Accuracy Improvements for Convolutional and Differential Distance Function Approximations

Alexander Belyaev; Pierre-Alain Fayolle

REVIEW 3 major objections 4 minor 21 references

Accuracy Improvements for Convolutional and Differential Distance Function Approximations

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper derives first- and second-order corrections to heat-method distance estimates by differentiating the screened Poisson equation with respect to its decay parameter, and reports that the normalized second-order formula is the most…

desk verdict A clean, honest incremental letter: new differential Taylor formulas for distance estimation with a genuinely unproven error analysis, so the empirical gains are plausible but not established. read the letter →

arxiv 2412.09200 v1 pith:ODO2YYAS submitted 2024-12-12 math.NA cs.CVcs.NA

Alexander Belyaev , Pierre-Alain Fayolle This is my paper · ORCID

classification math.NAcs.CVcs.NA MSC 65D18

keywords distancefunctionapproximationheatmethodscreenedPoissonequationLaplace'sTaylorextrapolationconvolutionaltransformeikonalgradientnormalization

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

This paper is about a basic geometry question: given a region, estimate for each interior point how far it is from the boundary. It offers two accuracy upgrades. For convolutional distance transforms, it blends the log-sum-exp and soft-min approximations using weights derived from Laplace's method. For differential methods, it differentiates the screened Poisson equation with respect to the decay parameter and uses the first- and second-order Taylor terms as new distance estimates. The paper's numerical evidence on 2D binary shapes says the normalized second-order estimate is the most accurate, with a flat error curve that makes it less sensitive to the choice of parameter.

What carries the argument

The machinery is the $\lambda$-derivative of the screened Poisson solution. Differentiating the boundary-integral representation with respect to $\lambda$ turns the logarithm of the heat kernel into a new distance estimator, $-v'_\lambda/v$, the continuous analogue of the identity that turns log-sum-exp into a soft minimum. Taylor expansion of $u(x;s)$ in $s = 1/\lambda$ around $s = 0$ then produces the first-order correction and the second-order correction. The gradient-normalization step completes the pipeline: normalize the gradient of the approximate distance and solve a Poisson equation to project the result back onto the space of distance-like functions.

What would settle it

Use a domain with an exactly known distance function, such as a disk or annulus, solve the screened Poisson equation and its $\lambda$-derivative equations numerically for many values of $\lambda$, and compare the three formulas with the exact distance; if the second-order formula does not consistently beat the first-order one as $\lambda$ grows, or if the improvement disappears when the heuristic gradient-normalization step is removed, the central accuracy claim would be disproved.

Watch

Extended reading notes

Core claim

The central claim is that for a bounded planar domain $\Omega$, the distance to the boundary can be estimated more accurately than the heat method's plain log formula by differentiating the screened Poisson solution with respect to $\lambda$. With $v$ solving $-\Delta v + \lambda^2 v = 0$ in $\Omega$ and $v = 1$ on $\partial\Omega$, the heat method uses $\mathrm{dist} \approx -(1/\lambda) \log v$. The paper's first correction is $\mathrm{dist} \approx -v'_\lambda/v$, where $v'_\lambda$ solves the equation obtained by differentiating once, and the second correction adds the term $-\frac{\lambda}{2}[v''_{\lambda\lambda}/v - (v'_\lambda/v)^2]$. These formulas arise from a Taylor expansion of $u = -(1/\lambda)\log v$ in $s = 1/\lambda$ near $s = 0$, and numerically, after the standard gradient-normalization projection step, the second-order formula gives the smallest $L^2$ and $L^\infty$ errors on all tested 2D shapes. The paper presents the formulas as practical improvements while noting that their rigorous asymptotic justification is not yet established.

Load-bearing premise

The new differential formulas assume that the solution of the regularized equation is smooth enough in $1/\lambda$ that its first- or second-order Taylor polynomial is accurate with no bound on the leftover error; the paper explicitly acknowledges that this assumption currently lacks mathematical justification.

