Datasetpublic

biglam/odor-olfactory-artwork-detection

ODOR — Object Detection for Olfactory References in Artworks 4,712 artwork images with 38,165 bounding-box annotations across 139 fine-grained categories of smell-related objects — flowers, fruit, censers, animals, vessels — drawn from European art. Computer vision on artworks is hard in ways photographic benchmarks are not: artistic abstraction, peripheral objects, and fine-grained distinctions between visually similar classes. ODOR is built to test exactly that.… See the full description on the dataset page: https://huggingface.co/datasets/biglam/odor-olfactory-artwork-detection.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

0likes46downloads

Dataset Card

ODOR — Object Detection for Olfactory References in Artworks

4,712 artwork images with 38,165 bounding-box annotations across 139 fine-grained categories of smell-related objects — flowers, fruit, censers, animals, vessels — drawn from European art.

Computer vision on artworks is hard in ways photographic benchmarks are not: artistic abstraction, peripheral objects, and fine-grained distinctions between visually similar classes. ODOR is built to test exactly that.

What's in a row

Each row carries the image, its detections, and its art-historical metadata — the Zenodo deposit ships these as separate files, and they are joined here:

—image, image_id, file_name, width, height
—objects — bbox (COCO xywh), category (ClassLabel, 139 names), area, iscrowd
—artist, title, iconography, `iconclass_code`, genre, material
—earliest_date, latest_date, photo_archive, image_credits, details_url, description, keywords, language

Where present, iconclass_code connects this dataset to `biglam/brill_iconclass` — but note it is sparse (see caveats).

Splits

Split	Images
train	4,264
test	448

Caveats

—Metadata coverage is uneven. artist is present on 96.6% of rows, but `iconclass_code` on only 541 of 4,264 (12.7%). Do not assume Iconclass coverage.
—Metadata is multilingual (language is de for much of the Städel material) and not normalised across source archives.
—earliest_date / latest_date are inconsistent in the source — some are years, some centuries (e.g. "18"). Left as strings rather than guessed at.
—Images come from several photo archives with differing terms; image_credits and details_url point back to the originals.

Source & credit

Mathias Zinnen, Prathmesh Madhu, Ronak Kosti et al. The Object Detection for Olfactory References (ODOR) Dataset (v3.0.2). Zenodo, 2024-04-26. <https://zenodo.org/records/11070878> — CC-BY-4.0.

Produced in the context of the Odeuropa project. This repository joins the deposit's COCO annotations with its meta.csv and converts to Parquet. Please cite the original authors.