Learned Radius Estimation for UDF-Based Point Cloud Reconstruction

Eito Ogawa; Hiroshi Watanabe

arxiv: 2606.18787 · v1 · pith:67T624YMnew · submitted 2026-06-17 · 💻 cs.CV

Learned Radius Estimation for UDF-Based Point Cloud Reconstruction

Eito Ogawa , Hiroshi Watanabe This is my paper

Pith reviewed 2026-06-26 21:57 UTC · model grok-4.3

classification 💻 cs.CV

keywords point cloud reconstructionunsigned distance fieldradius estimationsurface reconstructionlocal patch methodsneural network selectorfine-scale accuracy

0 comments

The pith

A neural selector learns a continuous per-query support radius to improve unsigned distance field surface reconstruction from point clouds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Local UDF methods reconstruct surfaces from point clouds but depend on a support radius that is usually fixed or set by a simple curvature rule, which struggles with varying local geometry. The paper trains a small network to output a continuous radius for each query point, using target values created by parabolic interpolation on cached error curves from a frozen backbone model. This selector is inserted into the existing pipeline without retraining the main UDF network. Experiments indicate higher accuracy on fine details compared with fixed or heuristic radii. The approach keeps the core reconstruction lightweight while adapting the radius to local conditions.

Core claim

The paper shows that a learned per-query radius selector, trained on off-grid target radii from parabolic interpolation of cached UDF error curves, can be plugged into a frozen LoSF-UDF backbone to predict continuous support radii and thereby raise fine-scale reconstruction accuracy on point cloud data.

What carries the argument

The learned per-query radius selector: a neural network that takes a query point and local patch and outputs a single continuous radius value used as the support radius for the UDF evaluation.

If this is right

Reconstruction accuracy improves on scenes with mixed curvature without changing the backbone model size.
The method supports continuous rather than discrete radius choices at query time.
Freezing the UDF backbone keeps training cost low while still adapting to heterogeneous geometry.
The selector can be swapped into other local-patch UDF pipelines that use a support radius.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same interpolation-based supervision could be tested on signed distance fields or other implicit representations.
Integrating the selector with online adaptation might allow radius adjustment during live scanning sessions.
Consumer devices that capture point clouds could ship the selector as a lightweight add-on to reduce manual parameter tuning.

Load-bearing premise

That the target radii obtained by parabolic interpolation of UDF error curves supply reliable and generalizable supervision for training the selector.

What would settle it

Measure reconstruction error on a held-out set of point clouds whose true per-point optimal radii are known from dense ground-truth surfaces; if the selector's predictions do not reduce error relative to the best fixed radius, the claim fails.

Figures

Figures reproduced from arXiv: 2606.18787 by Eito Ogawa, Hiroshi Watanabe.

**Figure 2.** Figure 2: Reconstruction images on ScanNet. From left to right and top to [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

read the original abstract

Surface reconstruction from point clouds is important for consumer-grade 3D capture, including AR/VR and indoor scanning. Local-patch Unsigned Distance Field (UDF) methods are lightweight and generalizable, but their accuracy depends on the support radius, traditionally fixed or selected by a one-dimensional curvature heuristic that cannot capture heterogeneous local geometry. We propose a learned per-query radius selector that predicts a continuous support radius and plugs into a frozen LoSF-UDF backbone. The selector is trained using off-grid target radii obtained by parabolic interpolation of cached UDF error curves. Experiments show improved fine-scale reconstruction accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Learned continuous radius selector trained on parabolic UDF error targets is the core idea, but the abstract supplies no numbers or method details so the gains and target reliability stay unverified.

read the letter

The paper's main move is replacing fixed or curvature-heuristic radii in local-patch UDF reconstruction with a learned per-query predictor that outputs a continuous support radius. The selector trains on off-grid targets taken from parabolic interpolation of cached UDF error curves and then slots into a frozen LoSF-UDF backbone.

This is new relative to the usual one-dimensional heuristics. It directly targets the problem that local geometry varies within a single scan, which fixed radii or simple curvature rules cannot handle well. Keeping the backbone frozen is a practical choice that lowers the barrier to trying the idea.

The soft spot is the complete absence of numbers. The abstract claims better fine-scale accuracy but gives no baselines, datasets, ablation results, or even the size of the reported gain. Without those, it is impossible to judge whether the improvement is real or worth the added complexity. The stress-test concern also lands: parabolic interpolation of cached error curves can only supply good targets if the curves are densely sampled and the fit actually recovers the true minimum rather than an artifact. The abstract says nothing about sampling density or caching procedure, so that assumption remains unchecked.

The work is aimed at practitioners doing consumer-grade 3D capture with UDF methods. A referee could usefully examine the experimental setup and test whether the interpolated targets generalize across different local geometries. I would send it for peer review to get those details evaluated rather than desk-reject on the abstract alone.

Referee Report

2 major / 1 minor

Summary. The paper claims that a learned per-query radius selector, trained on off-grid target radii obtained by parabolic interpolation of cached UDF error curves, can be plugged into a frozen LoSF-UDF backbone to improve fine-scale surface reconstruction accuracy from point clouds compared to fixed or heuristic radius selection.