Editorial extensions

If this is right

On 2D shapes, both new differential formulas reduce $L^2$ and $L^\infty$ distance errors relative to the heat method estimate; with gradient normalization, the second-order formula is the most accurate and the first-order formula is second.
The $L^2$ error of the second-order estimate stays nearly flat as $t = 1/\lambda^2$ varies from 0.2 to 10, so the second-order estimate is far less sensitive to the user-chosen parameter than the plain heat method.
The Laplace-blended convolutional estimate, formed by weighting the LogConv and SoftMin formulas with the weights derived from Laplace asymptotics, improves accuracy over either convolutional formula taken alone on the tested shape.
Where the regularized solution is already the exact distance function, the three formulas coincide, so the observed improvements are concentrated in regions where the heat solution departs from the true distance, such as near skeleton branch points.

Reading between the lines

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

The same $\lambda$-differentiation step could be applied to any distance estimator expressible as a Laplace-type integral, not just the screened Poisson equation; testing this on other integral kernels would show how general the mechanism is.
The flat error curve for the second-order formula suggests the second-order term is removing a systematic bias rather than random noise; a bias-variance decomposition on a domain with known exact distance could confirm this and guide the choice of $\lambda$.
A rigorous version of the Taylor extrapolation would need a bound on the remainder of the expansion; for convex domains with smooth boundaries, Green's function asymptotics might give an explicit $O(1/\lambda^3)$ estimate that would settle when the second-order formula converges.
The reported tests are planar, so the practical payoff for surfaces and geodesics remains open; if the formulas extend, their parameter insensitivity would be a stronger advantage there than the small absolute accuracy gain seen in 2D.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Colleague,

Quick read of arXiv:2412.09200. The paper does two genuinely new things: it blends LogConv and SoftMin distance estimates using Laplace asymptotics with weights (5), and it derives differential distance formulas (15) and (19) by differentiating the screened Poisson equation with respect to λ. The algebra linking the convolutional identity (8) to the differential formula (14) is clean, and the connection to Laplace's method in [10] is worth having. The authors are also admirably honest in Section VI that the Taylor extrapolation schemes lack rigorous mathematical justification.

That honesty matches the main soft spot. The load-bearing step is the Taylor extrapolation in s = 1/λ behind (17)/(18). The remainder is never bounded, and the stress-test note is right that this is not merely formal: the distance function is only Lipschitz at the medial axis, so singular behavior is plausible, and the recommended t = 5 gives s ≈ 2.24, far from the s → 0 limit. So the accuracy gain of normalized (19) over the heat method is an empirical observation on a handful of 2D binary images, not a consequence of the Taylor argument. The experiments themselves are not misleading: error curves are shown over 0.2 ≤ t ≤ 10, normalized versions are compared fairly against the normalized heat method, and the conclusions are phrased as performance on tested shapes. Still, no code is released (promised on acceptance), no error bars, no other domain types, so the generality is unproven.

The convolutional blending part is on firmer footing since it rests on established Laplace asymptotics, though the constant K is heuristic. That part is a modest but sensible incremental contribution.

Overall, this is a compact, readable letter that does not oversell. A serious referee should engage with it—the formal manipulations are solid, the empirical claims are honestly scoped, and the main request should be for a remainder estimate or at least a broader benchmark with released code before the accuracy improvements are treated as established.

My recommendation: send it to review, not desk reject.

Referee Report

3 major / 4 minor

Summary. This paper addresses the problem of approximating the distance to the boundary of a bounded domain. It proposes two types of improvements. First, for convolutional distance transforms, it uses Laplace's method to derive a blend (Eq. (5)) of the log-sum-exp approximation (6) and the soft-min approximation (7), with a free constant K. Second, for differential (heat-method-type) approximations, it derives first- and second-order corrections (15) and (19) by differentiating the screened Poisson equation (9) with respect to λ and interpreting the result as a Taylor extrapolation in s=1/λ of the solution to the rescaled equation (11). The corrected schemes are then combined with the standard gradient-normalization step of the heat method. Numerical experiments on 2D binary shapes suggest that the normalized second-order scheme (19) outperforms the standard heat method (12). The paper is clearly written, and the algebraic steps from (9) through (19) are internally consistent.

