$p$-Poisson surface reconstruction in curl-free flow from point clouds

The aim of this paper is the reconstruction of a smooth surface from an unorganized point cloud sampled by a closed surface, with the preservation of geometric shapes, without any further information other than the point cloud. Implicit neural representations (INRs) have recently emerged as a promising approach to surface reconstruction. However, the reconstruction quality of existing methods relies on ground truth implicit function values or surface normal vectors. In this paper, we show that proper supervision of partial differential equations and fundamental properties of differential vector fields are sufficient to robustly reconstruct high-quality surfaces. We cast the $p$-Poisson equation to learn a signed distance function (SDF) and the reconstructed surface is implicitly represented by the zero-level set of the SDF. For efficient training, we develop a variable splitting structure by introducing a gradient of the SDF as an auxiliary variable and impose the $p$-Poisson equation directly on the auxiliary variable as a hard constraint. Based on the curl-free property of the gradient field, we impose a curl-free constraint on the auxiliary variable, which leads to a more faithful reconstruction. Experiments on standard benchmark datasets show that the proposed INR provides a superior and robust reconstruction. The code is available at \url{https://github.com/Yebbi/PINC}.

Paper

Similar papers

Peer review

Reviewer 2pnE7/10 · confidence 5/52023-07-02

Summary

This content seems interesting. I like that the authors gave considerable thought to improving surface reconstruction from the perspective of vector field processing. The paper targets two challenging problems in surface reconstruction: (1) removing the requirement for surface normal, and (2) improving the over-smoothness. However, there are technical details that need to be addressed before I can accept it.

Strengths

(1) The problems are challenging. (2) Methodology is novel. (3) Improvements are observed.

Weaknesses

In a p-Poisson equation, you have the p-Laplacian instead of the Laplacian. This is equivalent to giving weights to the gradient of an implicit function. And this weight is based on the magnitude of the gradient itself. The geometric intuition behind this weight is not clearly stated.

Questions

major: (1) Add a comparison for the learned vector field. From equation (8), your G plays the role of the vector field derived from the oriented point cloud in PoissonRecon. It might be beneficial to compare G with the vector field formulated by surface normal. (2) In section 3.2, you mentioned the goal is to progressively enlarge lambda to infinity. If so, have you tried getting rid of the first term of equation (8)? I don’t think the method will fall apart. You would only be downgrading from ScreenedPoissonRecon to PoissonRecon. (3) The result in Figure 3 is a bit confusing. What’s your sampling strategy? Are you sampling all the vertices from the mesh? Because from the Screwstar, it seems you are not sampling uniformly. Could that partially be the reason the SIREN has a broken surface? Can you try to uniformly sample the star and see if there’s any improvement in the completeness? (4) The result in Figure 7 (a) had some clear artifacts. Why did that happen? When p = 2, your energy looks almost the same as ScreenedPoissonRecon. However, when I ran ScreenedPoissonRecon on this exact model, I did not see this artifact. minor: (1) The statement “irrotational flow” only appeared once in the title. If you want to use the statement, you need to be consistent in the writing. At least mention and explain it in the introduction. (2) In Chapter 4.3, you used the subtitle “Effect of curl-free constraint” twice. I believe you meant to say “Effect of minimal surface area constraint” for the second one.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

yes

Reviewer WkvV5/10 · confidence 5/52023-07-05

Summary

The paper introduces an intriguing approach to surface reconstruction, but further enhancements in terms of completeness and evaluations would strengthen its contribution to the field. Also, the gradient of the SDF acts as an auxiliary network output and incorporated the Poisson equation as a hard constraint. The proposed method was also used to propose a more accurate representation. They perform some experiments on standard benchmark datasets to demonstrate superior and robust reconstruction. In my view, the proposed method cannot achieve the best one numerically on average (table 1).

Strengths

This paper proposes a novel surface reconstruction method using the p-Poisson equation and a curl-free constraint, which is highly interesting. It demonstrates superior performance compared to previous works.

Weaknesses

The authors have submitted a revision of the full paper in the supplementary material, it may be considered a violation of the rules. It may be appropriate to consider resubmitting the paper to another venue due to the violation. There are several incomplete pieces of information. It would be beneficial to include evaluations to measure surface quality, such as normal consistency, as well as provide details on training and inference times.

