3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping

Junfeng Cheng, Yingkai Yang, Tania StathakiAAAI 2025 · DOI

Introduction

Archaeological research often involves the painstaking task of sorting through hundreds of ceramic or stone fragments recovered from excavation sites. Before reassembly can begin, researchers must manually group fragments that belong to the same original artifact—a process that is both time-consuming and prone to error. While 3D computer vision has made strides in part assembly (reconstructing a known object from its components), the prerequisite task of grouping fragments from a mixed set remains a significant challenge.

The primary difficulties in archaeological grouping include an unknown number of original objects, complex and irregular fracture geometries, and a lack of large-scale labeled datasets. To address these issues, researchers from Imperial College London have developed 3DPGS, a framework designed to automate the identification of fragment groups from mixed 3D mesh sets.

The 3DPGS framework operates in two distinct stages: Probabilistic Matching Graph Learning (PMGL) and Probabilistic Graph Search (PGS).

The 3DPGS architecture consists of a 3D Feature Extractor, a Relation Net for global context, and a Probabilistic Graph Search (PGS) module to identify groups.

Probabilistic Matching Graph Learning (PMGL)

The system first processes a "mix set" of 3D pieces, represented as point clouds. A 3D Feature Extractor (PointNet) derives geometric features from each piece. These features are then fed into a Relation Net based on EdgeConv Graph Neural Networks (GNNs). Unlike models that analyze pieces in isolation, the Relation Net considers the relationships between all fragments in the set, leveraging global context to determine piece identity. The output is a Probabilistic Matching Graph—an affinity matrix where each entry represents the probability that two specific fragments belong to the same object.

Probabilistic Graph Search (PGS)

In the second stage, the framework extracts distinct groups from the affinity matrix. Using a search algorithm governed by a probability threshold ($T_P$), the system clusters fragments with high mutual affinity. This approach allows the model to handle an arbitrary number of groups without prior knowledge of the total object count.

Datasets and Evaluation

To support this research, the authors introduced ArcPie (Archaeological Piece Dataset). This large-scale dataset features 980 original shapes across 11 classes, including bowls, steles, and human sculptures. Using the "Breaking Good" simulation, the researchers generated 100 distinct fracture patterns per mesh to ensure geometric diversity. ArcPie includes both "Seen" categories for training and "Unseen" categories to evaluate the model's ability to generalize to novel object types.

The study also introduces the Group-Level F1 (GL-F1) metric. Unlike traditional metrics that may award partial credit for incomplete groups, GL-F1 requires a predicted group to achieve a high Jaccard Similarity with the ground truth to be considered a success, providing a more rigorous measure of utility for real-world archaeology.

Experimental Results

3DPGS demonstrates significant improvements over previous state-of-the-art models such as G-FARS and Comp-Net. On the ArcPie "Seen" dataset, 3DPGS achieved a GL-F1 score of 0.806 (at a 0.5 threshold), outperforming G-FARS (0.700).

Qualitative comparison of 3DPGS against baseline methods on ArcPie and BBArtifact datasets, showing superior grouping accuracy on both seen and unseen categories.

Notably, the model maintained strong performance on "Unseen" categories with a score of 0.769, indicating that it learns the underlying logic of fracture patterns rather than simply memorizing specific shapes. Qualitative analysis shows that 3DPGS effectively separates complex fragments that typically cause other models to fail or incorrectly merge groups.

Conclusion and Future Work

3DPGS represents a major step toward automating the initial stages of archaeological reconstruction. By combining global context through GNNs with a flexible probabilistic search, the framework handles the inherent uncertainty of mixed fragment sets. Future research directions include integrating grouping with 6DoF pose estimation for full physical assembly, addressing the "noise" of eroded real-world shards, and deploying these algorithms on robotic systems for field-based artifact sorting.

Summary updated

BibTeX
@inproceedings{cheng2025dpgs,
  author = {Junfeng Cheng and Yingkai Yang and Tania Stathaki},
  title = {{3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i3.32246},
}

Loading related content...

Figure