PhysicsNeRF: Physics-Guided 3D Reconstruction from Sparse Views

Hasan Kurban; Hussein Alnuweiri; Mohamed Rayan Barhdadi

REVIEW 4 major objections 5 minor 21 references

PhysicsNeRF: Physics-Guided 3D Reconstruction from Sparse Views

T0 review · 4 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 PhysicsNeRF claims that a compact NeRF extended with four physics constraints can reconstruct 3D scenes from only eight training views, reaching an average training PSNR of 21.4 dB and a test PSNR of 15.2 dB.

desk verdict Plausible combination of known regularizers with a convincing ablation, but the load-bearing cross-view consistency loss is under-specified and the evaluation is too thin to trust the headline numbers yet. read the letter →

arxiv 2505.23481 v2 pith:VMHV47GB submitted 2025-05-29 cs.CV

Mohamed Rayan Barhdadi , Hasan Kurban , Hussein Alnuweiri This is my paper

classification cs.CV

keywords neuralradiancefieldssparse-viewreconstructionphysics-guidedpriorsdepthrankingcross-viewconsistencynovelviewsynthesisoverfittinggeneralizationgap

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

PhysicsNeRF is an attempt to make neural radiance fields work when only eight training views are available, a regime where standard NeRF overfits badly. The paper adds four complementary constraints to a deliberately small 0.67-million-parameter network: ordinal depth ranking from monocular estimates, agreement between rays that should meet at the same 3D point, volumetric sparsity, and a smoothness penalty. It reports the best test PSNR on Chair and Lego among the compared methods and an average of 21.4 dB training and 15.2 dB test PSNR, with each added constraint shrinking the train-test gap, on Lego from 13.5 dB down to 6.7 dB. The authors' deeper point is that the remaining 5.7–6.2 dB generalization gap is not an implementation artifact but a structural consequence of sparse supervision, so fixed-form physics priors can improve but not solve the task.

What carries the argument

The argument rides on the weighted sum $L_{total} = L_{rgb} + \alpha(t)(\lambda_{depth} L_{depth} + \lambda_{cv} L_{cv} + \lambda_{sparse} L_{sparse} + \lambda_{reg} L_{reg})$, with $\alpha(t)$ a piecewise-constant curriculum that ramps the physics terms up at 5k and 15k iterations. $L_{depth}$ compares ordinal depth signs between monocular estimates and predictions; $L_{cv}$ enforces equality between the radiance field's outputs for corresponding rays from different poses; $L_{sparse}$ penalizes softplus density following VolSDF; $L_{reg}$ smooths along rays. The compact dual-scale MLP (7 layers, 192 hidden units, 0.67M parameters) is itself part of the machinery, since capacity control is what keeps the under-determined problem from exploding. The observed collapse–recovery behavior near 20k iterations is reported as a signature of the optimization passing through different solution regimes as the constraints engage.

What would settle it

Retrain the model on the same 8-view protocol with the $L_{cv}$ correspondences produced by a published stereo matcher, or, if the correspondences come from the model's own geometry, with the term removed, and check whether the reported 15.0 dB test PSNR on Lego survives; also run the method on the five NeRF synthetic scenes missing from Table 1 and verify that the 21.4/15.2 dB averages still hold.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a compact physics-constrained NeRF can generalize from eight fixed views. On the NeRF synthetic scenes Chair, Lego, and Drums, it reports the best test PSNR on Chair (18.5 dB versus a 13.8 dB next-best) and on Lego (15.0 dB versus 13.0 dB), while on Drums it trails DietNeRF and SparseNeRF (12.0 versus 12.8 dB). The ablation shows each added constraint narrows the train-test gap, and the paper interprets the remaining 5.7–6.2 dB gap and the collapse–recovery dynamics as evidence that sparse supervision is fundamentally under-determined. The 21.4 dB average in the abstract is the training-split PSNR; the test average is 15.2 dB.

Load-bearing premise

The cross-view consistency loss silently assumes the method can reliably identify which rays from different cameras hit the same 3D point from only 8 views; the paper does not specify the mechanism, so the term's supervision could be circular or dependent on an unstudied matcher.

Editorial extensions

If this is right

If the 8-view result holds, sparse-capture 3D reconstruction becomes more practical: a handful of photos, not a dense sweep, could initialize a usable radiance field for agent interaction and simulation.
The shrinking train-test gap across the ablation (13.5, then 11.8, 9.9, 8.5, and finally 6.7 dB on Lego) implies that each added constraint contributes independent generalization signal, a recipe transferable to other under-determined 3D learning tasks.
A persistent 5.7–6.2 dB gap even with all priors active means sparse-view NeRF is bounded by information content, not by architecture; further gains require new supervision, not just more regularization.
The 0.67M-parameter design outperforming larger baselines on the easier scenes suggests capacity control is a first-class regularizer when views are scarce.

