Geometry-only 3D part decomposition

Hi-TOPS: Hierarchical Topology-aware Scoring Prior for 3D Part Decomposition

Ruoyu Wu^1,* Zhenhong Sun^2,* Xiaoming Gong³ Yuxin Xian⁴ Zhi Wang³ Yawen Chen¹ Huadong Mo^1,† Daoyi Dong⁵

¹ University of New South Wales ² Australian National University ³ Nanjing University ⁴ Southwestern University of Finance and Economics ⁵ University of Technology Sydney

^* Equal contribution ^† Corresponding author

Code arXiv · Soon Demo · Soon BibTeX

A composite gallery of meshes decomposed into colorful structural parts by Hi-TOPS. — Robust, geometry-driven part decomposition across diverse shapes and structural scales.

Hi-TOPS turns intrinsic mesh geometry into a hierarchical Flow-Freeze prior, guiding superquadric fitting toward coherent volumetric parts while preserving articulation seams and thin structures.

Overview

From mesh geometry to structural parts

Why Hi-TOPS

Abstract

Accurate 3D part decomposition requires separating shapes into structurally meaningful components with precise boundaries while preserving articulation seams and thin attachments. Existing approaches often suffer from a structural-scale mismatch: geometric evidence for separation is most reliable at the meso scale, yet many pipelines operate either too globally to respect joints or too locally to remain robust to noise. We propose Hi-TOPS, a Hierarchical Topology-aware Scoring Prior that aggregates complementary intrinsic cues into a multi-resolution Flow-Freeze field. Flow regions provide expandable support for primitive coverage, while Freeze regions restrict growth near articulations and thin structures. A TSDF-guided body-surface superquadric fitter then captures dominant cores and residual surface structures, followed by SQ-to-mesh assignment for connected, boundary-aligned parts. Across diverse benchmarks, Hi-TOPS delivers stable, editable decompositions without semantic supervision or 2D foundation priors.

Method

A hierarchical structural prior for decomposition

Three stages bridge stable meso-scale evidence and surface-aligned mesh parts.

Meso-scale voxelization

Curvature, normal variation, and point density reveal complementary structural evidence at several resolutions.

Hierarchical Flow-Freeze prior

Rule-based fusion preserves stable body support while refining boundary-sensitive regions at finer scales.

Body-surface SQ fitting

TSDF guidance expands superquadrics through Flow regions and stops them near Freeze boundaries before surface assignment.

Multi-scale voxelization at resolutions 16, 32 and 64 with curvature, point density and normal cues. — **Multi-scale intrinsic cues.** Coarse grids capture stable volumetric support; fine grids recover narrow joints and boundary details.

The complete Hi-TOPS pipeline from meso-scale voxelization through Flow-Freeze fusion and superquadric fitting to the segmented mesh. — **Hi-TOPS pipeline.** The hierarchical prior guides initialization, TSDF-guided inflation, and SQ-to-mesh assignment.

Results

Consistent across datasets, precise at articulations

Hi-TOPS remains competitive with learned systems while requiring no training data.

55.87PartNet mIoUBest among compared methods

51.63PartObjaverse-Tiny mIoUGeometry-only inference

47.22HY3D-Bench mIoUStrong generalization

3.5Cross-dataset std.Most consistent reported result

Qualitative comparison between ground truth, primitive methods, learning-based methods and Hi-TOPS. — **Qualitative comparison.** Hi-TOPS preserves coherent structures and articulation-aligned boundaries across diverse meshes.

Main comparison

Segmentation mIoU

Reported on three public part-annotated benchmarks. Higher is better.

Method	Training	PartObj.	PartNet	HY3D-B.
EMS	No	14.24	11.21	16.96
MPS	No	23.72	23.89	25.25
PrimAny	Yes	26.54	22.81	26.40
S²AM3D	Yes	30.64	29.77	22.20
SAMPart3D	Yes	46.74	29.48	43.90
PartSAM	Yes	49.15	30.16	55.13
PartField	Yes	66.77	45.57	44.56
PointSAM	Yes	43.27	27.66	32.94
Hi-TOPS	No	51.63	55.87	47.22

Structural fidelity

RI, VoI, and segmentation covering

Hi-TOPS delivers the strongest PartNet VoI and SC while maintaining stable performance across datasets.

Method	PartObjaverse-Tiny			PartNet
Method	RI ↑	VoI ↓	SC ↑	RI ↑	VoI ↓	SC ↑
MPS	0.719	2.370	0.295	0.734	2.216	0.338
EMS	0.373	1.770	0.283	0.372	1.897	0.277
PrimAny	0.744	2.453	0.292	0.746	2.561	0.273
SAMPart3D	0.818	1.300	0.541	0.574	1.949	0.333
PartSAM	0.772	1.563	0.481	0.791	1.814	0.427
PartField	0.815	1.088	0.548	0.837	1.625	0.493
Hi-TOPS	0.778	1.237	0.544	0.796	1.444	0.532

Values are transcribed from the current local manuscript. The public page can be updated when the final paper version is released.

Reference

Citation

This provisional BibTeX is generated from the current manuscript and can be updated with the arXiv identifier later.

@article{wu2026hitops,
  title   = {Hi-TOPS: Hierarchical Topology-aware Scoring Prior for 3D Part Decomposition},
  author  = {Wu, Ruoyu and Sun, Zhenhong and Gong, Xiaoming and Xian, Yuxin and Wang, Zhi and Chen, Yawen and Mo, Huadong and Dong, Daoyi},
  journal = {arXiv preprint},
  year    = {2026}
}