Pith. sign in

REVIEW 3 major objections 5 minor 53 references

CASA-SDF: Curriculum-Aware Spatial Adaptation with Curvature-Guided Density for Neural Implicit Surface Reconstruction

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Indoor 3D reconstruction improves by spatially varying the SDF-to-density sharpness and annealing unreliable monocular priors.

desk verdict A competent empirical paper with two genuinely new mechanisms, but the load-bearing locally-constant sharpness approximation is not adequately defended and the supporting illustration is in an absent supplement. read the letter →

arxiv 2607.13492 v1 pith:Z23O5ZMO submitted 2026-07-15 cs.CV cs.GR

Lei Yang , Weiqing Li , Zhiyong Su , Liang Xiao This is my paper · ORCID
classification cs.CVcs.GR
keywords neuralimplicitsurfacesindoorreconstructionsigneddistancefieldscurriculumlearninguncertainty-awareoptimizationcurvature-adaptivedensitymonocularpriorssurfacecompleteness
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

CASA-SDF seeks to break the planar-vs-thin-structure trade-off in indoor neural implicit surface reconstruction by adapting both where supervision comes from and how sharp the density mapping is. It claims that a pixel-wise curriculum which anneals monocular depth and normal losses early in unreliable regions — defined by a fusion of semantic and photometric uncertainty — lets photometric evidence refine geometry where priors lie, while a curvature-guided sharpness modulation raises the SDF-to-density bandwidth on high-curvature structures and keeps it low on flat walls. The paper reports that combining these two mechanisms yields the best completeness and recall on standard indoor benchmarks, with F-score gains concentrated on high-frequency structures (0.772 vs 0.724 for the strongest baseline) and no loss of planar stability. A sympathetic reader would care because it suggests the failure to reconstruct thin structures is not an unavoidable spectral-bias limit but a consequence of spatially indiscriminate prior weighting and representation bandwidth.

What carries the argument

The central mechanism is the spatially and temporally varying sharpness s(x,tau) = s_base (1 + g(tau) lambda_H tanh(|H_est(x)| R)^{alpha_H}) that replaces the global sharpness in the standard NeuS SDF-to-density transform T(t) = 1/(1+exp(-s(x,tau) f(t))), together with the pixel-wise curriculum weight w_prior(r,tau) = lambda_base exp(-eta tau/(1-U_total(r)+eps)) that gates monocular prior losses. Here U_total is the max of a semantic uncertainty (exponential of negative normalized von Mises–Fisher concentration) and a photometric uncertainty (1 - best-match warped SSIM), and |H_est| is a finite-difference normal-variation proxy on the tangent plane with step size 0.01 R. The sharpness modula

What would settle it

Take a synthetic scene with a known thin rod and ground-truth SDF. Train CASA-SDF and, along rays passing within 5 cm of the rod, record the position of the peak of the rendering weight w(t) relative to the zero crossing of the predicted SDF. If the weight peak deviates from the zero crossing by more than about 1 cm in a substantial fraction of such rays, and the deviation grows with the curvature proxy, then the local-constancy/stop-gradient approximation fails and the reconstructed surface is biased in exactly the regions the method aims to improve.

Watch

Extended reading notes

Core claim

The central claim is that indoor geometric heterogeneity is best handled as two coupled problems: how long each monocular prior should guide optimization, and how much bandwidth the SDF-to-density mapping should allocate at each point. The paper's specific discovery is that a conservative max-fusion of semantic uncertainty (from a von Mises–Fisher normal estimator) and photometric uncertainty (from multi-view patch SSIM warping) produces a pixel-wise curriculum w_prior(r,tau) that decays exponentially in training time, with decay rate controlled by uncertainty; and that multiplying the global sharpness by 1 + g(tau) lambda_H tanh(|H_est| R)^{alpha_H}, where |H_est| is a finite-difference nor

Load-bearing premise

The load-bearing premise is that treating the spatially varying sharpness s(x,tau) as locally constant when deriving the density rho(t) = -d/dt log T(t) — with a stop-gradient on the curvature term — leaves the rendered surface peak at the SDF zero-level set; the paper asserts 'no noticeable peak drift' but defers the supporting 1D illustration to a supplementary file that is not included in the arXiv version.

