Research map / 3D vision / 3D generation
Scalable Geometric Fracture Assembly via Co-creation Space among Assemblers
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 Co-creation Space where assemblers compete to write into a shared bottleneck workspace, plus a geometric collision loss, for scalable semantic-free geometric fracture assembly on PartNet and Breaking Bad.
Problem
Geometric fracture assembly without semantic information struggles to scale to many fragments; transformer pairwise attention (IET) scales quadratically and assemblers get confused and fall into local optima with similar fragments.
Approach
A shared, capacity-limited co-creation workspace updated by top-k attention-based write competition among assemblers (linear in number of assemblers, replacing pairwise attention), combined with a geometric collision loss that pushes apart fragments occupying the same location.
Key results
- On PartNet Chair, the method reaches SCD 7.00e-3, PA 53.59, and CA 38.97, versus IET's 9.40, 37.50, and 24.47.
- On Breaking Bad Artifact, SCD improves from 16.46e-3 (IET) to 15.49e-3 and PA from 17.64 to 19.78.
- With coarse-to-fine, Ours+CTF reaches PA 55.92, SCD 6.20e-3, CA 42.69, outperforming IET+CTF (41.71, 7.80, 31.13).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use co-creation workspace instead of no workspace (ablation) | Ours · PartNet Chair | PA: 45.89 -> 53.59 | Helped |
| use co-creation workspace instead of no workspace (ablation) | Ours · PartNet Chair | SCD (×10^-3): 08.10 -> 07.00 | Helped |
| use co-creation workspace instead of no workspace (ablation) | Ours · PartNet Chair | CA: 31.73 -> 38.97 | Helped |
| add collision loss (ablation) | Ours with k=10 · PartNet Chair | PA: 52.53 -> 53.59 | Helped |
| add collision loss (ablation) | Ours with k=10 · PartNet Chair | SCD (×10^-3): 07.10 -> 07.00 | Helped |
| add collision loss (ablation) | Ours with k=10 · PartNet Chair | CA: 38.40 -> 38.97 | Helped |
| set workspace capacity to k=10 rather than k=1, k=5, k=15, or k=20 (ablation) | Ours without collision loss · PartNet Chair | SCD (×10^-3): 07.60, 07.30, 07.32, 07.30 -> 07.10 | Helped |
| add coarse-to-fine processing (ablation) | Ours · PartNet Chair | PA: 53.59 -> 55.92 | Helped |
| add coarse-to-fine processing (ablation) | Ours · PartNet Chair | SCD (×10^-3): 07.00 -> 06.20 | Helped |
| add coarse-to-fine processing (ablation) | Ours · PartNet Chair | CA: 38.97 -> 42.69 | Helped |
| use Ours rather than IET | geometric fracture assembly · PartNet Chair | PA: 37.50 -> 53.59 | Helped |
| use Ours rather than IET | geometric fracture assembly · Breaking Bad Everyday | RMSE(T) (×10^-2): 15.07 -> 15.10 | Hurt |
Surprising findings
- Removing the workspace entirely (w/o any workspace) drops PA from 53.59 to 45.89 only when collision loss is also absent in comparison; the paper reports the workspace gives the largest gain, though the table mixes conditions.
Limitations
- Routing module learning is not studied and is deferred to future work.
- Collision loss weight and C require grid search; ablations are reported only on the Artifact category.
- Hardware and compute budget are not stated.
Experimental setup
Datasets: PartNet Breaking Bad
Benchmarks: PartNet Chair Breaking Bad Everyday Breaking Bad Artifact
Metrics: SCD PA CA RMSE(R) RMSE(T)
Baselines: LSTM Global CM DGL RGL IET
Builds on: IET Global Workspace Theory Min-of-N loss PointNet Transformer
Closest papers by meaning
- PuzzleFusion++: Auto-agglomerative 3D Fracture Assembly by Denoise and Verify (ICLR 2025 · similarity 0.87)
- PHFormer: Multi-Fragment Assembly Using Proxy-Level Hybrid Transformer (AAAI 2024 · similarity 0.85)
- Re-assembling the past: The RePAIR dataset and benchmark for real world 2D and 3D puzzle solving (NeurIPS 2024 · similarity 0.84)
- 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.84)
- 3D Geometric Shape Assembly via Efficient Point Cloud Matching (ICML 2024 · similarity 0.82)
- Imagine: Image-Guided 3D Part Assembly with Structure Knowledge Graph (AAAI 2025 · similarity 0.82)
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.