Significance. If the accuracy gains were rigorously established, the paper would make a useful contribution to distance estimation: the proposed modifications are simple, require only one or two additional Poisson solves, and could improve the accuracy of heat-method-based distance computations in applications. The Laplace-method perspective connecting convolutional distance transforms to asymptotic expansions is also instructive. However, the central theoretical claim rests on a Taylor extrapolation for which no remainder estimate is provided, and the numerical validation is limited to a few 2D shapes without error bars or sensitivity analysis. The paper cannot currently be regarded as establishing the claimed superiority of (19) over the standard heat method, although the empirical evidence, especially the error plots in Fig. 2, is suggestive.

major comments (3)

[Section VI and Eqs. (17)-(19)] The Taylor extrapolation underlying (15) and (19) is the load-bearing step of the paper, but it is not rigorously justified. Equations (17) and (18) assume that u(s), the solution of (11) with s=1/λ, is sufficiently smooth in s near 0 and that the remainder of the Taylor expansion at s=1/λ is negligible uniformly on Ω. No bound on the remainder is provided, and Section VI explicitly admits that these schemes 'are not supported by truly rigorous mathematical results.' This is not a purely formal concern: as s→0, u(s) converges to the distance function, which is only Lipschitz at the medial axis, so the correction terms u'_s and u''_ss can be singular there. Furthermore, the parameter values used in the experiments (e.g., t=5, so s≈2.24) are far from the s→0 limit in which the asymptotic argument would apply. Thus the theoretical derivation does not, by itself, establish that (19) improves on (12). I recommend either providing a uniform remainder estimate under explicit regularity assumptions, or substantially expanding the numerical evidence so that the accuracy claim is backed by convergence tests as s→0 and by experiments on domains with medial-axis singularities.
[Section V, Figs. 2 and 3] The numerical evaluation of the central claim is too narrow. The quantitative error plots in Fig. 2 are for a single 2D shape, and the error maps in Fig. 3 are presented without numerical error values or comparisons across a dataset. There are no error bars and no statistical summary over multiple shapes or boundary discretizations, so the statement that the normalized version of (19) 'demonstrates the best performance' is not supported with the usual precision expected of a numerical paper. Additionally, the paper does not compare the proposed schemes against other recent distance estimators (e.g., the ADMM-based scheme of [14], which is from the same authors) beyond the heat method. Since the main claim is empirical accuracy improvement, the authors should provide a more systematic experimental study, including error statistics for a set of reference shapes and a convergence study with respect to the discretization.
[Section III, Eq. (5)] The convolutional blending formula (5) depends on a constant K that is not determined analytically. The paper sets K=0.1 heuristically and reports only one experiment for this value. Because the weights α and β in (5) are monotone functions of K, the observed accuracy improvement of the blend over the individual approximations (6) and (7) could depend critically on this free parameter. A sensitivity analysis over a range of K values, or an adaptive way of estimating K, is needed to establish the robustness of the convolutional improvement. As written, the experimental support for this contribution is a single shape with K fixed, which is insufficient.

minor comments (4)

[Section II, Eq. (5)] The definitions of α and β in (5) use K, but K is introduced only as an unspecified positive constant in (4); the text later states that K is heuristic, but this dual role should be made explicit in the notation.
[Section IV, paragraph after Eq. (13)] The notation v'_λ and v''_λλ is used to denote derivatives with respect to λ; please state this explicitly at first use, as the same notation could be confused with a spatial derivative.
[Section V, paragraph on gradient normalization] There is a typo in 'dist(x∂Ω': the argument should be 'dist(x, ∂Ω)'.
[Section VI, reproducibility statement] The statement that the Matlab implementations 'will be made available upon acceptance' makes reproducibility conditional; please provide a permanent repository or include the code at submission time.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the Taylor extrapolations and Laplace blends are independent PDE manipulations; the admitted lack of remainder bounds is a rigor limitation, not a self-referential input.