Editorial extensions

If this is right

  • Any NeuS-style reconstruction pipeline can adopt spatially varying sharpness without new network components, at about 9% extra training time and 12% peak memory, since the curvature proxy is a finite-difference computation on existing normals.
  • Monocular prior supervision should be treated as a time-varying resource: strong early for coarse stabilization, released early where uncertainty is high, and retained where confidence is high.
  • The finite-difference normal-variation proxy for curvature is a cheap, scale-invariant signal that can be reused for adaptive ray sampling, loss reweighting, or mesh refinement.
  • The region-aware evaluation shows that global F-score is dominated by planar surfaces; without reporting high-frequency regions separately, gains on thin structures can be hidden in aggregate metrics.
  • The combination of SAUA and CALADT reduces large outlying errors in high-frequency regions even though global accuracy changes little, implying the two mechanisms remove a common failure mode rather than boosting all surfaces uniformly.

Reading between the lines

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

  • The local-constancy approximation for s could introduce a systematic bias in the weight peak location that is not captured by a 5 cm F-score; a sub-centimeter peak shift in high-curvature regions would be worth measuring directly.
  • The exponential curriculum schedule could be replaced by a learned or residual-driven schedule per pixel, which might further improve the photometric refinement phase in regions where priors remain unreliable.
  • The same two-mechanism template could transfer to sparse-view reconstruction, where reliability of priors and bandwidth demands are even more heterogeneous; this remains to be tested.
  • The progressive gate g(tau) makes a testable prediction: activating the spatial sharpness modulation from the start should destabilize planar regions, so an ablation that removes the gate should show increased planar noise.
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 CASA-SDF, a neural implicit surface reconstruction framework for indoor scenes, targeting the tension between planar smoothness and thin-structure preservation. Two mechanisms are introduced: (1) SAUA, a pixel-wise curriculum that anneals monocular normal/depth supervision based on a max-fusion of semantic (AngMF) and photometric (patch SSIM) uncertainties; and (2) CALADT, which modulates the NeuS SDF-to-density sharpness parameter s(x, tau) by a normal-variation curvature proxy with a progressive activation schedule. Experiments on ScanNet and Replica report competitive or best global F-scores (0.818 and 0.788), with the clearest gains in a GT-curvature-based high-frequency partition (0.772 vs. 0.724 for ND-SDF on ScanNet). Ablations and region-aware evaluations support the complementary role of the two components.

Significance. If the results hold, CASA-SDF provides a practical resolution to a known trade-off in indoor implicit reconstruction: it improves recall/completeness on thin structures without sacrificing planar stability, at a modest training-time cost (9.3% per-iteration overhead). The paper's strengths include a thoughtful region-aware evaluation protocol defined on ground-truth curvature, transparent ablation of both modules, and a clear discussion of the precision/recall trade-off. The reported high-frequency gains are substantial and plausible given the mechanism. However, the central technical claim depends on the validity of treating the spatially varying sharpness as locally constant in the NeuS density derivative; the paper's only defense is deferred to a supplementary that is not part of the arXiv submission. Additionally, the main tables aggregate numbers from multiple prior publications and report no variance, which weakens the statistical support for the headline 'best' claims. These issues are fixable but require additional evidence.

