NguyenDinhHieu/Artifact-PCDiff-Data

Pre-Columbian Artifact Point Clouds for 3D Completion

Dataset Summary

This processed 3D point-cloud dataset supports point-cloud completion and 3D reconstruction research on Pre-Columbian cultural heritage objects. Each example pairs an incomplete artifact scan with its removed-region target, providing training data for geometric restoration with Artifact-PCDiff.

The dataset is derived from mesh collections organized in the processing repository under SN_bowl, native, scanned, and MD40_bowl. The processing scripts and detailed generation instructions are available in the Artifact-PCDiff repository.

Processing

The documented processing pipeline performs these steps:

1. Samples each source mesh using Poisson-disk sampling, with 200,000 points by default.
2. Normalizes the sampled geometry into a 512-unit cube. The default transform centers the selected X and Z axes and places the object relative to the floor. These coordinates are normalized values, not physical measurements.
3. Creates train and test manifests. The default split assigns the scanned source group to test and randomly splits the other groups 70/30, unless a source already defines a split. The released CSV manifests should be treated as the authoritative split.
4. Simulates a missing lower region by removing about 20% of the object's vertical extent, with 10% variation. One broken/repair pair is generated per object by default.

Files and Schema

For use with the current Artifact-PCDiff training and evaluation code, arrange the runtime files as follows:

data/precol/
├── train.csv
├── test.csv
├── complete/<object_id>.obj
├── broken/<object_id>_0.npy
└── repair/<object_id>_0.npy

• train.csv and test.csv contain one complete-mesh path per row. Each path must include complete and point to the unsuffixed .obj mesh.
• complete/<object_id>.obj is the complete reference mesh.
• broken/<object_id>_0.npy contains the observed partial point cloud as XYZ coordinates.
• repair/<object_id>_0.npy contains the removed ground-truth region as XYZ coordinates.
• The object ID must match across the mesh and both point-cloud files. The current loader derives the matching _0.npy paths from the complete-mesh path.

The current loader reads the broken and repair arrays from .npy files. Any .ply files included in a release are visualization exports; they do not replace the paired .npy files required by the current training and metric-evaluation pipeline.

Intended Use

The data is intended for research on 3D point-cloud completion and geometric restoration. Training uses both the broken input and repair target. Inference uses the broken input. Full metric evaluation also requires the repair target.

Limitations

The generated damage removes points from the bottom region of an object. It does not represent the full range of real archaeological breakage, such as irregular fractures, multiple missing regions, erosion, or material degradation. The default split also places the scanned source group in the test set, so results can reflect differences between source groups as well as model generalization.

License and Provenance

No license is specified here for the dataset files. The MIT license in the code repository applies to its software and does not automatically license the source meshes or derived data. Confirm the terms of the original mesh collections before assigning a dataset license or redistributing the files.

Citation

If you use this dataset, cite the dataset repository and the associated method:

Nguyen Dinh Hieu and Phan Duy Hung. Artifact-PCDiff: Conditional Point-Cloud Diffusion for Missing-Region Completion of Fragmented 3D Artifacts.

Code and processing scripts: https://github.com/nguyendinhhieu1309/Artifact-PCDiff.