Photo corpus¶

The robustness and separability charts on the algorithm pages are drawn over two corpora. One is the synthetic corpus the tests use (verification methodology). The other is this one: 200 photographs from Wikimedia Commons, each released by its author under CC0 or into the public domain, as its own page on Commons states.

They were drawn from the files Commons' reviewers rated Quality images, spread over subjects (animals, architecture, arts, food, landscapes, nature, objects, people, text, transport) so that no one kind of picture dominates, with at most 3 from any one author. The site measures Commons' standard rendition of each, 1280 pixels wide (960 for a portrait), about a megapixel.

The files are not part of the repository and not part of the library's tests. tools/site/corpus_photos.tsv lists them with the SHA-256 of each, and the site's build downloads them once into a cache, from a copy kept for the site, and checks every file against its digest, so the charts are drawn from exactly the files listed here.