Benchmarking Face Recognition without Real Faces

arXiv cs.AI · 6 hours ago

Researchers evaluated whether synthetic face datasets can replace real photographs for benchmarking face recognition models, testing 12 synthetic datasets against 7 real benchmarks using 24 different pre-trained models. The two best synthetic datasets, MorphFace and Vec2Face, achieved agreement levels comparable to the natural disagreement already present among real benchmarks. Well-constructed synthetic datasets can now support reliable evaluation for face recognition, eliminating the need to use real facial data for both model training and performance testing.