notesum.ai

Published at December 3

Transformer-Metric Loss for CNN-Based Face Recognition

cs.CV

I.5.2

Released Date: December 3, 2024

Authors: Pritesh Prakash¹, Ashish Jacob Sam

Aff.: ¹Central Research Laboratory Bharat Electronics Limited, Ghaziabad, India

Arxiv: http://arxiv.org/pdf/2412.02198v1

Refer to caption

Configuration		High Quality
Train Data	Loss	LFW	AgeDB	CA-LFW	CP-LFW	CFP-FP
MS1MV2	CosFace[13]	99.81	98.11	95.76	92.28	98.12
	Transformer-Cosface	99.83	98.20	96.00	92.55	98.00
MS1MV2	ArcFace	99.83	98.28	95.45	92.08	98.27
	Transformer-Arcface	99.83	98.31	96.16	92.24	97.80
WebFace4M	AdaFace[13]	99.80	97.90	96.05	94.63	99.17
	Transformer-Adaface	99.82	98.02	96.07	93.90	98.88
WebFace4M	CosFace[47]	99.80	97.45	95.95	94.40	99.25
	Transformer-Cosface	99.82	98.00	96.00	93.95	99.00
WebFace4M	ArcFace[5]	99.83	97.95	96.00	94.35	99.19
	Transformer-Arcface	99.82	98.00	96.02	93.98	99.10