Researchers introduce TPO, a face-free presentation-attack-detection dataset built from recordings of vegetables subjected to print, replay, and recapture processes. Using foundation models, detectors trained on TPO achieved 92.70% average AUC across four face presentation-attack-detection benchmarks, outperforming training on synthetic faces and remaining competitive with models trained on real face datasets. The study shows the learned representations capture presentation-process characteristics rather than object semantics, and that incorporating face-free data improves cross-dataset performance.