Significance. If the interpolated targets prove reliable and the selector generalizes, the approach could address a key limitation of local UDF methods by adapting support radii to heterogeneous geometry, with potential benefits for consumer 3D applications. The frozen-backbone design is a practical strength that avoids retraining the full model.

major comments (2)

[Abstract] Abstract: the claim that 'experiments show improved fine-scale reconstruction accuracy' supplies no quantitative results, baselines, dataset details, or ablation studies, preventing verification of the central experimental claim.
[Abstract] Training procedure (described in abstract): the off-grid targets are generated by parabolic interpolation of cached UDF error curves; the manuscript must show that these targets recover the true error-minimizing radius (e.g., via denser sampling validation or error-curve analysis) rather than curve-shape artifacts, as this supervision signal is load-bearing for the learned selector.

minor comments (1)

[Abstract] LoSF-UDF acronym is used without expansion.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and will incorporate revisions to strengthen the presentation of results and validation of the supervision signal.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that 'experiments show improved fine-scale reconstruction accuracy' supplies no quantitative results, baselines, dataset details, or ablation studies, preventing verification of the central experimental claim.

Authors: We agree that the abstract lacks the necessary quantitative details to support the central claim. In the revised manuscript we will expand the abstract to report specific metrics (e.g., Chamfer distance and normal consistency improvements), name the datasets and baselines, and briefly reference the ablation studies that demonstrate the benefit of the learned selector. revision: yes
Referee: [Abstract] Training procedure (described in abstract): the off-grid targets are generated by parabolic interpolation of cached UDF error curves; the manuscript must show that these targets recover the true error-minimizing radius (e.g., via denser sampling validation or error-curve analysis) rather than curve-shape artifacts, as this supervision signal is load-bearing for the learned selector.

Authors: We acknowledge that explicit validation of the parabolic targets is required. The current manuscript does not contain a dedicated denser-sampling comparison or error-curve analysis; we will add such validation (including quantitative deviation statistics and example curves) in a new subsection of the experiments to confirm that the interpolated radii reliably approximate the true minima. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external supervision targets

full rationale

The paper trains a learned radius selector on target radii obtained via parabolic interpolation of cached UDF error curves, which are computed independently of the selector itself. These targets serve as supervision for training, after which the selector is plugged into a frozen LoSF-UDF backbone. No equation or step reduces the output to a fitted quantity defined by the model, nor does any load-bearing claim rest on self-citation chains or imported uniqueness theorems. The approach is self-contained with externally derived signals, consistent with the reader's assessment of no circular reasoning.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no explicit free parameters, axioms, or invented entities described.

pith-pipeline@v0.9.1-grok · 5615 in / 933 out tokens · 18604 ms · 2026-06-26T21:57:26.263372+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references · 1 canonical work pages · 1 internal anchor

[1]

Neural Unsigned Distance Fields for Implicit Function Learning,

J. Chibaneet al., “Neural Unsigned Distance Fields for Implicit Function Learning,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2020

2020
[2]

A Lightweight UDF Learning Framework for 3D Recon- struction Based on Local Shape Functions,

J. Huet al., “A Lightweight UDF Learning Framework for 3D Recon- struction Based on Local Shape Functions,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 1297–1307

2025
[3]

Accurate and Efficient Surface Reconstruction from Point Clouds via Geometry-Aware Local Adaptation,

E. Ogawaet al., “Accurate and Efficient Surface Reconstruction from Point Clouds via Geometry-Aware Local Adaptation,” inProc. IIEEJ Int. Conf. Image Electronics and Visual Computing (IEVC), 2026

2026
[4]

Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering,

F. Houet al., “Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering,”ACM Trans. Graph., vol. 42, no. 6, Art. 245, 2023

2023
[5]

ShapeNet: An Information-Rich 3D Model Repository

A. X. Changet al., “ShapeNet: An Information-Rich 3D Model Repos- itory,”arXiv:1512.03012, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015
[6]

Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single Images,

H. Zhuet al., “Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single Images,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2020

2020
[7]

ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,

A. Daiet al., “ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 2432–2443

2017

[1] [1]

Neural Unsigned Distance Fields for Implicit Function Learning,

J. Chibaneet al., “Neural Unsigned Distance Fields for Implicit Function Learning,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2020

2020

[2] [2]

A Lightweight UDF Learning Framework for 3D Recon- struction Based on Local Shape Functions,

J. Huet al., “A Lightweight UDF Learning Framework for 3D Recon- struction Based on Local Shape Functions,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 1297–1307

2025

[3] [3]

Accurate and Efficient Surface Reconstruction from Point Clouds via Geometry-Aware Local Adaptation,

E. Ogawaet al., “Accurate and Efficient Surface Reconstruction from Point Clouds via Geometry-Aware Local Adaptation,” inProc. IIEEJ Int. Conf. Image Electronics and Visual Computing (IEVC), 2026

2026

[4] [4]

Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering,

F. Houet al., “Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering,”ACM Trans. Graph., vol. 42, no. 6, Art. 245, 2023

2023

[5] [5]

ShapeNet: An Information-Rich 3D Model Repository

A. X. Changet al., “ShapeNet: An Information-Rich 3D Model Repos- itory,”arXiv:1512.03012, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015

[6] [6]

Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single Images,

H. Zhuet al., “Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single Images,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2020

2020

[7] [7]

ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,

A. Daiet al., “ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 2432–2443

2017