A Community Detection Approach to Cleaning Extremely Large Face Database

Table 2

Performance (in %) of the Inception-ResNet-v1 face recognition model trained with 5 different cleaning results.

Model	Training data	#Image	Pairwise verification		Close-set identification	Open-set identification
			Acc	VR@FAR = 0.1%	Rank-1	DIR@FAR = 1%	DIR@FAR = 0.1%
	LCNN result	5,049,824	99.63	99.43	98.06	94.01	82.08
	MS-Celeb-1M	8,456,240	99.40	97.97	96.50	86.36	59.31
	MS-C-1-noR	4,269,900	99.53	99.13	98.03	93.34	74.88
	MS-C-1	6,024,931	99.60	99.30	98.13	93.73	81.66
	C-MS-Celeb	6,464,016	99.67	99.53	98.31	95.40	86.31