# Oxford-IIIT Pet teaching subset

Dataset: Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman and C. V. Jawahar, **Cats and Dogs**, CVPR 2012.

Source: https://www.robots.ox.ac.uk/~vgg/data/pets/

The current official dataset page states Creative Commons Attribution-ShareAlike 4.0 International: https://creativecommons.org/licenses/by-sa/4.0/ . Copyright in the photographs remains with their original owners. Retain this attribution and license with redistributed assets and apply the same license to adaptations of these dataset assets. No endorsement by the authors or image owners is implied.

The archive's older `source/README.txt` says research purposes only and refers to original image-owner terms. It is retained verbatim for provenance alongside the current official webpage license statement; it is not rewritten as if the archive always contained the newer statement.

`source/*.jpg` and `source/*-trimap.png` are unmodified named members of the official archives. The eight local display PNGs and the RGB/trimap arrays in `subset.json` are resized adaptations generated by `scripts/build-pet-subset.py`, licensed CC BY-SA 4.0. The code itself is not part of this dataset license declaration.

Changes: RGB conversion; aspect-preserving resize to longest side 48; BOX resampling for RGB; nearest-neighbor for trimaps. Trimap 1 means pet foreground, 2 background, and 3 unclassified. Unclassified pixels must be excluded from training loss and scored-pixel denominators and their count reported. Downsampling can lose small details and alter boundaries.

Selection uses eight initial image archive entries, without using model outcomes. The five official trainval images are training examples. The three official test images are openly reused as development examples; this subset offers no untouched test and makes no breed/generalization claim. Pet foreground trimaps are not multi-instance annotations. Source animal-head boxes are not included; boxes later derived from masks must be identified as derived, not original head annotations.

`subset.json` records each source and display file's SHA-256 and byte length, original/resized dimensions, source split, species, source class ID and transformation. Original archive URLs:

- https://thor.robots.ox.ac.uk/pets/images.tar.gz
- https://thor.robots.ox.ac.uk/pets/annotations.tar.gz

Rebuild from official archives using the documented Python script and Pillow 12.3. Only explicitly named archive members are read; no archive paths are extracted to the local filesystem.