joseplcam/peqemo-artemis

Private, lossless PeqEMO source dataset: 81,192 images, 1,092,603 nested annotations, 1,092,603 native membership records, and 0 auxiliary assets.

Original image bytes and resolutions are preserved as embedded Parquet bytes. Every embedded image was checked against its SHA256. Source fields and nested annotations have matching source/export logical digests in manifest.json. Canonical rows are sorted by sample ID. Storage is approximately 500 MB per shard, Zstandard-compressed.

Load

from datasets import load_dataset, Image
ds = load_dataset("joseplcam/peqemo-artemis", "canonical", split="all", token=True)
stream = load_dataset("joseplcam/peqemo-artemis", "canonical", split="all", token=True, streaming=True)
raw = stream.cast_column("image", Image(decode=False))

all is the complete source collection, not a training split. Hugging Face reserves all as a load keyword: the physical split is named source, and split="all" selects its complete contents in both ordinary and streaming modes. For slices use source[:N]. Native protocols: official. Default source protocol: official. Native default fold: None. splits.parquet is the unmodified source membership table; each row also embeds its native memberships for streaming selection.

Source composition

• ArtEmis v1 (ArtEmis-v1): 454,684 human emotion + explanation annotations on 80,031 paintings.
• ArtEmis v2 (ArtEmis-v2): 237,998 contrastive emotion + explanation annotations on 50,905 target paintings, including the anchor painting in each annotation's payload_json.
• OLA (ArtEmis-OLA): 5,000 objective descriptions.
• Derived records: 394,650 aggregate_votes, distribution, and class_label annotations computed from the native labels.
• Anchor-only images: 271 images referenced only as a v2 contrastive anchor; each has a generated anchor_reference record.

Limitations and provenance

• ArtEmis and ArtEmis v2 are provided for non-commercial research use only.
• Images come from the official WikiArt archive distributed by the ArtEmis authors, in the original wikiart/<art_style>/<painting>.jpg layout.
• The official wikiart.zip has two corrupt entries: Baroque/rembrandt_woman-standing-with-raised-hands.jpg and Post_Impressionism/vincent-van-gogh_l-arlesienne-portrait-of-madame-ginoux-1890.jpg. The Rembrandt is referenced by ArtEmis and is replaced with the byte-identical WikiArt original (same size and CRC32); the Van Gogh is unreferenced and excluded. The rest of the archive is complete.
• The final 1.82 GB of wikiart.zip was fetched from the byte-identical Hugging Face mirror SujieHu/CDS_wikiart after the Google Drive file-level quota was exhausted.
• pixel_sha256 is SHA256 of the decoded RGB pixel bytes (PIL.Image.convert("RGB").tobytes()).
• Painting names are canonicalized to Unicode NFC; rows that appear with NFC/NFD variants across the source CSVs are merged into one painting.
• The released v2 Contrastive.csv has 237,998 rows over 50,905 target paintings; the v2 paper reports 260,533 annotations over 52,933 paintings. This repo preserves the released CSV exactly.

Original collection notes

artemis

81,192 images. 1,092,603 annotations.

• images/: image files (not materialized; original bytes are embedded in the Parquet image column).
• data/all-*.parquet: one row per painting, with its path and list of annotations.
• splits.parquet: original split and annotation memberships.
• source/: the official source CSVs used to build this repo.

from datasets import load_dataset
ds = load_dataset("joseplcam/peqemo-artemis", "canonical", split="all", token=True)
for row in ds:
    print(row["sample_id"], len(row["annotations"]))

Citation

• Achlioptas et al., ArtEmis: Affective Language for Visual Art, CVPR 2021.
• Mohamed et al., It is Okay to Not Be Okay: Overcoming Emotional Bias in Affective Image Captioning by Contrastive Data Collection, CVPR 2022.