acceptodds
Under review as a conference paper at ICLR 2027

PatchGF: Feed-Forward Geometric Fields from Unoriented Point Clouds

11 views

Abstract

Surface reconstruction and differential-quantity estimation from unoriented point clouds are often addressed separately, although both depend on the same local geometry. We present PatchGF, a feed-forward framework that jointly predicts patch-conditioned unsigned distance fields (UDFs), unoriented normals, mean-curvature magnitudes, and Gaussian curvatures. A shared Point Transformer encodes each patch; query-to-patch cross-attention predicts unsigned distances at nearby spatial locations, while dedicated heads estimate normals and curvatures. A single model is trained exclusively on synthetic quadratic patches and applied directly to unseen point clouds without fine-tuning or per-shape network optimization. Paired queries sampled along analytic normals, with curvature-adaptive offsets, supervise near-surface distance variation. At inference, fixed anchor-centered patches provide consistent local supports, and distance-weighted fusion combines overlapping predictions into a UDF over the observed surface region. Positive-level-set extraction and gradient-based projection provide an initial reconstruction, which can be refined through local surface fitting and adaptive remeshing guided by the predicted differential quantities. Experiments on standard benchmarks and real-world scans evaluate reconstruction and differential-quantity estimation and demonstrate accurate reconstruction under noise and outlier contamination.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_g7ZGOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a feedforward model for predicting a collection of local geometrical properties (UDFs, normals, mean-curvature magnitude and Gaussian curvatures) from unoriented point clouds. Authors use a single point transformer to jointly predict those, trained on a purely synthetic dataset of quadratic patches. At inference, this approach samples a set of patches using farthest-point sampling, and uses predicted quantities at those patches to create a mesh: per-patch UDFs are fused into a single distance field, and normals and curvatures are then used for mesh refinement. Evaluation is conducted on a set of standard synthetic benchmarks and small set of real-world scans.

Critical strengths
  • Strong performance on surface reconstruction and curvature accuracy. The method is either best or runner up on most metrics, and seems to be somewhat robust to noise. This is consequential since it suggests that a single model can jointly predict several quantities necessary for high-fidelity reconstructions, while generalizing from purely synthetic training data.
Critical weaknesses
  • Idea of predicting SDF/UDF properties with a point transformer / pointnet is far from novel, such as LoSF-UDF. Apart from the baselines already mentioned in the paper, Points2Surf and PPSurf seem like established methods that also operate on (noisy) point clouds. It also looks like a lot of the baselines used are not the most recent (e.g. recent MPF, PrimitiveUDF are not compared to). Proposed approach does not advance existing architecture (takes an off-the-shelf point transformer) in a meaningful way, does not address scalability (final pipeline still requires per-shape fitting which can take >11mins). Overall, it seems like this work might be of limited interest to wider community.
  • Results on normals are not SOTA, and method seems less robust to noise than some of the baselines, which makes the claims of the paper somewhat weaker.
Questions for authors
  • Robustness to non-uniform sampling: would be great to have a discussion on whether non-uniformity of input point clouds (which is common in real-world data) significantly changes the result?
  • How much of the gain is due to predicted UDF vs post-processing? E.g. it looks like in terms of inference-only UDF prediction PatchGF performs worse than some of the baselines?
  • Could you elaborate on the key fundamental distinction with LoSF-UDF? That method also trains a UDF predictor on synthetic quadratic patches. Without a direct ablation on components it is hard to judge the difference.
Reviewer_5bHDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a feed-forward point transformer that predicts unsigned distances, normals, and curvatures at query points. These differential quantities constrain the neural field and improve surface reconstruction. Experiments on several shape datasets demonstrate good reconstruction performance.

Critical strengths
  1. The method uses a single feed-forward model based on an expressive point transformer architecture.

  2. The model is trained only on synthetic mesh patches with with curvature-aware supervision and generalizes well across shapes and datasets.

  3. The same model predicts differential geometric properties of the underlying surface, including unoriented normals and mean and Gaussian curvatures, which in turn improve UDF-based surface reconstruction.

Critical weaknesses
  1. The evaluation on SRB does not sufficiently establish performance on real-world scans. Given the intended generalization of local patches, can the trained model handle large indoor scenes, such as rooms and offices from ScanNet or Matterport3D?

  2. The method achieves good surface reconstruction accuracy, but its estimation of differential quantities, particularly normals, appears less accurate and sensitive to noise. What causes this discrepancy, and how sensitive are the results to the local neighbourhood size K?

  3. Since the method performs multi-resolution mesh reconstruction, Wireframe visualisations would help assess how mesh resolution adapts across different regions of a single mesh.

  4. Since UDFs do not enforce watertightness, how frequently do unintended holes occur in the reconstructed meshes? Please show such examples and provide discussions.

  5. Providing visualisations of the predicted normals would be helpful.

Questions for authors

How many shapes and patches are used to train the model?

Reviewer_vHzKOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of efficient surface reconstruction and local geometric property estimation from unoriented point clouds. The authors aim to establish a unified local geometric representation that jointly predicts unsigned distance fields (UDFs), normals, and curvatures, and leverages these geometric cues to improve surface reconstruction quality. The main motivation is that UDFs, normals, and curvatures, although representing different geometric properties, are all determined by the underlying local surface geometry and can therefore be jointly learned from shared local features. The main contributions include a shared patch representation, a synthetic supervision strategy based on quadratic patches, and a reconstruction pipeline guided by the predicted geometric properties. Experimental results demonstrate the effectiveness of the proposed approach.

Critical strengths

Reasonable motivation. Jointly learning related geometric properties, such as UDFs, normals, and curvatures, and leveraging their complementary information to improve surface reconstruction is a reasonable and valuable direction.

Interesting synthetic supervision strategy. The method does not rely on complete shape-level UDF training data but instead uses synthetic quadratic patches to generate local training samples. In principle, this strategy can learn generalizable local geometric priors and improve generalization to unseen shapes.

Critical weaknesses

Insufficient analysis of the benefits of joint prediction. The paper does not sufficiently demonstrate the advantages of joint training over independently training each prediction task. Furthermore, the differences between the proposed method and related joint geometric prediction approaches, such as OscuFit, should be discussed more clearly to better establish the novelty of this work.

The source of performance improvements is unclear. It remains unclear whether the improvements mainly come from the joint prediction strategy or the feature extraction capability of the Point Transformer encoder. Controlled experiments are needed to better isolate their respective contributions.

Limited validation of generalization from synthetic quadratic patches. Quadratic patches mainly capture local second-order surface geometry and may not fully represent complex geometric structures in real-world point clouds, such as high-frequency details, sharp edges, and complex local structures. More experiments are needed to demonstrate whether a model trained exclusively on synthetic quadratic patches can generalize effectively to diverse and complex real-world geometry.

Insufficient evaluation of inference efficiency. Although the paper emphasizes feed-forward inference without per-shape optimization, a comparison of inference efficiency with existing methods is missing. The authors should report inference time, GPU memory consumption, and computational costs under different input sizes.

Questions for authors

Please see weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.