major comments (3)
  1. [Sec. 3.3 (Eq. 11, 'Remarks on Localization Fidelity')] The claim that treating s(x,tau) as locally constant in the density derivative preserves the 'symmetric kernel structure' of NeuS is not correct for the forward computation. With T(t)=sigmoid(s(t)f(t)) and rho(t)=-s(t) f'(t)(1-T(t)), the rendering weight becomes w(t)= -s(t) f'(t) T(t)(1-T(t)). For a ray crossing f=0, the factor s(t) is generally not symmetric about the zero crossing, so the peak of w(t) is displaced unless s'(t)=0 there. The stop-gradient only affects backpropagation; it does not alter this forward asymmetry. The progressive gate (tau0,tau1) mitigates but cannot eliminate the effect, and the bias is concentrated in the high-curvature regions that drive the reported gains in Table 7. The only defense is a sentence deferring to a supplementary 1D illustration, which is not included in this arXiv version. Please provide the 1D illustration or, more convincingly, measure the
  2. [Tables 1 and 2] The main comparative tables aggregate numbers from disparate publications and codebases, and the paper explicitly states that controlled reproduced results are in the Supplementary Material, which is absent from the arXiv submission. This weakens the central comparative claims: the global F-score difference over ND-SDF on ScanNet is 0.818 vs. 0.820, a margin that could easily be inverted under different mesh-extraction or culling protocols. Please include the controlled comparison in the main text (or at least make the supplement available and reference it with concrete numbers), and report the evaluation protocol used for the baselines in the same table.
  3. [All experimental tables (Tables 1-7)] All metrics are reported as point estimates without variance, per-scene breakdowns, or statistical tests. Given the small global F-score margins (e.g., 0.818 vs. 0.820 on ScanNet) and the known sensitivity of 5 cm F-scores to training stochasticity, the 'best' claims are not statistically grounded. The region-aware high-frequency difference (0.772 vs. 0.724) is larger, but still should be accompanied by variance or per-scene results. Please report repeated-run means with standard deviations, or at minimum per-scene scores, for the main comparisons and ablations.
minor comments (5)
  1. [Sec. 3.2, Eq. (5)] The co-visibility set V used for photometric uncertainty is not formally defined. Please clarify how neighboring source views are selected and how 'no valid source view' is determined.
  2. [Sec. 3.3, Eq. (8)] The tangent vectors t1 and t2 are not specified. Please state how they are computed (e.g., orthonormal basis from the normal at x) and how the finite-difference step delta is chosen consistently across scenes.
  3. [Sec. 4.5] The region-aware protocol uses GT-curvature percentile thresholds (75%/15%/10%) computed on the pooled ScanNet scenes. Sensitivity to these thresholds is deferred to the supplementary. Please report at least a one-line sensitivity check (e.g., varying the high-frequency cutoff) in the main text.
  4. [Fig. 4] The convergence plot shows mean F-score without error bars. Since this is one of the few figures showing training dynamics, adding per-scene spread would make the comparison more informative.
  5. [Throughout] The paper would benefit from a notation table for the many hyperparameters (lambda_base, eta, epsilon, kappa0, beta0/s_base, lambda_H, alpha_H, tau0, tau1). Some are described only in Sec. 4.2, which makes the method section harder to follow.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity; the only flagged item is an unsupported local-constancy claim, which is a correctness risk rather than a circular step.

full rationale

The paper's derivation chain is self-contained. SAUA's uncertainties come from an external frozen AngMF estimator (Eq. 4) and from multi-view SSIM patch warping (Eq. 5); these are offline inputs, and the time-dependent weight in Eq. 7 is a scheduling rule, not a fitted estimator of the reported F-scores. CALADT's curvature proxy (Eqs. 8-9) is computed by finite differences of the evolving SDF normals and modulates sharpness (Eq. 10), which then enters the NeuS-style volume-rendering weight. No equation defines the target metrics (Completeness, Recall, F-score) in terms of xi, s, or w_prior, so nothing reduces to its own output by construction. The region-aware evaluation partitions the GT mesh by curvature while CALADT uses the predicted SDF curvature; although both involve 'curvature,' they are distinct sources, and the observed high-frequency gains are empirical mechanism behavior, not a tautology. Self-citation is not load-bearing: the authors cite no prior work of their own, and the cited AngMF, Omnidata, NeuS, MonoSDF, DebSDF, and ND-SDF are external prior art. The one item worth flagging is not circular: Sec. 3.3 'Remarks on Localization Fidelity' admits the locally-constant-s approximation and defers the 'no noticeable peak drift' defense to a supplementary material not included in this arXiv version. That is an unverified modeling assumption and a correctness/robustness risk, but it is not a reduction of the result to its inputs. Accordingly, the circularity score is 0.

Assumptions & free parameters 12 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-chosen hyperparameters (annealing rate, modulation strength, gate intervals, loss weights) and on modeling assumptions about uncertainty calibration, SDF-gradient curvature proxies, and the stop-gradient treatment of spatially varying sharpness. No new physical entity is introduced.

