acceptodds
Under review as a conference paper at ICLR 2027

Projective Normal Fields: A Convex Optimization Method for Constructing Smooth UDFs

Abstract

Constructing a smooth approximation of an unsigned distance field (UDF) from a raw point cloud is challenging because the input provides neither surface connectivity nor consistently oriented normals. Methods that directly learn a scalar UDF must also handle its non-differentiability on the zero level set and weak supervision away from the samples, which can lead to unstable optimization and spatial artifacts. We introduce Projective Normal Fields (PNFs), an orientation-free representation and convex optimization framework for estimating bidirectional normals from point positions alone. Each normal axis is encoded by a rank-one projector, which is invariant to normal reversal. We relax the non-convex set of hard projectors to its convex hull: the symmetric positive-semidefinite matrices with unit trace. Each soft tensor defines a local quadratic distance model and retains the relative weights of candidate normal axes. We estimate a coherent PNF by combining local tangent-plane fitting, soft-PCA anchoring, and overlap regularization on a fixed neighborhood graph. With positive anchoring weights, the objective is strongly convex and admits a unique global minimizer. Principal eigenvectors provide bidirectional normals, while the corresponding eigengaps provide spectral confidence indicators. We use these indicators to select and weight directional sources for heat diffusion, followed by Poisson integration to construct a regularized UDF approximation. By separating local geometry estimation from scalar-field construction, PNF avoids directly fitting the non-differentiable UDF. Experiments demonstrate reduced sensitivity to neighborhood size, competitive reconstruction under noise and outliers, and improved accuracy near non-manifold junctions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point clouds, 3D reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MA4AOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Projective Normal Fields (PNFs), an orientation-free representation for estimating bidirectional normals from raw point clouds, and a convex optimization framework for constructing smooth unsigned distance field (UDF) approximations. Each normal axis is encoded by a rank-one projector, which is invariant to normal reversal. The non-convex set of hard projectors is relaxed to its convex hull—symmetric positive-semidefinite matrices with unit trace—yielding a strongly convex objective with a unique global minimizer under positive anchoring weights. Principal eigenvectors provide bidirectional normals, and the corresponding eigengaps serve as spectral confidence indicators, which are used to weight directional sources for heat diffusion, followed by Poisson integration to construct a regularized UDF approximation. Experiments report reduced sensitivity to neighborhood size, competitive reconstruction under noise and outliers, and improved accuracy near non-manifold junctions.

Critical strengths
  1. Well-motivated problem.
  2. Technically sound convex relaxation.
  3. Elegant orientation-free design.
  4. Clear mathematical exposition. But ...
Critical weaknesses

The paper is not a representation learning contribution (primary reason for rejection).

The authors frame PNF as an “orientation-free representation” and the resulting UDF as a “compact and flexible representation.” However, the paper does not constitute a representation learning contribution in any meaningful sense. The distinction is not terminological but methodological:

  1. No learned parameters. PNF is obtained by solving a convex optimization problem independently for each input point cloud. There is no training procedure, no gradient-based parameter updates, and no shared parameters across samples.

  2. No cross-sample generalization. Each point cloud is solved in isolation. There is no train/test split, no encoder that transfers to unseen shapes, and no latent space that can be interpolated or reused.

  3. No data-driven prior. The method relies entirely on geometric priors and regularization terms (local tangent-plane fitting, soft-PCA anchoring, overlap regularization). No prior is learned from a data distribution.

  4. The UDF is constructed, not learned. Heat diffusion followed by Poisson integration is a deterministic post-processing pipeline. It is not a learned implicit field.

In short, PNF is a per-instance, optimization-based representation construction, not representation learning. ICLR’s scope is centered on learning-based methods and representation learning, broadly construed. A purely optimization-driven geometry processing pipeline, however elegant, does not fit this scope unless it is explicitly framed as a learning contribution—for example, by learning the PNF representation with a network, or embedding it as a differentiable module inside a learning system. The current submission does none of these.

Questions for authors

it is a good paper, and submit it to CVPR or some other place.

Reviewer_iomBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper reconstructs surfaces, including open, thin, and non-manifold ones, from raw point clouds without normals. Following its predecessor VAD, it first estimates a normal at every point and then turns these normals into an unsigned distance field (UDF) and a mesh. Its contribution is the normal-estimation step. Because a normal's sign is unknown, each normal n is stored as the sign-free matrix nnᵀ. Relaxing these matrices to their convex hull (positive-semidefinite, unit trace) lets all normals be estimated jointly by one convex optimization, which fits local tangent planes, stays close to a soft PCA estimate, and makes neighboring local surface models agree. Because the convex problem has a unique optimum, the result does not depend on initialization. Each normal is the top eigenvector of its matrix, and the gap between the top two eigenvalues serves as a confidence. Normals with confidence below 0.1 are discarded before VAD-style heat diffusion, integration, and meshing. In experiments, the normals are more accurate and less sensitive to neighborhood size than PCA, and reconstruction error is lower than VAD.