full rationale

The derivation chain is self-contained and none of the claimed improvements reduces to its inputs by construction. The convolutional combination uses Laplace asymptotics (2) and (4) with a heuristic constant K; K is a free parameter rather than a fitted value, and the weights in (5) are derived by canceling leading asymptotic terms. The numerical improvement is an empirical observation, not a consequence forced by the formulas. The differential formulas (15) and (19) are obtained by differentiating the screened Poisson solution with respect to λ and by Taylor extrapolation of u(s) in s=1/λ. They are related to (12) only through the algebraic identity -v'_λ/v = u + λ u'_λ, which is a reformulation, not an assumption of the conclusion. The paper explicitly states that (12), (15), and (19) coincide only when u is already the exact distance function; that statement is used as a consistency check, not as a derivation of the accuracy claim. The admitted lack of rigorous remainder bounds for the Taylor extrapolations, noted in Section VI, is a correctness and robustness limitation: it weakens the support for the claimed accuracy improvements but does not make the claim presuppose itself. No load-bearing self-citation is present: [11] and [12] are external mathematical results, and the authors' own prior work [14] is cited only as a source of Matlab scripts.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claims rest on standard asymptotics and PDE tools plus two unproven or heuristic choices. The Taylor extrapolation lacks remainder control, and the convolutional blend needs a constant K that is set to 0.1 without sensitivity analysis. No new physical or mathematical entities are introduced.

free parameters (2)

K in the blend weights (5) = 0.1 (heuristic choice)
The weights alpha and beta in Eq (5) depend on K, which in principle can be derived from phi(p) and its derivatives at the minimizer but is unknown for a general shape. The paper sets K=0.1 for all experiments and does not study sensitivity.
lambda (via t=1/lambda^2) for each scheme = t=1 for (12), t=5 for (19), t=2 for the visualization; Fig 2 sweeps t in [0.2, 10]
Accuracy depends on the time or scale parameter. The choice t=5 for (19) is admitted to be nonoptimal, so the favorable comparison is partly parameter-dependent, though (19) shows a flat error curve over a wide range.

assumptions (3)

domain assumption Laplace's method (1)-(4) applies to the boundary integral with phi(p)=||x-p|| for x inside a bounded smooth domain.
Section III implicitly assumes a unique, nondegenerate global minimizer of ||x-p|| on the boundary and sufficient smoothness of the kernel. Distance functions in nonconvex domains with multiple closest boundary points may violate this assumption.
ad hoc to paper The function u(s) solving (11), with s=1/lambda, is smooth enough in s near 0 and the first- or second-order Taylor polynomial at s=1/lambda approximates u(0) well.
Equations (17)-(19) use Taylor extrapolation with no remainder estimates. The authors explicitly call the lack of mathematical justification a limitation in Section VI.
domain assumption After computing a distance estimate w, normalizing the gradient and solving the Poisson problem Delta w_n = div(nabla w / |nabla w|) improves accuracy.
This step is inherited from the heat method [12] and is applied to every differential variant in Section V. The reported gains depend on this normalization procedure.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Accuracy Improvements for Convolutional and Differential Distance Function Approximations." pith.science (2026). https://pith.science/paper/ODO2YYAS

@misc{pith2026241209200,
  author       = {Pith},
  title        = {Pith review of: Accuracy Improvements for Convolutional and Differential Distance Function Approximations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODO2YYAS}},
  note         = {Machine review of arXiv:2412.09200}
}

read the original abstract

Given a bounded domain, we deal with the problem of estimating the distance function from the internal points of the domain to the boundary of the domain. Convolutional and differential distance estimation schemes are considered and, for both the schemes, accuracy improvements are proposed and evaluated. Asymptotics of Laplace integrals and Taylor series extrapolations are used to achieve the improvements.

Figures

Figures reproduced from arXiv: 2412.09200 by the authors.

