Research map / 3D vision / 3D generation
Re-assembling the past: The RePAIR dataset and benchmark for real world 2D and 3D puzzle solving
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.
Introduces RePAIR, a real-world 2D and 3D puzzle-solving dataset of 1070 scanned fresco fragments from Pompeii with archaeologist-made ground-truth reassemblies, and benchmarks baselines that perform poorly.
Problem
Existing puzzle-solving and fracture reassembly benchmarks are mostly synthetic or small-scale and do not capture realistic erosion, irregular fractures, and missing pieces of real archaeological artifacts.
Approach
A real-world dataset of eroded, irregularly fractured Pompeii fresco fragments digitized in 3D and rendered to 2D, with ground-truth 6DoF poses manually reassembled by archaeologists, plus standardized geometric and neighbor-consistency metrics for benchmarking.
Key results
- In 2D, Derech et al. achieves Qpos 0.037 and F1 0.471, while the proposed Genetic Optimization and Greedy Geometric Matching baselines reach F1 0.394 and 0.351.
- In 3D on RePAIR, Global achieves the best Qpos of 0.127 and F1 0.683, with no single method dominating across metrics.
- Models trained on Breaking Bad and tested on RePAIR perform much worse (e.g., best Qpos 0.066) than RePAIR-trained models tested on Breaking Bad (best Qpos 0.296).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Genetic Optimization instead of Derech et al. | RePAIR 2D puzzle solving · RePAIR 2D | Qpos: 0.037 -> 0.047 | Helped |
| Use Genetic Optimization instead of Derech et al. | RePAIR 2D puzzle solving · RePAIR 2D | F1: 0.471 -> 0.394 | Hurt |
| Use Greedy Geom Match instead of Derech et al. | RePAIR 2D puzzle solving · RePAIR 2D | RMSE (R°): 80.964 -> 76.987 | Helped |
| Use Greedy Geom Match instead of Derech et al. | RePAIR 2D puzzle solving · RePAIR 2D | Qpos: 0.037 -> 0.023 | Hurt |
| Use DiffAssemble instead of Global | RePAIR 3D puzzle solving · RePAIR 3D | Qpos: 0.127 -> 0.099 | Hurt |
| Use DiffAssemble instead of Global | RePAIR 3D puzzle solving · RePAIR 3D | F1: 0.683 -> 0.599 | Hurt |
| Train on RePAIR rather than Breaking Bad | DGL · Breaking Bad test subset | Qpos: 0.063 -> 0.296 | Helped |
| Train on Breaking Bad rather than RePAIR | DGL · RePAIR test set | Qpos: 0.296 -> 0.063 | Hurt |
Surprising findings
- Models trained on RePAIR generalize better to Breaking Bad than the reverse, suggesting RePAIR's complexity is not captured by synthetic datasets.
- Performance is surprisingly similar between 2D and 3D settings, even though the 2D method (Derech et al.) is the only one using texture information.
Limitations
- Baselines are not designed for eroded, texture-poor real fractures, and all tested methods perform poorly.
- Ground truth is manually produced by archaeologists and the full true solution is unknown, with only about 117 groups annotated.
- Dataset is licensed for non-commercial use with attribution only, and the paper does not report error bars.
Experimental setup
Datasets: RePAIR Breaking Bad
Benchmarks: RePAIR 2D puzzle solving RePAIR 3D puzzle solving Breaking Bad generalization
Metrics: Qpos RMSE rotation RMSE translation Qprecision Qrecall F1
Baselines: Derech et al. archaeological puzzle solver Genetic Optimization Greedy Geometric Matching Global LSTM DGL SE(3)-Equiv. DiffAssemble
Builds on: Crossing Cuts Breaking Bad dataset DiffAssemble SE(3)-Equivariant assembly
Closest papers by meaning
- 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.88)
- PuzzleFusion++: Auto-agglomerative 3D Fracture Assembly by Denoise and Verify (ICLR 2025 · similarity 0.86)
- Scalable Geometric Fracture Assembly via Co-creation Space among Assemblers (AAAI 2024 · similarity 0.84)
- DiffAssemble: A Unified Graph-Diffusion Model for 2D and 3D Reassembly (CVPR 2024 · similarity 0.83)
- Puzzles: Unbounded Video-Depth Augmentation for Scalable End-to-End 3D Reconstruction (NeurIPS 2025 · similarity 0.82)
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 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.