Reading between the lines

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

The paper's reported tables cover only three of the eight standard NeRF synthetic scenes; until the remaining five are evaluated, the 21.4/15.2 dB averages and the broader 'outperforms prior methods' claim are only partially verified.
The cross-view consistency term requires knowing which rays from different cameras meet at the same 3D point, but the paper never defines the correspondence mechanism; my reading is that any implementation must choose between an external matcher (whose accuracy under extreme sparsity is untested) and self-supervision from the model's own geometry (which risks reinforcing errors).
A natural extension the authors imply but do not test is making the constraint weights $\alpha(t)$ scene-adaptive instead of piecewise-constant; the observed correlation between geometric complexity and generalization gap suggests the optimal schedule likely depends on the scene.

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.

Desk Editor's Note

read the letter

Here's my read. The paper combines four known regularizers (depth ranking from SparseNeRF, cross-view consistency, sparsity from VolSDF, smoothness) in a compact dual-scale NeRF and reports solid PSNR gains on three synthetic scenes with 8 views. The ablation in Table 2 is the most convincing part: each constraint improves test PSNR monotonically on Lego, and the full model beats RegNeRF and SparseNeRF by 2-3 dB. The code is linked. That is a legitimate, if incremental, contribution.

The soft spots are real but not fatal. The biggest one is Eq. (2). I have no idea how they get pairs of rays that "project to the same 3D point" from 8 views without metric depth. The stress-test note holds up on reading: if it's ground-truth depth, the method is oracle-based; if it's their own predicted geometry, the +1.6 dB from this term in Table 2 could be self-confirmation. The paper needs to say where these correspondences come from, and what Fθ(r) actually returns (integrated color or pointwise field value). This is load-bearing because the claimed advantage over RegNeRF/SparseNeRF is largely the last few dB.

Second, the evaluation is thin: three of eight NeRF-synthetic scenes, one run each, no error bars, no explanation for the missing five. The abstract says a "5.7-6.2 dB generalization gap" but Table 1 shows 4.7, 6.7, 7.2, with an average of 6.2. That inconsistency undercuts confidence in the numbers, even though nothing in the table looks impossible.

Third, the theory section is mostly generic PAC talk, not specific to NeRF or to their constraints. That is a minor overclaim, not a flaw in the experiments.

Overall: the central idea is plausible and the ablation supports it, but the under-specified cross-view loss and the thin evaluation keep me from trusting the headline number yet. I'd send this to a serious referee — the method is worth engaging with — but I'd expect heavy revision: specify the correspondence mechanism, run all eight scenes with multiple seeds, and fix the gap reporting. It's a fair paper, just not a finished one.

Referee Report

4 major / 5 minor

Summary. PhysicsNeRF extends Neural Radiance Fields with a compact 0.67M-parameter dual-scale MLP and four auxiliary losses for 8-view novel-view synthesis: monocular depth ranking (Eq. 1), cross-view consistency (Eq. 2), volumetric sparsity (Eq. 3), and gradient regularization (Eq. 4). The paper reports train/test/gap PSNR on three NeRF-synthetic scenes (Chair, Lego, Drums), an ablation on Lego, and training-dynamics observations. The central claim is that these physics-guided constraints outperform NeRF, RegNeRF, DietNeRF, and SparseNeRF under 8-view supervision and reduce the train-test PSNR gap.

Significance. The paper addresses an important problem—severe overfitting in sparse-view NeRF—and the proposed combination of depth ranking, sparsity, and cross-view consistency is a reasonable inductive-bias strategy. If the reported test-PSNR gains (e.g., Chair 18.5 dB vs. 13.8 dB for DietNeRF) are reproducible, they would be a useful contribution. The paper also makes a concrete attempt to compare constraint ablations and to analyze collapse-recovery dynamics. However, the empirical evidence is currently incomplete: only 3 of 8 benchmark scenes are reported, all numbers are single-run, and the abstract's headline 21.4 dB is the average training PSNR rather than the test PSNR. The cross-view consistency term, which Table 2 credits with a 1.6 dB improvement, is not specified at the implementation level. These issues must be addressed before the claims can be evaluated.

major comments (4)