**Figure 1.** (a) A 2-D shape (binary image) and its 1-D slice which are used to evaluate the proposed combination of the LogConv and SoftMin approximations [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Left: 1-D slices of the distance function approximations (12), (15), and (19). Middle: dependencies of the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Distance function error maps for several binary shapes (binary images). Top row: the heat method [12] with [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

[10]

Laplace meets Moreau: Smooth approximation to infimal convolutions using Laplace’s method,

R. J. Tibshirani, S. W. Fung, H. Heaton, and S. Osher, “Laplace meets Moreau: Smooth approximation to infimal convolutions using Laplace’s method,” arXiv preprint arXiv:2406.02003 , 2024

arXiv 2024
[14]

An ADMM-based scheme for distance function approximation,

A. Belyaev and P.-A. Fayolle, “An ADMM-based scheme for distance function approximation,” Numerical Algorithms , vol. 84, pp. 983–996, 2020

work page 2020
[1]

ReSDF: Redistancing implicit surfaces using neural networks,

Y . Park, C. h. Song, J. Hahn, and M. Kang, “ReSDF: Redistancing implicit surfaces using neural networks,” Journal of Computational Physics, vol. 502, 2024

work page 2024
[2]

A closest point method for PDEs on manifolds with interior boundary conditions for geometry processing,

N. King, H. Su, M. Aanjaneya, S. Ruuth, and C. Batty, “A closest point method for PDEs on manifolds with interior boundary conditions for geometry processing,” ACM Transactions on Graphics , 2024

work page 2024
[3]

A framework for solving parabolic partial differential equations on discrete domains,

L. Mattos Da Silva, O. Stein, and J. Solomon, “A framework for solving parabolic partial differential equations on discrete domains,” ACM Transactions on Graphics , 2024

work page 2024
[4]

A heat diffusion perspective on geodesic preserving dimensionality reduction,

G. Huguet, A. Tong, E. De Brouwer, Y . Zhang, G. Wolf, I. Adelstein, and S. Krishnaswamy, “A heat diffusion perspective on geodesic preserving dimensionality reduction,” in 37th Conference on Neural Information Processing Systems (NeurIPS 2023) , 2023

work page 2023
[5]

A convex optimization framework for regularized geodesic distances,

M. Edelstein, N. Guillen, J. Solomon, and M. Ben-Chen, “A convex optimization framework for regularized geodesic distances,” in ACM SIGGRAPH 2023 Conference Proceedings , 2023, pp. 2:1–2:11

work page 2023
[6]

A heat method for generalized signed distance,

N. Feng and K. Crane, “A heat method for generalized signed distance,” ACM Transactions on Graphics , vol. 43, no. 4, pp. 92:1–92:16, 2024

work page 2024

Show all 21 references

[7]

Stabilizing the optimization of neural signed distance functions and finer shape rep- resentation,

H. Yang, Y . Sun, G. Sundaramoorthi, and A. Yezzi, “Stabilizing the optimization of neural signed distance functions and finer shape rep- resentation,” in 37th Conference on Neural Information Processing Systems (NeurIPS 2023) , 2023

2023
[8]

A fast algorithm for all-pairs-shortest-paths suitable for neural networks,

M. Meister, “A fast algorithm for all-pairs-shortest-paths suitable for neural networks,” arXiv preprint arXiv:2308.07403 , 2023

2023 arXiv
[9]

Fast convolutional distance transform,

C. Karam, K. Sugimoto, and K. Hirakawa, “Fast convolutional distance transform,” IEEE Signal Processing Letters , vol. 26, no. 6, pp. 853–857, 2019

2019
[11]

On the behavior of the fundamental solution of the heat equation with variable coefficients,

S. R. S. Varadhan, “On the behavior of the fundamental solution of the heat equation with variable coefficients,” Comm. Pure Appl. Math. , vol. 20, pp. 431–455, 1967

1967
[12]

Geodesics in heat: A new approach to computing distance based on heat flow,

K. Crane, C. Weischedel, and M. Wardetzky, “Geodesics in heat: A new approach to computing distance based on heat flow,” ACM Transactions on Graphics , vol. 32, pp. 152:1–152:11, 2013

2013
[13]

The heat method for distance computation,

——, “The heat method for distance computation,” Communications of the ACM, vol. 60, no. 112, pp. 90–99, 2017

2017
[15]