Questions

1. Your method appears to be computationally intensive. Can you provide information on the training and inference times compared to other methods? 2. While your results look promising, including quantitative results on surface quality would be valuable. 3. It would be unfair to directly compare the proposed method with approaches that utilize normal information. However, conducting experiments and comparisons with methods such as "shape as points" from the URL [https://pengsongyou.github.io/sap] would provide valuable insights and contribute to a more comprehensive evaluation of the proposed method. 4. What are lines 457-460?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

1 poor

Presentation

1 poor

Contribution

3 good

Limitations

The limitation has been roughly discussed in the paper. Poisson-based methods, including the proposed approach, are unable to handle open surfaces.

Reviewer FaEi7/10 · confidence 4/52023-07-06

Summary

This paper considers the problem of reconstruction of a smooth surface from an unorganised point cloud. The proposed approach is based on neural implicit function while without normal information. The main contribution of this work is that they demonstrate that proper supervision of partial differentiable equation and fundamental properties of differential vector fields are enough to reconstruct high-quality surfaces. A novel part is to develop a variable splitting structure by introducing a gradient of the SDF as an auxiliary variable and a curl-free constraint on the auxiliary variable. The experimental results somehow demonstrate the effectiveness on some aspects.

Strengths

I like the idea by introducing auxiliary variable to solve the optimization problem under the framework of neural implicit function. Actually, they all both solve optimization problem. Therefore, the optimization strategy in numerical algorithm can be adopted for neural implicit function based representation. This paper shows a good example in this aspect, and might inspire some interesting along this direction.

Weaknesses

The results shown in Tab. 1 and Tab. 2 are not good enough. Can you explain why the performance is not good enough for different dataset and metrics?

Questions

With additional auxiliary variable, does the optimization take long time to converge? If yes, please list the computation time with more details.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

As listed in the above.

Reviewer ymPA6/10 · confidence 2/52023-07-06

Summary

The paper presents a surface reconstruction method that uses only raw point clouds. It enforces the Poisson surface equation implicitly over the SDF representation of the surface. As a consequence, it obtains smooth surfaces with preserved details without any 3D supervision or apriori knowledge of normals. The experiments show that the performance is comparable to sota methods which require data beyond raw point clouds.

Strengths

The use of p-poisson equation to describe SDF is well-motivated. The use of auxiliary variables relating to the gradient and curl of SDF is interesting and convincingly reduces the computational complexity. Ablation study is a plus. The performance is comparable to the methods that use either 3D supervision or oriented normals. The performance on noisy data is generally better than sota. I think it is due to the fact the method uses only raw point clouds that serves as a benefit here. Normal computation on noisy point clouds can be disproportionately erroneous.

Weaknesses

There is no theoretical motivation/argument provided to choose p-poisson over eikonal equation to describe SDFs. IGR[23] can perform surface reconstruction without normals as well. It is not clear whether the authors used the Normals while evaluating IGR. A comparison with both: IGR with and without normals should have been considered.

Questions

see weakness section

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

to some extent, the limitations are discussed.

Reviewer rqMu6/10 · confidence 4/52023-07-07

Summary

The paper considers the task of surface reconstruction from point clouds without normals with INRs. They consider solving for the SDF by solving the p-Poisson equation (with manifold constraint) as p\to\infty. However as the obvious loss function form of this is difficult to optimise, they consider a variable splitting strategy. They also do this w.r.t. ensuring that their gradient solution is conservative. Finally they add a minimal surface area regularisation term.