[Abstract; Table 1] The abstract states that PhysicsNeRF 'achieves 21.4 dB average PSNR using only 8 views, outperforming prior methods' and that a generalization gap of 5.7–6.2 dB is consistently observed. Table 1 reports train/test/gap values of 23.2/18.5/4.7 (Chair), 21.7/15.0/6.7 (Lego), and 19.2/12.0/7.2 (Drums), so the 21.4 dB figure is the average training PSNR, not the test PSNR, and the test average is 15.2 dB. The stated generalization-gap range also does not match the reported scene gaps (4.7, 6.7, and 7.2 dB). This discrepancy affects the paper's central claim and must be corrected in the abstract and in Section 4.2, which repeats the 5.7–6.2 dB range.
[3.2, Table 1] The NeRF synthetic dataset contains eight scenes, but Table 1 reports results only for Chair, Lego, and Drums. The text says results are 'averaged across all objects' and concludes that PhysicsNeRF 'consistently outperforms baselines,' yet no numbers are given for the five omitted scenes. With only single-run results and no standard deviation or repeated-seed information, the claim of consistent superiority is not supported by the presented evidence. The authors should provide the full eight-scene table, or explicitly restrict the claims to the three reported scenes.
[2.2.2, Eq. (2)] The cross-view consistency loss is the load-bearing addition in the ablation (Table 2 reports +1.6 dB on Lego), but the manuscript never specifies how the ray pairs (r_{k,1}, r_{k,2}) that 'project to the same 3D point' are obtained for eight sparse views. If they are derived from ground-truth depth or an external matcher, that source must be described; if they are derived from the model's own geometry, the loss is not independent supervision. In addition, Fθ(r) is ambiguous: comparing integrated ray colors conflates legitimate view-dependent effects and occlusions, while comparing pointwise field values requires knowing the shared 3D location. Without this specification, Eq. (2) is not reproducible and the attribution of the Table 2 improvement to this term is unverifiable.
[4.1] Section 4.1 claims the generalization gap scales as O(sqrt(|θ|/N)) and cites Srivastava et al. (2014) and Truong et al. (2022), but no derivation connects this generic bound to the specific NeRF architecture, the 8-view setting, or the proposed constraints. The discussion is qualitative and does not establish the stated 'theoretical analysis' contribution. If the theory is meant to be a contribution, it needs a concrete statement of the hypothesis class and a derivation; otherwise it should be reframed as intuition rather than a formal result.

minor comments (5)

[2.2.1, Eq. (1)] Equation (1) uses sgn and a pairwise ranking loss ℓrank, but ℓrank is never defined; please specify the exact loss (e.g., hinge versus logistic) and how the pixel pairs P are sampled.
[2.2.3, Eq. (3)] In Section 2.2.3, 'Following V olSDF' contains a spacing typo, and the relationship of Eq. (3) to VolSDF's actual regularization should be clarified, since VolSDF uses a different density parameterization and loss formulation.
[3.3, Figure 2] Figure 2 says collapse–recovery occurs around 20k iterations, while the α(t) schedule in Eq. (5) changes at 5k and 15k; the connection between the schedule and the observed dynamics should be stated explicitly.
[Abstract; 2.2.3] The abstract and Section 2.2.2 refer to a 'RegNeRF-style consistency' constraint, but Eq. (4) is a gradient regularizer; if this term is intended to correspond to RegNeRF, the relationship should be explained, otherwise the wording should be changed.
[3.2, Tables 1 and 2] Tables 1 and 2 report PSNR to one decimal place without reporting the number of random seeds or the standard deviation; please state whether these are single runs and, if so, add at least mean ± std over multiple seeds.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PhysicsNeRF's constraints are external priors and the claimed PSNR is evaluated on held-out views.

full rationale

The central claim is an empirical comparison on held-out test views (Table 1), not a quantity derived from its own inputs. Each auxiliary loss is an independent regularizer drawn from prior external work: depth ranking (Eq. 1) follows SparseNeRF; the cross-view term (Eq. 2) is an additional consistency penalty whose correspondence mechanism is under-specified but is not shown to reduce to Fθ itself; sparsity (Eq. 3) follows VolSDF; the gradient penalty (Eq. 4) is a standard smoothness prior. The model is trained on 8 input views and evaluated on novel test views, so test PSNR is not statistically forced by a fitted parameter. There is no load-bearing self-citation chain: the cited theoretical results (Rademacher complexity, generalization bounds) are external, and the authors do not invoke their own prior uniqueness theorems. The reported 21.4 dB figure is the training PSNR while the held-out test average is 15.2 dB; the abstract's '5.7-6.2 dB' gap range also conflicts with Table 1's per-scene gaps (4.7/6.7/7.2 dB). These are reporting inconsistencies, not circular reductions. The unstated source of rk,1/rk,2 in Eq. 2 is a reproducibility/validity concern that would matter if the correspondence pairs came from oracle depth or from the model's own geometry, but the paper does not specify either mechanism, so under the hard rules no circular step can be demonstrated.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper rests on several domain assumptions about the reliability of external depth estimates and geometric correspondences, plus hand-chosen hyperparameters. No new physical entities or mathematically invented objects are introduced; the constraints are standard priors imported from prior work.