Survey of meshless and generalized finite element methods: A unified approach,

I. Babu ˇska, U. Banerjee, and J. E. Osborn, “Survey of meshless and generalized finite element methods: A unified approach,”Acta Numerica, vol. 12, pp. 1–125, 2003

2003
[16]

Reconstruction of solid models from oriented point sets,

M. Kazhdan, “Reconstruction of solid models from oriented point sets,” in Proceedings of the third Eurographics symposium on Geometry processing, 2005, pp. 73–82

2005
[17]

Poisson surface recon- struction,

M. Kazhdan, M. Bolitho, and H. Hoppe, “Poisson surface recon- struction,” in Proceedings of the fourth Eurographics symposium on Geometry processing, vol. 7, no. 4, 2006

2006
[18]

Screened Poisson surface reconstruction,

M. Kazhdan and H. Hoppe, “Screened Poisson surface reconstruction,” ACM Transactions on Graphics , vol. 32, no. 3, pp. 1–13, 2013

2013
[19]

Geodesic via asymmetric heat diffusion based on Finsler metric,

F. Yang, L. Chai, D. Chen, and L. Cohen, “Geodesic via asymmetric heat diffusion based on Finsler metric,” in Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Revised Selected Papers, Part V 14 , 2019, pp. 371–386

2018
[20]

Finsler-Laplace-Beltrami operators with application to shape analysis,

S. Weber, T. Dag `es, M. Gao, and D. Cremers, “Finsler-Laplace-Beltrami operators with application to shape analysis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 3131–3140

2024
[21]

Walk on spheres for PDE-based path planning,

R. I. C. Muchacho and F. T. Pokorny, “Walk on spheres for PDE-based path planning,” arXiv preprint arXiv:2406.01713 , 2024

2024 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [10]

Laplace meets Moreau: Smooth approximation to infimal convolutions using Laplace’s method,

R. J. Tibshirani, S. W. Fung, H. Heaton, and S. Osher, “Laplace meets Moreau: Smooth approximation to infimal convolutions using Laplace’s method,” arXiv preprint arXiv:2406.02003 , 2024

arXiv 2024

[2] [14]

An ADMM-based scheme for distance function approximation,

A. Belyaev and P.-A. Fayolle, “An ADMM-based scheme for distance function approximation,” Numerical Algorithms , vol. 84, pp. 983–996, 2020

work page 2020

[3] [1]

ReSDF: Redistancing implicit surfaces using neural networks,

Y . Park, C. h. Song, J. Hahn, and M. Kang, “ReSDF: Redistancing implicit surfaces using neural networks,” Journal of Computational Physics, vol. 502, 2024

work page 2024

[4] [2]

A closest point method for PDEs on manifolds with interior boundary conditions for geometry processing,

N. King, H. Su, M. Aanjaneya, S. Ruuth, and C. Batty, “A closest point method for PDEs on manifolds with interior boundary conditions for geometry processing,” ACM Transactions on Graphics , 2024

work page 2024

[5] [3]

A framework for solving parabolic partial differential equations on discrete domains,

L. Mattos Da Silva, O. Stein, and J. Solomon, “A framework for solving parabolic partial differential equations on discrete domains,” ACM Transactions on Graphics , 2024

work page 2024

[6] [4]

A heat diffusion perspective on geodesic preserving dimensionality reduction,

G. Huguet, A. Tong, E. De Brouwer, Y . Zhang, G. Wolf, I. Adelstein, and S. Krishnaswamy, “A heat diffusion perspective on geodesic preserving dimensionality reduction,” in 37th Conference on Neural Information Processing Systems (NeurIPS 2023) , 2023

work page 2023

[7] [5]

A convex optimization framework for regularized geodesic distances,

M. Edelstein, N. Guillen, J. Solomon, and M. Ben-Chen, “A convex optimization framework for regularized geodesic distances,” in ACM SIGGRAPH 2023 Conference Proceedings , 2023, pp. 2:1–2:11

work page 2023

[8] [6]

A heat method for generalized signed distance,

N. Feng and K. Crane, “A heat method for generalized signed distance,” ACM Transactions on Graphics , vol. 43, no. 4, pp. 92:1–92:16, 2024

work page 2024