Critical strengths
  • Storing each normal as the matrix Mᵢ = nᵢnᵢᵀ, relaxed to any positive-semidefinite matrix with unit trace, makes the squared distance to a tangent plane a linear function of the unknown Mᵢ. Every term of the objective is therefore linear or the square of a linear expression. The relaxed problem is thus convex with a unique solution that does not depend on initialization, unlike VAD’s non-convex optimization.
  • PNF is more accurate than PCA in almost every setting and varies much less as the neighborhood changes.
Critical weaknesses
  • Without the overlap terms, the method reduces exactly to PCA. Yet the overlap weights are never reported, and the terms are never ablated.
  • The reconstruction metrics are one-directional (reference to reconstruction only), so fragments and double layers are not counted. Also, only means are reported, and the margins over VAD on clean and noisy data are small.
Questions for authors
  • Without the overlap terms, the method reduces to PCA. I expect an ablation study to ablate the overlap terms (all removed, and each removed in turn) to show how much each contributes.
  • I think the same 0.1 cut-off should be applied to PCA and VAD normals in Table 4, and show how sensitive the results are to the threshold.
  • Could you report symmetric (two-sided) Chamfer and Hausdorff distances, so that spurious fragments and double layers are also counted?
  • What hyperparameter values (the objective weights, neighborhood size, and input normalization) were used in each experiment?
  • How are the PSD and unit-trace constraints enforced during gradient descent, what step size is used, and do the 500 iterations converge to the unique minimizer?
  • Some closely related work should be mentioned and discussed:
    • Guy, G., & Medioni, G. (1997). Inference of surfaces, 3D curves, and junctions from sparse, noisy, 3D data. IEEE TPAMI, 19(11), 1265-1277.
    • Alliez, P., Cohen-Steiner, D., Tong, Y., & Desbrun, M. (2007). Voronoi-based variational reconstruction of unoriented point sets. Proc. Eurographics Symposium on Geometry Processing (SGP), 39-48.
Reviewer_Zu6nOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The authors claim that current state-of-the-art methods focused on directly predicting a scalar UDF suffer from the commonly known zero level set and weak supervision. This leads to a worse optimization resulting in significant artifacts in the reconstruction phase form point cloud to mesh. The authors claim that this can be mitigated through an orientation-free representation and convex optimization of normals from points alone. From my understanding the idea behind this work is to reduce the impact of the zero-level set by introducing an invariant normal optimization method for each point in 3D. The normal is invariant to its reversal thus suffering less from this during the UDF estimation process.

Critical strengths

I think the mathematical foundation is strong and provides solid grounding of the main claim the authors make. The results, while sparse, seem to show improvements over PCA across all distance metrics.

Critical weaknesses

Major:

  • While PNF outperforms PCA across all given samples the number of samples is incredible sparse (Table 1) . Additionally, in Table 4. we see additional evaluation, but considering the VAD and PNF share a similar structure for a uniform confidence the performance is nearly the same
  • The confidence value choice seems a little magical to me could you please provide additional insights on the confidence value and how much search it requires to obtain the correct threshold, how well it translates to other shapes etc.?
  • Significant similarity to VAD while not having significant difference in most scenarios (expect outlier cases which seem to be a significant positive of the paper but have not been evaluated enough to be relevant)
  • Majorly, we are lacking enough analysis on the impact of the second-order in the convex optimization. Currently, the work is lacking a clean insight/observation that might be relevant for the larger community. For me the Hessian most likely impacts the performance positively, as such I would like to have a comparison with Neural Singular Hessian [1] (even though its not a perfect method comparison), as it might be that both works contain the same insight but different approach in solving the problem (thus requiring additional comparisons=
  • What is the computational complexity and latency etc.? This has not been discussed is usually a main limitations in methods that aim to solve the zero-level set problem of UDFs
  • More evaluation on more data, there is enough 3d datasets by now (like ShapeNet, ModelNet etc.) with >100 shapes this is required to have a statistically relevant analysis

Comment:

  • I think its a nice contribution but may lack enough evaluation and additional insights for the conference level

[1] Wang, Z., Zhang, Y., Xu, R., Zhang, F., Wang, P., Chen, S., Xin, S., Wang, W., & Tu, C. (2023). Neural-Singular-Hessian: Implicit Neural Representation of Unoriented Point Clouds by Enforcing Singular Hessian. ACM Transactions on Graphics (TOG), 42, 1 - 14.

Questions for authors

Have been asked under weakness section. Please address all those points I am always open to changing my opnion!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.