Neural Toroidal Patch: On Smooth Meshes, the Presence of a Local Analytic Base Matters More Than Its Form or the Network Architecture

Abstract In neural surface representation, how much does the presence of a geometric prior matter relative to its specific form or the network architecture? We study this question using the Neural Toroidal Patch (NTP), a representation built on the osculating torus, a local approximant that makes second‐order contact with the target surface by matching position, normal, and both principal curvatures at each seed vertex. NTP decomposes a triangle mesh into toroidal patches by curvature‐based fitting and iterative refinement; an MLP is then trained to predict only a small scalar residual along the torus normal, with the torus itself accounting for most of the local geometry. Holding the patch decomposition fixed and the training protocol uniform, no tested architectural augmentation improves over a flat three‐layer MLP with 41K parameters, and a hierarchical design with 3.7× more parameters performs no better, scoring slightly but consistently lower across all five random seeds. An osculating torus, an osculating paraboloid, and a tangent plane reach comparable accuracy, within about 1.2 × at supervised vertices and roughly 2× when the surface is queried continuously between vertices, so the specific form of the analytic base contributes less to final accuracy than its presence. In contrast, removing the base entirely, using a capacity‐matched Direct MLP, is an order of magnitude worse; a control that keeps the analytic per‐patch pose but removes the curved base surface remains substantially worse. The base also behaves as a data‐efficient regularizer, its advantage over a base‐free network widening as supervision thins and persisting when the base parameters are fit on the training vertices alone. The learned residual also transfers across shapes. A shared MLP trained on eight meshes adapts to four held‐out meshes by tuning only per‐patch latent codes, roughly halving their error, whereas a randomly initialized frozen baseline barely improves. For the smooth meshes studied here, the presence of an analytic base matters more than its specific form or the network architecture.

Authors

Institutions

Publication Details

Journal
Computer Graphics Forum
Published
2026-11-13
DOI
https://doi.org/10.1111/cgf.70642
Primary Topic
3D Shape Modeling and Analysis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Neural Toroidal Patch: On Smooth Meshes, the Presence of a Local Analytic Base Matters More Than Its Form or the Network Architecture

Youngjin Park, Jinyoung Choi
4 links
Computer Graphics Forum
3D Shape Modeling and Analysis
article

Neural Toroidal Patch: On Smooth Meshes, the Presence of a Local Analytic Base Matters More Than Its Form or the Network Architecture

Youngjin Park, Jinyoung Choi
article en

Abstract

Abstract In neural surface representation, how much does the presence of a geometric prior matter relative to its specific form or the network architecture? We study this question using the Neural Toroidal Patch (NTP), a representation built on the osculating torus, a local approximant that makes second‐order contact with the target surface by matching position, normal, and both principal curvatures at each seed vertex. NTP decomposes a triangle mesh into toroidal patches by curvature‐based fitting and iterative refinement; an MLP is then trained to predict only a small scalar residual along the torus normal, with the torus itself accounting for most of the local geometry. Holding the patch decomposition fixed and the training protocol uniform, no tested architectural augmentation improves over a flat three‐layer MLP with 41K parameters, and a hierarchical design with 3.7× more parameters performs no better, scoring slightly but consistently lower across all five random seeds. An osculating torus, an osculating paraboloid, and a tangent plane reach comparable accuracy, within about 1.2 × at supervised vertices and roughly 2× when the surface is queried continuously between vertices, so the specific form of the analytic base contributes less to final accuracy than its presence. In contrast, removing the base entirely, using a capacity‐matched Direct MLP, is an order of magnitude worse; a control that keeps the analytic per‐patch pose but removes the curved base surface remains substantially worse. The base also behaves as a data‐efficient regularizer, its advantage over a base‐free network widening as supervision thins and persisting when the base parameters are fit on the training vertices alone. The learned residual also transfers across shapes. A shared MLP trained on eight meshes adapts to four held‐out meshes by tuning only per‐patch latent codes, roughly halving their error, whereas a randomly initialized frozen baseline barely improves. For the smooth meshes studied here, the presence of an analytic base matters more than its specific form or the network architecture.

Computer Graphics Forum
Pusan National University (KR)
Ministry of Trade, Industry and Energy, Korea Institute for Advancement of Technology, National Research Foundation of Korea, Busan Metropolitan City, Ministry of Science and ICT, South Korea, Institute for Information and Communications Technology Promotion
3D Shape Modeling and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon