Research map / 3D vision / Face and shape
Category-Level Multi-Part Multi-Joint 3D Shape Assembly
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 hierarchical graph network that jointly reasons over part geometry and peg-hole joints to assemble multi-part shapes with aligned joints, plus a joint-annotated PartNet benchmark.
Problem
Multi-part shape assembly ignores contacting joints; the task requires both valid shape structure and correct peg-hole joint matching and alignment, which involves a large combinatorial matching space and error compounding.
Approach
Decouples the task into a part graph (shape structure from part geometry) and a joint graph (peg-hole matching), fused via max-pooled joint-centric part features and refined over coarse-to-fine graph iterations, with joint-aware shape and joint losses that include a matching algorithm for congruent parts.
Key results
- Ours achieves average Part Accuracy 69.2 and Shape CD 0.006 versus best baseline B-GNN 61.7 and 0.007 under original setting.
- Ours achieves average Joint Accuracy 51.4 versus 16.3 for best baseline under joint-input setting (B-Complement) and 26.9 for B-GNN in original setting.
- Ablation shows removing the first part-graph iteration drops Part Accuracy from 72.81 to 23.71 on chairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the hierarchical graph assembly method rather than the strongest listed baseline | Multi-part multi-joint shape assembly · Average | Shape Chamfer Distance: 0.024 -> 0.006 | Helped |
| Use the hierarchical graph assembly method rather than the strongest listed baseline | Multi-part multi-joint shape assembly · Average | Joint Matching Accuracy: 16.3 -> 51.4 | Helped |
| Add the first part-graph iteration (ablation) | Ours full · Chair | Part Accuracy: 23.71 -> 72.81 | Helped |
| Add joint embedding (ablation) | Ours full · Chair | Part Accuracy: 33.36 -> 72.81 | Helped |
| Add the joint-matching algorithm (ablation) | Ours full · Chair | Joint Accuracy: 37.50 -> 57.18 | Helped |
| Add Lflip (ablation) | Ours full · Chair | Joint Accuracy: 39.71 -> 57.18 | Helped |
| Add Lcoarse (ablation) | Ours full · Chair | Joint Accuracy: 44.82 -> 57.18 | Helped |
| Add Lfine (ablation) | Ours full · Chair | Joint Accuracy: 52.01 -> 57.18 | Helped |
Surprising findings
- Adapting baselines to joint inputs and losses causes them to collapse into clustered shapes, reducing their part accuracy to roughly one-third of the original setting.
Limitations
- Only handles simple peg-hole joint type
- Does not address sequential or programmatic assembly planning
- Assumes point cloud parts with joints from PartNet, with joint detection based on a Chamfer distance heuristic
Experimental setup
Datasets: PartNet
Benchmarks: joint-augmented PartNet chairs, tables, cabinets
Metrics: Part Accuracy Shape Chamfer Distance Joint Accuracy Joint Chamfer Distance
Baselines: B-Global B-LSTM B-Complement B-GNN B-NSM B-Joinable
Builds on: PointNet Hungarian matching B-Joinable (Willis et al.) PartNet
Closest papers by meaning
- Imagine: Image-Guided 3D Part Assembly with Structure Knowledge Graph (AAAI 2025 · similarity 0.90)
- Two by Two: Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation (CVPR 2025 · similarity 0.87)
- Neural Assembler: Learning to Generate Fine-Grained Robotic Assembly Instructions from Multi-View Images (AAAI 2025 · similarity 0.87)
- 3D Geometric Shape Assembly via Efficient Point Cloud Matching (ICML 2024 · similarity 0.86)
- Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing (CVPR 2024 · similarity 0.86)
- BiAssemble: Learning Collaborative Affordance for Bimanual Geometric Assembly (ICML 2025 · similarity 0.84)
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.