free parameters (12)
  • lambda_base = 0.1
    Initial weight for monocular prior losses in Eq. (7); chosen by hand.
  • eta = 4.0
    Global annealing speed in Eq. (7); controls how fast prior supervision decays; no sensitivity analysis in main text.
  • epsilon = 1e-3
    Stability constant in Eq. (7) denominator; prevents division by zero; chosen by hand.
  • kappa0 = median of AngMF kappa over training pixels
    Normalization constant in Eq. (4); computed once from frozen AngMF predictions; rescales semantic uncertainty across scenes.
  • beta0 / s_base = 0.3 / 3.333
    Base bandwidth/sharpness in Eq. (10); global NeuS-style sharpness before curvature modulation.
  • lambda_H = 1.0
    Strength of curvature modulation in Eq. (10); selected by hand.
  • alpha_H = 2.0
    Exponent on xi in Eq. (10); controls how sharply curvature maps to bandwidth; no sensitivity analysis.
  • tau0, tau1 = 0.3, 0.6
    Progressive gate ramp endpoints; delays spatial sharpness modulation until curvature estimates stabilize.
  • delta = 0.01 * R
    Finite-difference step in Eq. (8); scale-invariant curvature proxy step size.
  • surface proximity threshold = 5 cm
    Activates curvature estimation only for samples with |f(x)| < 5cm; controls computational cost and proxy locality.
  • loss weights lambda_eik, lambda_d, lambda_n = 0.05, 0.005, 0.003
    Weights in Eq. (12); standard MonoSDF-style tuning.
  • region partition thresholds = 75/15/10 percentiles
    GT curvature percentiles defining planar/transitional/high-frequency regions in Sec. 4.5; chosen on pooled evaluation scenes.
assumptions (5)
  • domain assumption SDF gradients n(x) are unit normals and finite differences of n along tangent directions approximate local curvature magnitude.
    Used in Eq. (8) to compute H_est; neural SDFs with Softplus and Eikonal penalty only approximately satisfy the unit-gradient property, and one-sided finite differences on a noisy field are a rough proxy.
  • domain assumption AngMF concentration kappa and patch SSIM warping measure the true unreliability of monocular priors.
    SAUA Eqs. (4)-(6) assume both uncertainty signals correlate with prior error; no calibration or correlation analysis is provided.
  • domain assumption s(x,tau) can be treated as locally constant and stop-gradients do not shift rendering-weight peaks.
    Sec. 3.3 drops gradient-of-s terms from rho(t) and claims no peak drift, deferring evidence to absent supplementary; this is the key modeling assumption for CALADT.
  • ad hoc to paper The exponential annealing schedule and tanh curvature compression are suitable functional forms.
    Eqs. (7), (9), (10) are heuristic; the regimes (slow decay for confident, fast decay for uncertain) are designed by hand rather than derived.
  • domain assumption Monocular priors are metric-aligned and helpful early, harmful late, while photometric loss can refine geometry once priors are released.
    Motivates the curriculum in Eq. (12); assumes optimization dynamics behave as claimed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CASA-SDF: Curriculum-Aware Spatial Adaptation with Curvature-Guided Density for Neural Implicit Surface Reconstruction." pith.science (2026). https://pith.science/paper/Z23O5ZMO

@misc{pith2026260713492,
  author       = {Pith},
  title        = {Pith review of: CASA-SDF: Curriculum-Aware Spatial Adaptation with Curvature-Guided Density for Neural Implicit Surface Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z23O5ZMO}},
  note         = {Machine review of arXiv:2607.13492}
}
read the original abstract

Neural implicit representations have emerged as a powerful paradigm for 3D reconstruction. However, high-fidelity indoor surface reconstruction remains a significant challenge, primarily due to the pronounced \emph{geometric heterogeneity} of indoor scenes. Large texture-less planar regions typically require stronger regularization to suppress high-frequency artifacts, while thin structures demand sharper, more adaptive representations to mitigate the spectral bias of multi-layer perceptrons (MLPs) and prevent over-smoothing. Existing approaches often rely on spatially indiscriminate prior supervision and a scene-global SDF-to-density transformation, which constrains their ability to balance planar smoothness and detail preservation. In this paper, we propose CASA-SDF (Curriculum-Aware Spatial Adaptation for SDF), a unified framework that addresses this challenge via complementary adaptations of supervision and representation capacity. Specifically, Hybrid Spatially-Adaptive Uncertainty Annealing (SAUA) fuses semantic and photometric uncertainties to construct a pixel-wise curriculum for monocular prior supervision. This strategy maintains regularization in reliable regions while attenuating unreliable supervision early in training to enable data-driven photometric refinement. Meanwhile, Curvature-Aware Locally Adaptive Density Transformation (CALADT) progressively modulates the sharpness of the SDF-to-density mapping via a curvature proxy to enhance the representation of thin structures. Extensive experiments on benchmark indoor datasets demonstrate that CASA-SDF improves surface completeness and detail recovery on high-frequency structures, without compromising the stability of planar surfaces.

Figures

Figures reproduced from arXiv: 2607.13492 by the authors.

