biglam/artinsight-painting-deterioration

ArtInsight — Easel Painting Deterioration Detection

20 high-resolution full-frame photographs of easel paintings, annotated by expert restorers with the areas of damage they see. 2,909 annotations in total.

Conservation assessment is normally done by eye, by specialists, one painting at a time. This dataset is an attempt to make it learnable.

Two damage types, two disjoint sets of paintings

Subset	Images	What it marks
LPL	8	Lacuna from Loss of the Painting Layer — paint is gone and the support panel shows through
stucco	12	Filler (animal glue + calcium sulfate) used to level lacunae before chromatic reintegration

No painting appears in both subsets, so subset is effectively a property of the image.

Split	Images	Polygons	Points
train	14	1,137	862
test	6	910	0

Fields

• image — full-frame painting photograph (large; up to ~3700×2800)
• subset — lpl or stucco (ClassLabel)
• objects — list of {polygon, bbox, shape, damage}
  • polygon — [[x, y], ...]
  • bbox — xywh
  • shape — polygon or point (see below)
  • damage — lpl or stucco (ClassLabel)
• file_name, width, height

Caveats

Read this one before you trust the headline number.

• "2,909 polygons" is not accurate. The deposit's readme reports 2,909 polygons. In fact the annotations are 2,047 polygons + 862 VIA point annotations — a point is a single x/y click with no extent. All 862 points sit in stucco/train. They are kept here and tagged via shape, because discarding them would lose annotation, but they are not regions and will produce zero-area boxes. Filter on shape == "polygon" if you need segmentable regions.
• The readme's per-split prose is wrong too — it says "6 train / 2 test" images per subset, but its own file lists (and the data) give lpl 6/2 and stucco 8/4. The file lists are correct.
• Some polygon coordinates fall outside the image frame (negative values appear). They are preserved as deposited rather than clipped.
• 20 images. This is a small, specialist evaluation set, not a training corpus. Expect to fine-tune, not train from scratch.
• The deposit also ships model.h5, a trained damage-detection model. It is not included here — a model belongs in a model repo, not a dataset. See the authors' code at https://github.com/frangam/artdet.

Load

from datasets import load_dataset

ds = load_dataset("biglam/artinsight-painting-deterioration", split="train")

# 862 of the 2,909 annotations are zero-extent VIA points, not regions.
# Filter to real polygons before training a segmentation model.
def polygons_only(row):
    row["objects"] = [o for o in row["objects"] if o["shape"] == "polygon"]
    return row

seg = ds.map(polygons_only)

Source & credit

Francisco M. García-Moreno, José Manuel del Castillo de la Fuente, Luis Rodrigo Rodríguez-Simón, María Visitación Hurtado-Torres. ArtInsight Dataset: Precision Annotations for Easel Paintings' Deterioration Detection. Universidad de Granada. Zenodo, 2025-06-11. https://zenodo.org/records/15640972 — CC-BY-4.0.

This repository parses the VIA annotations into Parquet. Please cite the original authors.

@dataset{garciamoreno_2025_artinsight,
  author    = {Garc{\'i}a-Moreno, Francisco M. and del Castillo de la Fuente, Jos{\'e} Manuel and
               Rodr{\'i}guez-Sim{\'o}n, Luis Rodrigo and Hurtado-Torres, Mar{\'i}a Visitaci{\'o}n},
  title     = {{ArtInsight Dataset: Precision Annotations for Easel Paintings' Deterioration Detection}},
  year      = {2025},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.15640972}
}

Deposit: https://doi.org/10.5281/zenodo.15640972 · Authors' code: https://github.com/frangam/artdet