free parameters (3)

Progressive constraint schedule alpha(t) = 0.008, 0.025, 0.08
Piecewise-constant schedule in Eq. 5, chosen by hand with no sensitivity analysis or justification reported.
Constraint weights lambda_i = Not reported
Fixed weights for the four auxiliary losses in Eq. 5; values are never given, so the balance between terms is unknown to the reader.
Model capacity (192 hidden units, 7 layers, dual-scale) = 0.67M parameters
Architecture size selected to balance capacity and generalization without ablating other sizes; it directly influences the observed train-test gap.

assumptions (4)

domain assumption MiDaS monocular depth estimates provide sufficient ordinal depth accuracy to supervise the depth ranking loss.
Eq. 1 uses sign comparisons of MiDaS depth as ground truth; if depth ranking is noisy, the loss can bias geometry. No validation of MiDaS accuracy on the synthetic dataset is provided.
domain assumption Corresponding rays across different views can be reliably identified for the cross-view consistency loss.
Eq. 2 assumes such correspondences exist and are known; the paper does not describe the matching procedure under sparse views.
domain assumption Natural scene radiance fields are sparse and smooth, so sparsity and gradient penalties improve generalization.
Eqs. 3-4 impose these priors; they are plausible for synthetic objects, but no evidence confirms they hold in all tested scenes.
domain assumption The generalization bound O(sqrt(|theta|/N)) from Srivastava et al. applies to radiance fields and explains the observed gaps.
Section 4.1 invokes this bound without adapting it to NeRF's continuous volume rendering setup and without quantitative comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PhysicsNeRF: Physics-Guided 3D Reconstruction from Sparse Views." pith.science (2026). https://pith.science/paper/VMHV47GB

@misc{pith2026250523481,
  author       = {Pith},
  title        = {Pith review of: PhysicsNeRF: Physics-Guided 3D Reconstruction from Sparse Views},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VMHV47GB}},
  note         = {Machine review of arXiv:2505.23481}
}

read the original abstract

PhysicsNeRF is a physically grounded framework for 3D reconstruction from sparse views, extending Neural Radiance Fields with four complementary constraints: depth ranking, RegNeRF-style consistency, sparsity priors, and cross-view alignment. While standard NeRFs fail under sparse supervision, PhysicsNeRF employs a compact 0.67M-parameter architecture and achieves 21.4 dB average PSNR using only 8 views, outperforming prior methods. A generalization gap of 5.7-6.2 dB is consistently observed and analyzed, revealing fundamental limitations of sparse-view reconstruction. PhysicsNeRF enables physically consistent, generalizable 3D representations for agent interaction and simulation, and clarifies the expressiveness-generalization trade-off in constrained NeRF models.

Figures

Figures reproduced from arXiv: 2505.23481 by the authors.

**Figure 1.** Architecture of PhysicsNeRF. Dual-scale coordinate encoding is combined with four physics-based constraints: (1) MiDaS-based depth ranking, (2) cross-view coherence, (3) sparsity regularization, and (4) RegNeRF-style consistency. The model uses a balanced 0.67M-parameter design optimized for sparse-view reconstruction. Physics-informed neural networks (PINNs) (Raissi et al., 2017) provide a framework for incorporati… view at source ↗

**Figure 2.** Training Dynamics Evolution. PSNR progression showing collapse-recovery patterns at 20k iterations during progressive constraint engagement, revealing critical insights into physicsguided optimization landscapes and fundamental overfitting challenges. 3.3. Training Dynamics and Collapse-Recovery Analysis [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 17 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

Curriculum learning

Bengio, Y., Louradour, J., Collobert, R., and Weston, J. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, pp.\ 41--48, 2009

2009
[3]

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

Deng, K., Liu, A., Zhu, J.-Y., and Ramanan, D. Depth-supervised NeRF : Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

work page 2022
[4]

PIE-NeRF : Physics-informed explicit neural radiance fields

Feng, W., Zhang, X., Liang, D., Chen, Y., and Zhao, W. PIE-NeRF : Physics-informed explicit neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

work page 2023
[5]