Figure 1
Figure 1. Breaking the trade-off in indoor reconstruction via curriculum-aware spatial adaptation. Indoor scenes exhibit geometric heterogeneity, which creates a fundamental conflict for standard reconstruction methods. (a) Input Views. (b) A strong baseline (e.g., DebSDF [1]) suffers from rigid global regularization scheme, resulting in noisy artifacts on texture-less planar regions (red box) and over-smoothed, broken geomet… view at source ↗
Figure 2
Figure 2. Overview of CASA-SDF. The pipeline consists of two complementary mechanisms. (a) SAUA (Sec. 3.2): a pre-trained Angular von Mises–Fisher (AngMF) estimator estimates aleatoric uncertainty 𝜅, which drives a pixel-wise curriculum for annealing monocular prior supervision. Reliable regions (e.g., walls) retain regularization, whereas unreliable regions (e.g., edges) transition earlier to photometric refinement. (b) CALA… view at source ↗
Figure 3
Figure 3. Visualization of the SAUA curriculum. Unlike static weighting, the proposed pixel-wise weights follow different trajectories over training. High-uncertainty regions (red) are annealed rapidly to release unreliable constraints, whereas confident regions (blue) retain regularization for a longer period. relatively gentle and delayed decay. It retains super￾visory influence significantly longer during the op￾timization… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: through two complementary perspectives: conver￾gence behavior and per-iteration overhead [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative mesh comparisons on ScanNet [49] for two representative scenes. CASA-SDF better preserves thin structures (e.g., chair legs), small objects (e.g., lamps), and wall-mounted details, while reducing floating artifacts on texture-less planar regions. Reference …
Figure 6
Figure 6. Figure 6: Qualitative normal comparisons on ScanNet [49]. CASA-SDF reconstructs smoother and more coherent normals on tabletop surface regions while preserving sharp boundaries. Synergistic behavior of SAUA and CALADT. Although adding SAUA or CALADT alone slightly degrades Accu￾…
Figure 7
Figure 7. Figure 7: Qualitative mesh comparisons on Replica [50]. Top-down views of room-scale geometry and zoomed local reconstructions are shown. Compared with the baselines, CASA-SDF better preserves the global room structure while recovering sharper and more complete fine details on t…
Figure 8
Figure 8. Figure 8: Qualitative normal comparisons on Replica [50]. CASA-SDF produces smoother wall and tabletop normals while retaining clear geometric discontinuities. with their conservative fusion strategy. In [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Visualization of Hybrid-SAUA uncertainty fusion. (a) Input RGB view. (b) Semantic uncertainty sem highlights boundaries and thin structures where CNN-based priors are unreliable. (c) Photometric uncertainty photo highlights regions where multi-view photometric consis…
Figure 10
Figure 10. Figure 10: Visualization of the spatially adaptive sharpness field 𝑠(𝐱). CALADT assigns stronger smoothing (small 𝑠) to planar regions and stronger sharpening (large 𝑠) to detailed structures. independent of predicted topology or local noise. Specifi￾cally, we estimate a normali…
Figure 11
Figure 11. Figure 11: Region-aware decoupled evaluation. (a) Ground-truth geometric partitioning into planar, transitional, and high￾frequency regions using the fixed GT-based curvature protocol described in Sec. 4.5. (b) Completeness error map of a representative baseline (MonoSDF), exhib…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 1 canonical work pages

  1. [1]

    Debsdf: Delvingintothedetailsandbiasofneuralindoorscenereconstruction

    Yuting Xiao, Jingwei Xu, Zehao Yu, and Shenghua Gao. Debsdf: Delvingintothedetailsandbiasofneuralindoorscenereconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  2. [2]

    Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  3. [3]

    PengWang,LingjieLiu,YuanLiu,ChristianTheobalt,TakuKomura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction.arXiv: Computer Vision and Pattern Recognition,arXiv: Computer Vision and Pattern Recognition, Jun 2021

    2021

  4. [4]

    Volume renderingofneuralimplicitsurfaces.AdvancesinNeuralInformation Processing Systems, 34:4805–4815, 2021

    Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume renderingofneuralimplicitsurfaces.AdvancesinNeuralInformation Processing Systems, 34:4805–4815, 2021

    2021

  5. [5]

    Pixelnerf: Neural radiance fields from one or few images

    Alex Yu et al. Pixelnerf: Neural radiance fields from one or few images. InCVPR, 2021

    2021

  6. [6]

    Mvsnerf: Fast generalizable radiance field recon- struction from multi-view stereo

