Learning to Estimate Surface Area from Voxelised 3D Objects: A Comparative Study of Classical Geometry, 3D Convolutional Networks and PointNet

Abstract Surface area is a fundamental descriptor in three-dimensional morphology, with relevance to porous materials, biomedical imaging, particulate systems and computational geometry. In voxelised data, however, surface area is not directly represented: it must be inferred from a discrete occupancy field, and conventional estimators exhibit resolution- and noise-dependent bias. This work investigates whether learned regressors can estimate the surface area of a 3D object more accurately than direct geometric estimators when only its binary voxel representation is available. A synthetic dataset of 15,000 geometric instances was generated across five shape families—spheres, ellipsoids, cylinders, superellipsoids and Gaussian-random-field (GRF) blobs—with three voxel resolutions (32 3 , 64 3 , 128 3 ) and three levels of synthetic boundary perturbation (0, 0.02 and 0.05), giving 135,000 volume variants. Ground-truth areas were obtained analytically for spheres and cylinders and from a higher-resolution Marching Cubes reference mesh for the remaining families. A 3D convolutional neural network (CNN) was trained independently at each resolution, while PointNet was trained on 2,048 surface points sampled from Marching Cubes meshes at 32 3 . On the held-out test set, PointNet achieved an MAEof0.1953andMAPEof2.73%, outperforming the matched CNN (MAE 0.2813, MAPE 5.25%), Marching Cubes (MAE 0.7889, MAPE 13.95%) and voxel-face counting (MAE 3.8983, MAPE 58.07%). The CNN remained substantially more accurate than the classical baselines at all three resolutions, although its aggregate error was not monotonic with resolution. The results demonstrate the promise of learned geometric priors for surface-area estimation while exposing important limitations arising from shape imbalance, reference-mesh uncertainty, architecture scaling and the dependence of PointNet preprocessing on Marching Cubes.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11715919/v1
Primary Topic
3D Shape Modeling and Analysis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Learning to Estimate Surface Area from Voxelised 3D Objects: A Comparative Study of Classical Geometry, 3D Convolutional Networks and PointNet

Mayank Goswami, Tanmay Kumar
1 links
Research Square
3D Shape Modeling and Analysis
preprint

Learning to Estimate Surface Area from Voxelised 3D Objects: A Comparative Study of Classical Geometry, 3D Convolutional Networks and PointNet

Mayank Goswami, Tanmay Kumar
preprint en

Abstract

Abstract Surface area is a fundamental descriptor in three-dimensional morphology, with relevance to porous materials, biomedical imaging, particulate systems and computational geometry. In voxelised data, however, surface area is not directly represented: it must be inferred from a discrete occupancy field, and conventional estimators exhibit resolution- and noise-dependent bias. This work investigates whether learned regressors can estimate the surface area of a 3D object more accurately than direct geometric estimators when only its binary voxel representation is available. A synthetic dataset of 15,000 geometric instances was generated across five shape families—spheres, ellipsoids, cylinders, superellipsoids and Gaussian-random-field (GRF) blobs—with three voxel resolutions (32 3 , 64 3 , 128 3 ) and three levels of synthetic boundary perturbation (0, 0.02 and 0.05), giving 135,000 volume variants. Ground-truth areas were obtained analytically for spheres and cylinders and from a higher-resolution Marching Cubes reference mesh for the remaining families. A 3D convolutional neural network (CNN) was trained independently at each resolution, while PointNet was trained on 2,048 surface points sampled from Marching Cubes meshes at 32 3 . On the held-out test set, PointNet achieved an MAEof0.1953andMAPEof2.73%, outperforming the matched CNN (MAE 0.2813, MAPE 5.25%), Marching Cubes (MAE 0.7889, MAPE 13.95%) and voxel-face counting (MAE 3.8983, MAPE 58.07%). The CNN remained substantially more accurate than the classical baselines at all three resolutions, although its aggregate error was not monotonic with resolution. The results demonstrate the promise of learned geometric priors for surface-area estimation while exposing important limitations arising from shape imbalance, reference-mesh uncertainty, architecture scaling and the dependence of PointNet preprocessing on Marching Cubes.

Research Square
Indian Institute of Technology Roorkee (IN)
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