface points', without quantitative evidence.
  • ad hoc to paper Normalized similarity measures follow a normal distribution.
    Sec. 3.3.1 assumes normality to justify a threshold; no distribution fit is shown.
  • ad hoc to paper Adaptive sampling based on a fixed precomputed distribution is beneficial throughout the first 5000 iterations.
    The view-consistency distribution is computed without updating as NeRF geometry evolves; the method is only active early in training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A View-consistent Sampling Method for Regularized Training of Neural Radiance Fields." pith.science (2026). https://pith.science/paper/YXYGEGWY

@misc{pith2026250704408,
  author       = {Pith},
  title        = {Pith review of: A View-consistent Sampling Method for Regularized Training of Neural Radiance Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YXYGEGWY}},
  note         = {Machine review of arXiv:2507.04408}
}
read the original abstract

Neural Radiance Fields (NeRF) has emerged as a compelling framework for scene representation and 3D recovery. To improve its performance on real-world data, depth regularizations have proven to be the most effective ones. However, depth estimation models not only require expensive 3D supervision in training, but also suffer from generalization issues. As a result, the depth estimations can be erroneous in practice, especially for outdoor unbounded scenes. In this paper, we propose to employ view-consistent distributions instead of fixed depth value estimations to regularize NeRF training. Specifically, the distribution is computed by utilizing both low-level color features and high-level distilled features from foundation models at the projected 2D pixel-locations from per-ray sampled 3D points. By sampling from the view-consistency distributions, an implicit regularization is imposed on the training of NeRF. We also utilize a depth-pushing loss that works in conjunction with the sampling technique to jointly provide effective regularizations for eliminating the failure modes. Extensive experiments conducted on various scenes from public datasets demonstrate that our proposed method can generate significantly better novel view synthesis results than state-of-the-art NeRF variants as well as different depth regularization methods.

Figures

Figures reproduced from arXiv: 2507.04408 by the authors.

Figure 1
Figure 1. View-consistent sampling. Our central idea is to pre-compute a view-consistency distribution along rays and to perform importance sampling according to this distribution. As a result, the sampling will concentrate around surface points instead of random points in the capture volume. implicit regularization and preventing the overfitting prob￾lem [62]. Our contribution is therefore a novel view-consistent sampling te… view at source ↗
Figure 2
Figure 2. Visualization of the feature distillation process. For the two test images from Megadepth dataset [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the effectiveness of the view-consistency metric on the B [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: We show comparisons of VS-NeRF to the main competitors and the corresponding ground truth images from held-out test [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: We show performances of VS-NeRF and competitors with increasing number of views over Mip-NeRF360 dataset and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 54 canonical work pages

  1. [1]

    Barron, B

    J. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin- Brualla, and P. Srinivasan. Mip-Nerf: A Multiscale Repre- sentation for Anti-Aliasing Neural Radiance Fields. InInter- national Conference on Computer Vision, pages 5855–5864,

  2. [2]

    Barron, B

    J. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman. Mip-Nerf 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pat- tern Recognition, 2022. 1, 3, 7

  3. [3]

    Zip-nerf: Anti-aliased grid-based neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 19697–19705, 2023. 2

  4. [4]

    Zoedepth: Zero-shot trans- fer by combining relative and metric depth

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot trans- fer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 7

  5. [5]

    Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo

    Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo. arXiv preprint arXiv:2401.11673, 2024. 7

  6. [6]

    Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124–14133, 2021. 1, 2

  7. [7]

    A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su. Tensorf: Tenso- rial Radiance Fields. In European Conference on Computer Vision, pages 333–350, 2022. 2, 7

  8. [8]

    Depth-supervised nerf: Fewer views and faster train- ing for free

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ra- manan. Depth-supervised nerf: Fewer views and faster train- ing for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882– 12891, 2022. 1, 2, 7

Show all 65 references
  1. [9]

    Roma: Revisiting ro- bust losses for dense feature matching

    Johan Edstedt, Qiyu Sun, Georg B ¨okman, M ˚arten Wadenb¨ack, and Michael Felsberg. Roma: Revisiting ro- bust losses for dense feature matching. arXiv preprint arXiv:2305.15404, 2023. 3

  2. [10]

    Fridovich-Keil, A

    S. Fridovich-Keil, A. Yu, M. Tancik, Q. Chen, B. Recht, and A. Kanazawa. Plenoxels: Radiance fields without neural networks. In Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2

  3. [11]

    Hartley and A

    R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 1

  4. [12]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. In Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 3

  5. [13]

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick. Masked Autoencoders Are Scalable Vision Learners. In Conference on Computer Vision and Pattern Recognition , pages 16000–16009, 2022. 2

  6. [14]

