Research map / Point clouds / Point cloud classification
Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a part-whole-hierarchy message passing network that groups parts into unsupervised super-parts, predicts latent super-part poses, and uses them to guide part pose prediction for generative 3D part assembly on PartNet.
Problem
Generative 3D part assembly requires predicting 6-DoF poses for parts without semantic labels, but prior methods ignore part-whole hierarchies that could constrain and simplify pose prediction.
Approach
Group geometrically similar parts into unsupervised super-parts, treat super-part poses as latent variables learned only from part-pose supervision, and inject the transformed point clouds and latent super-part poses into a part encoder via cross-level and within-level attention.
Key results
- Chair part accuracy 64.13% and connectivity 49.28% at Chamfer threshold 0.01 versus 62.80%/48.45% for IET, with lower SCD (0.51e-2 vs 0.54e-2).
- Mean part accuracy on chairs 76.79% vs 74.92% for IET; mean connectivity accuracy 65.32% vs 62.37%.
- Human study: ours preferred over IET 95.2%, RGL 97.6%, DGL 99.5% of comparisons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the part-whole hierarchy message passing network | Part assembly model · PartNet Chair | mPA: 74.92 -> 76.79 | Helped |
| Add the part-whole hierarchy message passing network | Part assembly model · PartNet Chair | mCA: 62.37 -> 65.32 | Helped |
| Add the part-whole hierarchy message passing network | Part assembly model · PartNet Table | PA (%): 61.67 -> 64.83 | Helped |
| Add the part-whole hierarchy message passing network | Part assembly model · PartNet Lamp | mPA: 52.80 -> 53.31 | Helped |
| Add the part-whole hierarchy message passing network | Part assembly model · PartNet Lamp | mCA: 75.94 -> 78.15 | Helped |
| Add the super-part encoder (ablation) | Part assembly model · PartNet Table | PA: 61.04 -> 64.83 | Helped |
| Add the super-part encoder (ablation) | Part assembly model · PartNet Table | CA: 55.39 -> 58.45 | Helped |
| Add the super-part encoder (ablation) | Part assembly model · PartNet Table | SCD: 0.0036 -> 0.0028 | Helped |
| Remove the part translation loss Lt (ablation) | Part assembly model · PartNet Table | assembly performance | Hurt |
| Remove the rigid rotation loss Lr (ablation) | Part assembly model · PartNet Table | part rotation accuracy | Hurt |
| Remove the shape assembly loss Ls (ablation) | Part assembly model · PartNet Table | assembly coherence | Hurt |
Surprising findings
- Simply augmenting parameters without the super-part encoder decreased performance relative to the base model in the ablation.
Limitations
- Super-parts from bounding-box grouping may lack semantic meaning.
- Lamp category shows minor gains, attributed to limited geometric variation.
- Many inaccurate assemblies are still visually valid, so metrics may not capture plausibility.
Experimental setup
Datasets: PartNet
Benchmarks: PartNet
Metrics: SCD part accuracy (PA) connectivity accuracy (CA) mPA mCA QDS WQDS MMD
Baselines: B-Global B-LSTM B-Complement DGL Score RGL IET
Builds on: IET DGL RGL PointNet Min-of-N loss
Closest papers by meaning
- Imagine: Image-Guided 3D Part Assembly with Structure Knowledge Graph (AAAI 2025 · similarity 0.91)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- PartNeXt: A Next-Generation Dataset for Fine-Grained and Hierarchical 3D Part Understanding (NeurIPS 2025 · similarity 0.88)
- AutoPartGen: Autoregressive 3D Part Generation and Discovery (NeurIPS 2025 · similarity 0.88)
- Neural Assembler: Learning to Generate Fine-Grained Robotic Assembly Instructions from Multi-View Images (AAAI 2025 · similarity 0.87)
- PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.