screenpad-data

Data for the screenpad take-home PAD project: 1000 bona fide face images (Bonafide/, apparently FFHQ, CC BY-NC-SA 4.0) and 23 phone recaptures of some of them shown on a screen (Screens/, same filename as the source image; iPhone 15 x14, Galaxy A12 x9). The attack images were provided by the assessors for the assignment.

Path	Contents	Fetch
raw/screenpad-raw.tar, raw/screenpad-raw.sha256	the original files, byte-identical, in one uncompressed tar, with per-file SHA-256	screenpad pull-data (verifies every file)
data/	the same files uploaded loose (first upload; slow to fetch, kept for browsing)
cache/crops_{loose,tight}_384[_aligned].npz	preprocessed face crops (384 px; loose = 1.25 x, tight = 1.0 x the face box; aligned = eye line levelled), each verified byte-exact against a fresh recompute	screenpad pull-data --cache
cache/synth_{loose,tight}_384_train-val.npz, cache/synth_params_384_train-val.csv	synthetic screen recaptures and camera-only controls of train/validation faces, with every sampled simulator parameter	screenpad pull-data --cache
cache/shrink_{loose,tight}_384_aligned.npz	resized bona fide crops for the resampling-shortcut test

Splits (by image ID) are made by screenpad splits; no test image is used to make any cache except the plain crops.