    Anpei Chen et al. Mvsnerf: Fast generalizable radiance field recon- struction from multi-view stereo. InICCV, 2021

    2021

  7. [7]

    Sparseneus: Fast generalizable neural surface reconstruction from sparse views

    Xiaoxiao Long et al. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. InECCV, 2022

    2022

  8. [8]

    Sparis: Neural implicit surface reconstruction of indoor scenes from sparse views

    Yulun Wu, Han Huang, Wenyuan Zhang, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. Sparis: Neural implicit surface reconstruction of indoor scenes from sparse views. InProceedings of the AAAI Conference on Artificial Intelligence, 2025

    2025

Show all 53 references
  1. [9]

    Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction.Advances in neural informa- tion processing systems, 35:25018–25032, 2022

    Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction.Advances in neural informa- tion processing systems, 35:25018–25032, 2022. L. Yang et al.:Preprint submitted to...

    2022

  2. [10]

    Neuris:Neuralreconstruc- tion of indoor scenes using normal priors

    JiepengWang,PengWang,XiaoxiaoLong,ChristianTheobalt,Taku Komura,LingjieLiu,andWenpingWang. Neuris:Neuralreconstruc- tion of indoor scenes using normal priors. InEuropean Conference on Computer Vision, pages 139–155. Springer, 2022

    2022

  3. [11]

    Psdf: Prior-driven neural implicit surface learning for multi-view recon- struction.IEEETransactionson VisualizationandComputer Graph- ics, pages 1–16, 2024

    Wanjuan Su, Chen Zhang, Qingshan Xu, and Wenbing Tao. Psdf: Prior-driven neural implicit surface learning for multi-view recon- struction.IEEETransactionson VisualizationandComputer Graph- ics, pages 1–16, 2024. doi: 10.1109/TVCG.2024.3444035

    2024

  4. [12]

    Global and hierarchical geometry consistency priors for few-shot nerfs in indoor scenes

    XiaotianSun,QingshanXu,XinjieYang,YuZang,andChengWang. Global and hierarchical geometry consistency priors for few-shot nerfs in indoor scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20530– 20539, June 2024

    2024

  5. [13]

    Pmvc: Promoting multi-view consistency for 3d scene reconstruction

    Chushan Zhang, Jinguang Tong, Tao Jun Lin, Chuong Nguyen, and Hongdong Li. Pmvc: Promoting multi-view consistency for 3d scene reconstruction. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3678–3688, 2024

    2024

  6. [14]

    ND-SDF: Learning normal deflection fields for high-fidelity indoor reconstruction

    Ziyu Tang, Weicai Ye, Yifan Wang, Di Huang, Hujun Bao, Tong He, and Guofeng Zhang. ND-SDF: Learning normal deflection fields for high-fidelity indoor reconstruction. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?i...

    2025

  7. [15]

    Multiview neural surface reconstruc- tion by disentangling geometry and appearance

    Lior Yariv, Yoni Kasten, et al. Multiview neural surface reconstruc- tion by disentangling geometry and appearance. InNeurIPS, 2020

    2020

  8. [16]

    InProceedingsoftheIEEE/CVF conference on computer vision and pattern recognition, pages 3504– 3515, 2020

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger.Differentiablevolumetricrendering:Learningimplicit3drep- resentationswithout3dsupervision. InProceedingsoftheIEEE/CVF conference on computer vision and pattern recognition, pages 3504– 3515, 2020

    2020

  9. [17]

    Neuralangelo: High- fidelity neural surface reconstruction

    ZhaoshuoLi,ThomasMüller,AlexEvans,RussellHTaylor,Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High- fidelity neural surface reconstruction. InIEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2023

    2023

  10. [18]

    HF-neus: Im- proved surface reconstruction using high-frequency details

    Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. HF-neus: Im- proved surface reconstruction using high-frequency details. In Al- ice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Information Processing Systems, 2022. URLhttps://openreview...

