biglam/chambord-mortar-joint-segmentation

Château de Chambord — Mortar Joint Segmentation

425 image patches (512×512) cut from high-resolution photographs of the Château de Chambord, France, each paired with a binary mask marking the mortar joints between stones.

Built-heritage conservation needs to measure how mortar joints erode over time. That means segmenting them first.

Split	Patches
train	333
test	92

Fields

• image — 512×512 RGB patch
• mask — 512×512 binary mask (mode L, values {0, 255}); white = mortar joint
• file_name

The dataset covers flat-relief, vertical-relief and horizontal-relief façade configurations.

How the masks were made

Stone masks were generated with a custom QGIS plugin built on SAM (Segment Anything Model), manually curated by heritage experts, then inverted to obtain the mortar-joint labels. So the annotations are model-assisted and human-corrected, not drawn from scratch.

Caveats

• This is built heritage (a building), not a collection object — a different kind of cultural-heritage data than manuscripts or artworks.
• Masks are derived by inversion from stone masks, so mortar-joint boundaries inherit whatever imprecision the stone segmentation had.
• All patches come from a single monument. Expect limited generalisation to other masonry, stone types, or lighting.

Load

from datasets import load_dataset

ds = load_dataset("biglam/chambord-mortar-joint-segmentation", split="train")

row = ds[0]
row["image"].size, row["mask"].size   # 512x512 patch and its binary joint mask

Source & credit

Sylvie Treuillet, Xavier Brunetaud, Xavier Desquesnes et al. Château de Chambord mortar joint segmentation dataset. Zenodo, 2026-03-14. https://zenodo.org/records/19025457 — CC-BY-4.0.

This repository converts the deposit to Parquet. Please cite the original authors.

@dataset{treuillet_2026_chambord_mortar,
  author    = {Treuillet, Sylvie and Brunetaud, Xavier and Desquesnes, Xavier and Lucho, Stuardo},
  title     = {{Ch{\^a}teau de Chambord mortar joint segmentation dataset}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19025457}
}

Deposit: https://doi.org/10.5281/zenodo.19025457