    Sape: Spatially-adaptive progressive encoding for neural optimization

    Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. Sape: Spatially-adaptive progressive encoding for neural optimization. Advances in Neural Infor- mation Processing Systems, 34:8820–8832, 2021. 2

  7. [15]

    Putting nerf on a diet: Semantically consistent few-shot view synthesis

    Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885–5894, 2021. 2

  8. [16]

    Tensoir: Tensorial inverse rendering

    Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Song- fang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165–174, 2023. 2, 8

  9. [17]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 7, 8

  10. [18]

    J. Kerr, C. Kim, K. Goldberg, A. Kanazawa, and M. Tancik. LERF: Language Embedded Radiance Fields. In Interna- tional Conference on Computer Vision, 2023. 2, 8

  11. [19]

    Infonerf: Ray entropy minimization for few-shot neural volume ren- dering

    Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume ren- dering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 12912– 12921, 2022. 1, 2

  12. [20]

    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

  13. [21]

    Kobayashi, E

    S. Kobayashi, E. Matsumoto, and V . Sitzmann. Decompos- ing NeRF for Editing via Feature Field Distillation. In Ad- vances in Neural Information Processing Systems, 2022. 3

  14. [22]

    Nerfacc: Efficient sampling accelerates nerfs

    Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18537–18546, 2023. 7

  15. [23]

    Megadepth: Learning single- view depth prediction from internet photos

    Zhengqi Li and Noah Snavely. Megadepth: Learning single- view depth prediction from internet photos. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 2041–2050, 2018. 4

  16. [24]

    Z. Li, T. M ¨uller, A. Evans, R. Taylor, M. Unberath, M. Liu, and C. Lin. Neuralangelo: High-Fidelity Neural Surface Re- construction. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 8

  17. [25]

    Diffusion hyperfeatures: Searching through time and space for semantic correspondence

    Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holyn- ski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Ad- vances in Neural Information Processing Systems, 36, 2024. 3

  18. [26]

    Nerf in the wild: Neural radiance fields for uncon- strained photo collections

    Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duck- worth. Nerf in the wild: Neural radiance fields for uncon- strained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

  19. [27]

    Mod- ulated periodic activations for generalizable local functional representations

    Ishit Mehta, Micha ¨el Gharbi, Connelly Barnes, Eli Shecht- man, Ravi Ramamoorthi, and Manmohan Chandraker. Mod- ulated periodic activations for generalizable local functional representations. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages ...

  20. [28]

    Ben Mildenhall, S. P. P., M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng. NeRF: Representing Scenes as Neu- ral Radiance Fields for View Synthesis. In European Con- ference on Computer Vision, 2020. 1, 2

  21. [29]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1–15, 2022. 2, 7

  22. [30]

    Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs

    Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition,...

  23. [31]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El- Nouby, R. Howes, P. Huang, H. Xu, V . Sharma, S. Li, W. Galuba, M. Rabbat, M. Assran, N. Ballas, G. Synnaeve, I. Misra, H. Jegou, J. Mairal, P. Labatut, A. Joulin...

  24. [32]

    Prados and O

    E. Prados and O. Faugeras. Shape from Shading: A Well- Posed Problem? In Conference on Computer Vision and Pattern Recognition, 2005. 1

  25. [33]

    Radford, J

    A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learn- ing transferable visual models from natural language super- vision. In International Conference on Machine Learning . PMLR, 2021. 4

  26. [34]

    Dense depth pri- ors for neural radiance fields from sparse input views

    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 Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12892–12901, 2022. 2

  27. [35]

    Sarlin, D

    P.E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich. Superglue: Learning Feature Matching with Graph Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2020. 3

  28. [36]

    Zerorf: Fast sparse view 360deg reconstruction with zero pretrain- ing

    Ruoxi Shi, Xinyue Wei, Cheng Wang, and Hao Su. Zerorf: Fast sparse view 360deg reconstruction with zero pretrain- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 21114–21124,

  29. [37]

    Sitzmann, J

    V . Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wet- zstein. Implicit Neural Representations with Periodic Acti- vation Functions. In Advances in Neural Information Pro- cessing Systems, 2020. 2

  30. [38]

    Vip-nerf: Visibility prior for sparse input neural radiance fields

    Nagabhushan Somraj and Rajiv Soundararajan. Vip-nerf: Visibility prior for sparse input neural radiance fields. In ACM SIGGRAPH 2023 Conference Proceedings , pages 1– 11, 2023. 2

  31. [39]

    Simplenerf: Regularizing sparse input neural radiance fields with simpler solutions

    Nagabhushan Somraj, Adithyan Karanayil, and Rajiv Soundararajan. Simplenerf: Regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia 2023 Conference Papers, pages 1–11, 2023. 2

  32. [40]

