Research map / Graph learning / Graph algorithms and clustering
3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping
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 the Archaeological Piece Grouping benchmark with the ArcPie dataset (980 fractured 3D shapes, 11 classes) and 3DPGS, a probabilistic matching graph plus graph-search method for grouping mixed 3D fragments, with a new Group-Level F1 metric.
Problem
Separating mixed, shuffled 3D archaeological fragments into their original object groups when the number of groups is unknown, with a lack of suitable archaeological fracture data and metrics that ignore group completeness.
Approach
Learn a probabilistic matching graph from pairwise dot products of relation-network features over all pieces, then group pieces via a threshold-based graph search (PGS) rather than autoregressive sampling; also introduce GL-F1 to score group completeness.
Key results
- On ArcPie Seen, 3DPGS reaches PL-F1 ssa 0.854 and GL-F1 0.5 of 0.806, versus 0.826 and 0.700 for the prior best G-FARS.
- On ArcPie Unseen, 3DPGS achieves GL-F1 0.5 of 0.769 versus 0.646 for G-FARS.
- On BBArtifact, 3DPGS achieves GL-F1 0.5 of 0.727 versus 0.598 for G-FARS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 3DPGS instead of G-FARS | 3DPGS · ArcPie Seen | GL-F1 0.5: 0.700 -> 0.806 | Helped |
| use 3DPGS instead of G-FARS | 3DPGS · ArcPie Unseen | GL-F1 0.5: 0.646 -> 0.769 | Helped |
| use 3DPGS instead of G-FARS | 3DPGS · BBArtifact | GL-F1 0.5: 0.598 -> 0.727 | Helped |
| add the relation net (ablation) | 3DPGS · ArcPie Seen | PL-F1 ssa: 0.693 -> 0.854 | Helped |
| add the relation net (ablation) | 3DPGS · ArcPie Unseen | PL-F1 ssa: 0.681 -> 0.829 | Helped |
| use dot product rather than an MLP to produce the matching graph (ablation) | 3DPGS · ArcPie Seen | PL-F1 ssa: 0.845 -> 0.854 | Helped |
| use dot product rather than an MLP to produce the matching graph (ablation) | 3DPGS · ArcPie Unseen | PL-F1 ssa: 0.818 -> 0.829 | Helped |
| set the probability threshold to 0.65 or 0.75 (ablation) | 3DPGS PGS · ArcPie | grouping performance across reported metrics | Helped |
Surprising findings
- Removing the relation net causes a large drop (e.g., Seen GL-F1 0.75 falls from 0.547 to 0.163), and replacing the dot product with an MLP also lowers scores.
Limitations
- Group-level F1 cannot be replaced by Average Precision because the model confidence scores are not valid, so only F1 is reported.
- Performance is sensitive to the probability threshold TP, which must be tuned (0.75 used).
Experimental setup
Datasets: ArcPie Breaking Bad Artifact (BBArtifact)
Benchmarks: ArcPie Seen ArcPie Unseen BBArtifact
Metrics: PL-F1 (ssa/oa) GL-F1 0.5 GL-F1 0.75 GL-F1 [0.5:0.95]
Baselines: GRU-Mask Comp-Net G-FARS
Builds on: G-FARS Breaking Bad dataset EdgeConv PointNet
Closest papers by meaning
- Re-assembling the past: The RePAIR dataset and benchmark for real world 2D and 3D puzzle solving (NeurIPS 2024 · similarity 0.88)
- VaseVQA-3D: Benchmarking 3D VLMs on Ancient Greek Pottery (ICLR 2026 · similarity 0.84)
- Scalable Geometric Fracture Assembly via Co-creation Space among Assemblers (AAAI 2024 · similarity 0.84)
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.83)
- Component-Level Segmentation for Oracle Bone Inscription Decipherment (AAAI 2025 · similarity 0.82)
- Archaeoscape: Bringing Aerial Laser Scanning Archaeology to the Deep Learning Era (NeurIPS 2024 · 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.