    2022

  11. [19]

    Voxurf: Voxel-based efficient and accurate neural surface reconstruction

    Tong Wu et al. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. InICLR, 2023

    2023

  12. [20]

    Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices

    Radu Rosu et al. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. InCVPR, 2023

    2023

  13. [21]

    Bakedsdf: Meshing neural sdfs for real-time view synthesis

    Lior Yariv et al. Bakedsdf: Meshing neural sdfs for real-time view synthesis. InSIGGRAPH, 2023

    2023

  14. [22]

    Streetsurf: Extending multi-view implicit surface reconstruction to street views

    Haoyu Guo et al. Streetsurf: Extending multi-view implicit surface reconstruction to street views. InICLR, 2024

    2024

  15. [23]

    S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields

    Zehao Xie et al. S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields. InICCV, 2023

    2023

  16. [24]

    ShengYe,YubinHu,MatthieuLin,Yu-HuiWen,WangZhao,Yong- Jin Liu, and Wenping Wang. Indoor Scene Reconstruction With Fine-GrainedDetailsUsingHybridRepresentationandNormalPrior Enhancement .IEEE Transactions on Visualization & Computer Graphics, 31(09):5275–5287, 2025. ISSN 1941-0506

    2025

  17. [25]

    Ref-neus: Ambiguity-reduced neural implicit surface learning for indoor scenes with reflection

    Wenhang Ge, Tao Hu, et al. Ref-neus: Ambiguity-reduced neural implicit surface learning for indoor scenes with reflection. InICCV, 2023

    2023

  18. [26]

    Nc-sdf: Enhancing indoor scene reconstruction using neural sdfs with view-dependent normal compensation

    Ziyi Chen, Xiaolong Wu, and Yu Zhang. Nc-sdf: Enhancing indoor scene reconstruction using neural sdfs with view-dependent normal compensation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5155–5165, 2024

    2024

  19. [27]

    Neusample: Neural sample importance for neural radiance fields

    Jiemin Fang et al. Neusample: Neural sample importance for neural radiance fields. InCVPR, 2024

    2024

  20. [28]

    Springer Na- ture Switzerland, December 2024

    Gonçalo Dias Pais, Valter Piedade, Moitreya Chatterjee, Marcus Greiff, and Pedro Miraldo.A Probability-Guided Sampler for Neural Implicit Surface Rendering, page 164–182. Springer Na- ture Switzerland, December 2024. ISBN 9783031729133. doi: 10.1007/978-3-031-72913-3_10. URLht...

    2024 doi

  21. [29]

    Ray- distance volume rendering for neural scene reconstruction.arXiv preprint arXiv:2408.15524, 2024

    Ruihong Yin, Yunlu Chen, Sezer Karaoglu, and Theo Gevers. Ray- distance volume rendering for neural scene reconstruction.arXiv preprint arXiv:2408.15524, 2024

    2024 arXiv

  22. [30]

    Neural 3d scene reconstruc- tion with the manhattan-world assumption

    Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruc- tion with the manhattan-world assumption. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5511–5520, June 2022

    2022

  23. [31]

    Neural 3d scene reconstruction with indoor planar priors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):6355–6366, 2024

    Xiaowei Zhou, Haoyu Guo, Sida Peng, Yuxi Xiao, Haotong Lin, Qianqian Wang, Guofeng Zhang, and Hujun Bao. Neural 3d scene reconstruction with indoor planar priors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):6355–6366, 2024. doi: 10.1109/TPAMI.2024.3379833

    2024

  24. [33]

    Repurposing diffusion-based image generators for monocular depth estimation

    Bingxin Ke et al. Repurposing diffusion-based image generators for monocular depth estimation. InCVPR, 2024

    2024

  25. [34]

    LiheYangetal.Depthanythingv2.arXivpreprintarXiv:2406.09414, 2024

    2024 arXiv

  26. [35]

    Wonder3d: Single image to 3d using cross- domain diffusion

    Xiaoxiao Long et al. Wonder3d: Single image to 3d using cross- domain diffusion. InCVPR, 2024

    2024

  27. [36]

    Wildfusion:Learning3d-awarelatentdiffusion for view synthesis in the wild

    NicholasKolkinetal. Wildfusion:Learning3d-awarelatentdiffusion for view synthesis in the wild. InCVPR, 2024

    2024

  28. [37]

    Nerfprior: Learning neural radiance field as a prior for indoor scene reconstruction

    Wenyuan Zhang, Emily Yue-ting Jia, Junsheng Zhou, Baorui Ma, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. Nerfprior: Learning neural radiance field as a prior for indoor scene reconstruction. In ProceedingsoftheComputerVisionandPatternRecognitionConfer- ence, pages 11317–11327, 2025

    2025

  29. [38]

    Sat-dn: Implicit surface reconstruction from multi-view satel- lite images with depth and normal supervision.arXiv preprint arXiv:2502.08352, 2025