    C. Sun, M. Sun, and H. Chen. Direct voxel grid optimiza- tion: Super-fast convergence for radiance fields reconstruc- tion. In Conference on Computer Vision and Pattern Recog- nition, pages 5459–5469, 2022. 2

  33. [41]

    J. Sun, Z. Shen, Y . Wang, H. Bao, and X. Zhou. LoFTR: Detector-Free Local Feature Matching with Transformers. In Conference on Computer Vision and Pattern Recognition,

  34. [42]

    Tancik, P

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. In Advances in Neural Information Processing Systems, 2020. 2

  35. [43]

    Tancik, E

    M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa. Nerfstudio: A Modular Framework for Neural Radiance Field Development. In ACM SIGGRAPH,

  36. [44]

    Sparf: Neural radiance fields from sparse and noisy poses

    Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4190–4200, 2023. 2, 3

  37. [45]

    Pynerf: Pyramidal neural radiance fields

    Haithem Turki, Michael Zollh ¨ofer, Christian Richardt, and Deva Ramanan. Pynerf: Pyramidal neural radiance fields. Advances in Neural Information Processing Systems , 36,

  38. [46]

    Tyszkiewic, P

    M. Tyszkiewic, P. Fua, and E. Trulls. DISK: Learning Lo- cal Features with Policy Gradient. In Advances in Neural Information Processing Systems, 2020. 3

  39. [47]

    Scade: Nerfs from space carving with ambiguity-aware depth estimates

    Mikaela Angelina Uy, Ricardo Martin-Brualla, Leonidas Guibas, and Ke Li. Scade: Nerfs from space carving with ambiguity-aware depth estimates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16518–16527, 2023. 2

  40. [48]

    Dig- ging into depth priors for outdoor neural radiance fields

    Chen Wang, Jiadai Sun, Lina Liu, Chenming Wu, Zhelun Shen, Dayan Wu, Yuchao Dai, and Liangjun Zhang. Dig- ging into depth priors for outdoor neural radiance fields. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1221–1230, 2023. 2

  41. [49]

    Sparsenerf: Distilling depth ranking for few-shot novel view synthesis

    Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zi- wei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9065–9076,

  42. [50]

    Ibr- net: Learning multi-view image-based rendering

    Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibr- net: Learning multi-view image-based rendering. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and ...

  43. [51]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 7

  44. [52]

    Masked feature predic- tion for self-supervised visual pre-training

    Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature predic- tion for self-supervised visual pre-training. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668–14678, 2022. 2 10

  45. [53]

    S- volsdf: Sparse multi-view stereo regularization of neural im- plicit surfaces

    Haoyu Wu, Alexandros Graikos, and Dimitris Samaras. S- volsdf: Sparse multi-view stereo regularization of neural im- plicit surfaces. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 3556–3568,

  46. [54]

    Reconfusion: 3d reconstruction with diffusion priors

    Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patt...

  47. [55]

    Revealing the dark secrets of masked im- age modeling

    Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked im- age modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 14475– 14485, 2023. 3

  48. [56]

    C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neu- ral surface reconstruction

    Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhao- jie Zeng, Junle Wang, and Wei Yang. C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neu- ral surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, page...

  49. [57]

    Freenerf: Im- proving few-shot neural rendering with free frequency reg- ularization

    Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Im- proving few-shot neural rendering with free frequency reg- ularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254–8263,

  50. [58]

    K. M. Yi, E. Trulls, V . Lepetit, and P. Fua. LIFT: Learned Invariant Feature Transform. In European Conference on Computer Vision, 2016. 3

  51. [59]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578–4587, 2021. 1, 2

  52. [60]

    Monosdf: Exploring monocu- lar geometric cues for neural implicit surface reconstruc- tion

    Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sat- tler, and Andreas Geiger. Monosdf: Exploring monocu- lar geometric cues for neural implicit surface reconstruc- tion. Advances in neural information processing systems , 35:25018–25032, 2022. 1, 2

  53. [61]

    Cor-gs: sparse-view 3d gaussian splatting via co-regularization

    Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatting via co-regularization. In European Conference on Computer Vision, pages 335–352. Springer, 2024. 2

  54. [62]

    Zhang, G

    K. Zhang, G. Riegler, N. Snavely, and V . Koltun. Nerf++: Analyzing and Improving Neural Radiance Fields. In arXiv Preprint, 2020. 1, 2

  55. [63]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018. 7

  56. [64]

    ibot: Image bert pre-training with online tokenizer

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 ,

  57. [65]

    Hanxin Zhu, Tianyu He, Xin Li, Bingchen Li, and Zhibo Chen. Is vanilla mlp in neural radiance field enough for few- shot view synthesis? In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 20288–20298, 2024. 2 11

Pith tools

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