Plenoxels : Radiance fields without neural networks

Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., and Kanazawa, A. Plenoxels : Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 5501--5510, 2022

work page 2022
[6]

and Zisserman, A

Hartley, R. and Zisserman, A. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003

work page 2003
[7]

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

Jain, A., Tancik, M., and Abbeel, P. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 5885--5894, 2021

work page 2021
[8]

Y., Jatavallabhula, K

Li, X., Qiao, Y.-L., Chen, P. Y., Jatavallabhula, K. M., Lin, M., Jiang, C., and Gan, C. PAC-NeRF : Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In International Conference on Learning Representations (ICLR), 2023

work page 2023

Show all 21 references

[9]

Neural sparse voxel fields

Liu, L., Gu, J., Zaw Lin, K., Chua, T.-S., and Theobalt, C. Neural sparse voxel fields. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pp.\ 15651--15663, 2020

2020
[10]

S., Barron, J

Martin-Brualla, R., Radwan, N., Sajjadi, M. S., Barron, J. T., Dosovitskiy, A., and Duckworth, D. NeRF-W : Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 7210--7219, 2021

2021
[11]

P., Tancik, M., Barron, J

Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., and Ng, R. NeRF : Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020

2020
[12]

Instant neural graphics primitives with a multiresolution hash encoding

M \"u ller, T., Evans, A., Schied, C., and Keller, A. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41 0 (4): 0 1--15, 2022. SIGGRAPH 2022 Best Paper Award

2022
[13]

T., Mildenhall, B., Sajjadi, M

Niemeyer, M., Barron, J. T., Mildenhall, B., Sajjadi, M. S., Geiger, A., and Radwan, N. RegNeRF : Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 5470--...

2022
[14]

D-NeRF : Neural radiance fields for dynamic scenes

Pumarola, A., Corona, E., Pons-Moll, G., and Moreno-Noguer, F. D-NeRF : Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 10318--10327, 2021

2021
[15]

Raissi, M., Perdikaris, P., and Karniadakis, G. E. Physics informed deep learning (part I ): Data-driven solutions of nonlinear partial differential equations. arXiv preprint arXiv:1711.10561, 2017

2017 arXiv
[16]

T., Mildenhall, B., Srinivasan, P

R \"o ssle, B., Barron, J. T., Mildenhall, B., Srinivasan, P. P., and Nie ner, M. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 12892--12901, 2022

2022
[17]

M., Curless, B., Diebel, J., Scharstein, D., and Szeliski, R

Seitz, S. M., Curless, B., Diebel, J., Scharstein, D., and Szeliski, R. A comparison and evaluation of multi-view stereo reconstruction algorithms. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1: 0 519--528, 2006

2006
[18]

Dropout: A simple way to prevent neural networks from overfitting

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15 0 (1): 0 1929--1958, 2014

1929
[19]

On the rademacher complexity of neural radiance fields

Truong, P., Rakotosaona, M.-J., Manhardt, F., and Tombari, F. On the rademacher complexity of neural radiance fields. arXiv preprint arXiv:2211.11738, 2022

2022 arXiv
[20]

C., and Liu, Z

Wang, G., Chen, Z., Loy, C. C., and Liu, Z. SparseNeRF : Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

2023
[21]

Volume rendering of neural implicit surfaces

Yariv, L., Gu, J., Kasten, Y., and Lipman, Y. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021

2021

Pith tools

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

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

Curriculum learning

Bengio, Y., Louradour, J., Collobert, R., and Weston, J. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, pp.\ 41--48, 2009

2009

[3] [3]

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

Deng, K., Liu, A., Zhu, J.-Y., and Ramanan, D. Depth-supervised NeRF : Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

work page 2022

[4] [4]

PIE-NeRF : Physics-informed explicit neural radiance fields

Feng, W., Zhang, X., Liang, D., Chen, Y., and Zhao, W. PIE-NeRF : Physics-informed explicit neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

work page 2023

[5] [5]

Plenoxels : Radiance fields without neural networks

Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., and Kanazawa, A. Plenoxels : Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 5501--5510, 2022

work page 2022

[6] [6]

and Zisserman, A

Hartley, R. and Zisserman, A. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003

work page 2003

[7] [7]

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

Jain, A., Tancik, M., and Abbeel, P. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 5885--5894, 2021

work page 2021

[8] [8]

Y., Jatavallabhula, K

Li, X., Qiao, Y.-L., Chen, P. Y., Jatavallabhula, K. M., Lin, M., Jiang, C., and Gan, C. PAC-NeRF : Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In International Conference on Learning Representations (ICLR), 2023

work page 2023