notesum.ai

Published at December 9

VariFace: Fair and Diverse Synthetic Dataset Generation for Face Recognition

cs.CV

cs.LG

Released Date: December 9, 2024

Authors: Michael Yeung¹, Toya Teramoto¹, Songtao Wu², Tatsuo Fujiwara¹, Kenji Suzuki¹, Tamaki Kojima¹

Aff.: ¹Sony Group Corporation; ²Sony R&D Center China

Arxiv: http://arxiv.org/pdf/2412.06235v1

Refer to caption

Dataset	Size (ID $\times$ images/ID)	LFW	CFP-FP	CPLFW	AgeDB	CALFW	Average	Real Gap
CASIA-WebFace	0.5M ( $\approx 10.5K\times 47$ )	0.9950	0.9536	0.9007	0.9487	0.9368	0.9470	0.0000
SynFace [53]	0.5M ( $10K\times 50$ )	0.8407	0.6337	0.6355	0.5910	0.6937	0.6789	-0.2681
IDNet* [32]	0.5M ( $10K\times 50$ )	0.9258	0.7540	0.7425	0.6388	0.7990	0.7913	-0.1557
ExFaceGAN* [9]	0.5M ( $10K\times 50$ )	0.9350	0.7384	0.7160	0.7892	0.8298	0.8017	-0.1453
DigiFace* [2]	0.5M ( $10K\times 50$ )	0.9540	0.8740	0.7887	0.7697	0.7862	0.8345	-0.1125
VIGFace* [31]	0.5M ( $10K\times 50$ )	0.9660	0.8666	0.7503	0.8250	0.8342	0.8484	-0.0986
IDiff-Face* [8]	0.5M ( $10K\times 50$ )	0.9800	0.8547	0.8045	0.8643	0.9065	0.8820	-0.0650
DCFace* [30]	0.5M ( $10K\times 50$ )	0.9855	0.8533	0.8262	0.8970	0.9160	0.8956	-0.0514
ID³* [36]	0.5M ( $10K\times 50$ )	0.9768	0.8684	0.8277	0.9100	0.9073	0.8980	-0.0490
Arc2Face* [49]	0.5M ( $10K\times 50$ )	0.9881	0.9187	0.8516	0.9018	0.9263	0.9173	-0.0297
Vec2Face* [72]	0.5M ( $10K\times 50$ )	0.9887	0.8897	0.8547	0.9312	0.9357	0.9200	-0.0270
VariFace (ours)	0.5M ( $10K\times 50$ )	0.9938	0.9460	0.8882	0.9438	0.9305	0.9405	-0.0065