    Tianle Liu, Shuangming Zhao, Wanshou Jiang, and Bingxuan Guo. Sat-dn: Implicit surface reconstruction from multi-view satel- lite images with depth and normal supervision.arXiv preprint arXiv:2502.08352, 2025

    2025 arXiv

  30. [39]

    Geo- metricprior-guidedneuralimplicitsurfacereconstructioninthewild,

    Lintao Xiang, Hongpei Zheng, Bailin Deng, and Hujun Yin. Geo- metricprior-guidedneuralimplicitsurfacereconstructioninthewild,

  31. [40]

    Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations.2021 International Conference on 3D Vision (3DV), pages 972–981, 2021

    Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno- Noguer. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations.2021 International Conference on 3D Vision (3DV), pages 972–981, 2021. URLhttps://api. semanticscholar.org/CorpusID...

    2021

  32. [41]

    Cf-nerf: Camera parameter free neural radiance fields with probabilistic uncertainty

    Yuxin Shen et al. Cf-nerf: Camera parameter free neural radiance fields with probabilistic uncertainty. InECCV, 2022

    2022

  33. [42]

    Activenerf: Learning where to see with uncertainty estimation

    Xuran Pan et al. Activenerf: Learning where to see with uncertainty estimation. InECCV, 2022

    2022

  34. [43]

    Activermap:Radiancefieldforactivemappingand planning

    TianruoGaoetal. Activermap:Radiancefieldforactivemappingand planning. InICRA, 2023

    2023

  35. [44]

    Fisherrf: Active view selection via fisher informa- tion in neural radiance fields

    Yuang Jiang et al. Fisherrf: Active view selection via fisher informa- tion in neural radiance fields. InCVPR, 2024

    2024

  36. [45]

    Bayessdf: Surface-based laplacian uncertainty estima- tion for 3d geometry with neural signed distance fields, 2025

    Rushil Desai. Bayessdf: Surface-based laplacian uncertainty estima- tion for 3d geometry with neural signed distance fields, 2025. URL https://arxiv.org/abs/2507.06269

    2025 arXiv

  37. [46]

    InCVPR, 2024

    JingyangZhangetal.Regsdf:Regularizingneuralsurfacereconstruc- tion with surface-based geometric consistency. InCVPR, 2024

    2024

  38. [47]

    Estimating andexploitingthealeatoricuncertaintyinsurfacenormalestimation

    Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating andexploitingthealeatoricuncertaintyinsurfacenormalestimation. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2021. doi: 10.1109/iccv48922.2021.01289. URLhttp: //dx.doi.org/10.1109/iccv48922...

    2021

  39. [48]

    Improving neural implicit surfaces geometry with patch warping

    François Darmon, Bénédicte Bascle, et al. Improving neural implicit surfaces geometry with patch warping. InCVPR, 2022

    2022

  40. [49]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    AngelaDai,AngelX.Chang,ManolisSavva,MaciejHalber,Thomas Funkhouser, and Matthias Niessner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jul 2017. doi: 10.1109/cvpr.2017.261. URLhttp://d...

    2017 doi

  41. [50]

    Julian Straub, ThomasJ. Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob Engel, Raul Mur-Artal, CarlYuheng Ren, Sandeep Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Ying Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyle...

    2019

  42. [51]

    Pixelwiseviewselectionforunstructuredmulti-view stereo

    Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and MarcPollefeys. Pixelwiseviewselectionforunstructuredmulti-view stereo. InComputerVision–ECCV2016:14thEuropeanConference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501–518. Spri...

    2016

  43. [52]

    Helix- surf: A robust and efficient neural implicit surface learning of indoor sceneswithiterativeintertwinedregularization

    ZhihaoLiang,ZhangjinHuang,ChangxingDing,andKuiJia. Helix- surf: A robust and efficient neural implicit surface learning of indoor sceneswithiterativeintertwinedregularization. InProceedingsofthe IEEE/CVFConferenceonComputerVisionandPatternRecognition, pages 13165–13174, 2023

    2023

  44. [53]

    Learning a room with the occ-sdf hybrid: Signed distance function mingled with occupancy aids scene repre- sentation

    Xiaoyang Lyu, Peng Dai, Zizhang Li, Dongyu Yan, Yi Lin, Yifan Peng, and Xiaojuan Qi. Learning a room with the occ-sdf hybrid: Signed distance function mingled with occupancy aids scene repre- sentation. InProceedings of the IEEE/CVF international conference on computer vision,...

    2023

  45. [2025]

    URLhttps://arxiv.org/abs/2505.07373

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.