ItsArca/persian-cats-cat-data

Persian cats from the cat data browser group

A collection of 403 Persian-labeled original images, assembled on 2026-10-28 from the four sources selected in the user's Chrome cat data group. Repository: ItsArca/persian-cats-cat-data. Visibility: private.

The default configuration contains 384 unique images from sources that state a license. The license_unverified configuration contains 18 unique images from a repository that does not specify a license. The annotation_unverified configuration contains 1 image with source annotations flagged for review, when present. Every image retains source labels and provenance; inclusion does not constitute expert breed verification or rights clearance.

Sources and attribution

Source	Persian source records	License and details
FeliBreed V2	158	CC BY 4.0. Mehraj Hossain Mahi; Muhtasimul Hoque Arnab; Anzir Rahman Khan (2025), FeliBreed – A Multi-Breed Cat Image Dataset, Mendeley Data, V2, doi:10.17632/ng6tv57j5c.2. Labels taken from the Persian source folder.
Oxford-IIIT Pet	200	Current official website states CC BY-SA 4.0 and retains original-owner copyright. The original annotation archive README contains an older research-only note and asks users to respect original website terms. Images retrieved through the timm HF mirror; mirror revision is retained per record.
CAT V1 by FOLY, Roboflow	27	Project states CC BY 4.0. 27 Persian-bearing original project images collected. All 78 originals were inspected. The 180 generated version-1 variants are a separate scope and are not included in addition to these originals.
li-yan/cat-by-breed-v1	18	No license, README, or original attribution specified at pinned revision 80be8b7dbd91c947445da487d6e035997d39b1a0. Some filenames refer to stock-photo services. Kept in the separate license_unverified configuration; verify permission before reuse or publication.

Oxford citation: Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and Dogs. IEEE Conference on Computer Vision and Pattern Recognition, 2012.

The collection does not apply one blanket license to all images. Follow the license and attribution of each source record, including CC BY-SA obligations where applicable. Dataset organization, provenance, and QA are newly generated; underlying images and annotations retain their source rights.

Structure and loading

Images are embedded in Parquet as Hugging Face Image features. Downloaded image file bytes are preserved without re-encoding. No browser session data, local machine paths, or credentials are included.

from datasets import load_dataset
data = load_dataset("ItsArca/persian-cats-cat-data", name="default", split="train", token=True)
# Optional, only after reviewing the source-rights limitations:
unverified = load_dataset("ItsArca/persian-cats-cat-data", name="license_unverified", split="train", token=True)

When present, annotation_unverified can be loaded with the same call and config name. Its source metadata records annotation_review_required and quality_flags; small or contradictory source boxes are retained as evidence, not accepted as verified breed labels.

All images use one train pool. This is a collection for downstream dataset design, not an established benchmark split. Original source_split values remain in each sources entry. Repeated animals, related shots, source duplicates, and Roboflow augmentations can leak between random splits; create subject-aware splits before evaluation. A collection containing only the Persian class cannot by itself train or evaluate a Persian-versus-other-breed classifier.

Each row contains image_id, image, label, original decoded dimensions, byte and canonical-pixel SHA-256, a perceptual dHash, source_record_ids, detailed sources, and optional annotations. source-records.jsonl has every 403 collected source record, including records that collapse to a duplicate image. sources.metadata_json preserves additional source fields and attribution. No metadata was inferred from image appearance.

Oxford trimap is an optional embedded image: 1=foreground, 2=background, 3=unclassified. head_bbox retains the source Pascal VOC [xmin, ymin, xmax, ymax] coordinates and names its source record. Roboflow annotations, when available, are preserved in object_annotations_json with their source record and coordinate metadata. These are head boxes or source object annotations, not newly generated whole-cat boxes.

Curation and verification

Source records are merged only when their exact file hashes or EXIF-transposed RGB pixel hashes match. Pixel hashes include width, height, and RGB mode in the hash header. The primary file's original bytes are kept and every source record remains mapped to its output image. Similarity candidates use 64-bit horizontal dHash with Hamming distance <=4, grouped transitively; they are flagged in similarity_group, not automatically removed. Candidates may include unrelated images.

• Collected source records: 403.
• Unique embedded images: 403.
• Exact-byte/pixel duplicate records collapsed: 0.
• Perceptual similarity candidate groups: 2.
• Verified embedded image decodes: 403/403.
• Oxford trimaps retained: 200; head boxes retained: 100.

qa-report.json records counts, checksums, duplicate mappings, source coverage, and similarity candidates. Breed labels are inherited from sources; no expert breed verification was performed.

Similarity flags for earlier images

The two rows in provenance-updates.jsonl add similarity-group flags to previously uploaded image rows. Join by image_id and replace the fields supplied by the overlay. qa-report.json lists the complete groups; near matches are retained, so keep related images together when creating evaluation splits. All source mappings are also retained in source-records.jsonl.