Strengths

- Variable splitting is a nice approach to the issues with neural networks and automatic differentiation - Model is backed strongly by theoretical intuition - Auxilary variables sharing the network structure is nice - Decent results

Weaknesses

- It would be nice to have some intuition as to why you propose to use the p-Poisson equation rather than other PDEs like the Eikonal equation. At first glance it seems that the reason is because it is possible to describe it as a variational problem as shown in equation 2, however that doesn't get used by your method. Another argument you posit is that without the vanishing viscosity method a normal eikonal PDE based solution may produce a non-unique weak solution, why is it clear that your method does not produce a non-unique weak solution? Is it because of the curl-free constraint being enforced? - How important is the enforcement of the loose eikonal constraint within the curl-free constraint? It somewhat diminishes the story of trying to solve the SDF problem using a different PDE to the eikonal equation. - It is not clear why the curl-free constraint is needed. $\nabla u$ is curl-free by design, so isn't minimising $||\nabla u - G||$ in (8) sufficient? Why is a separate auxiliary variable necessary, apart from enforcing a loose Eikonal constraint by construction? Doesn't the argument about needed to set $\lambda_1$ infinitely large to enforce the constraint apply to $\lambda_3$ as well? - The qualitative diagrams (Figure 5-7) for the ablation study are great for intuition and understanding, however you should have quantitative results on what happens when you remove each of those components (especially for curl-free) I like this type of approach, willing to increase the score if clarity on the necessity of both the curl-free constraint and the eikonal construction constraint as well as quantitative ablations are given.

Questions

- What is $F$? Is it kept to the example given, $1/3x$? - I am confused why the curl-free constraint needed. By equation 10, the new curl-free constraint ensures that $G=\nabla v$ for some $v$, but why allow it to be some $v$ that is not the $u$ being outputed as the INR value, and/or instead constrain it to be similar to the current INR value $u(x)$? - I would like more intuition on the role of the curl-free constraint. Figure 5 seems to show that it forces the model to pay attention to detail more, however I don't see why this is the case theoretically. Is it because the eikonal term is loosely being enforced within the curl-free objective by construction? - The good results of IGR on Thingi10K seems unlikely given its bad performance on SRB, it is almost as good as your model. Is there a property of Thingi10K that causes this? Is there a reason your model and IGR would be so similar on Thingi10K? - The performance of your model on Daratech in SRB seems a bit confusing, in the results it does really badly on $d_C$ however it looks fairly good in Figure 6?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Some discussion given. No clear potential negative societal impact or broader societal impacts to discuss

Reviewer 2pnE2023-08-13

The rebuttal has been thorough. I recommend authors add additional explanation and try to be as clear as possible to convey the motivation/intuition of adopting p-Poisson equation in the final manuscript. Regardless, this work has enough novelty. I change my score to 7, and recommend acceptance of this paper.

Reviewer WkvV2023-08-16

Reply

After reading the other reviews and your responses, I think some of the concerns are addressed well. Here, more comprehensive numerical evaluations are provided to demonstrate the quantitative performance of the proposed methods, I suggest the author add these experiments to the revised paper, including the numerical evaluations, training/inference time, and discussion with SAP. All of the responses have addressed my major concerns, instead of the unfair submission. After that, I am positive about the submission and will change my score to accept when I ignore the unfair submission. I have raised the issues to the ac and sac, and I have no other comments if it is uncritical for the submission.

Reviewer rqMu2023-08-18

Thanks for addressing my concerns, final comments and score change

The new explanation for the theoretical motivation for p-poisson makes a lot more sense, please have that more clear in the paper, it is the most important part of your paper. It would be cool if the uniqueness was able to be shown for your method with a toy problem, e.g. for a very simple set of points in 2D you show that under different initialisations an eikonal loss guided network will converge to drastically different solutions, however with the same initialisations your loss leads to a single (or at much less varied) solution. Same thing with the reason for the curl-free constraint: please improve the explanation in the paper, and it would be great if you can show a toy example showing that often L_2's minimisation does't converge to a curl-free field in practice (while theoretical counter-examples are great, since everything is happening with neural networks which are biased to very smooth approximations due to gradient descent, it would be great to show it is a practical consideration as well). Though its not completely neccessary as your new Table A1 indicates this too. Thanks for providing Tables A1-3, they provide a lot of context about your method. I reccommend having A1 in the main paper alongside the visualisation (maybe A3 as well, and A2 can go to supplementary). As the authors have sufficiently addressed my concerns, I am increasing my score from 4 to 6. I hope they consider my comments for improving the paper (whichever are reasonable for them to do).

Reviewer ymPA2023-08-18

Thank you for the rebuttal. It addresses most of my concerns. I am going to maintain my rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC