A Community Detection Approach to Cleaning Extremely Large Face Database

Table 1

Comparison with other methods in terms of cleaning MS-Celeb-1M. Ours is able to achieve high cleanness, rich data diversity, and large data scale at the same time.

Method	k	ρ	ζ	Cleanness	Diversity	#Subject	#Image
Ours	-	5	-	93.6%	0.5811	97,646	7,244,505
	-	10	-	97.2%	0.5513	91,180	6,024,931
	-	15	-	98.3%	0.5449	84,297	5,738,961
	-	20	-	98.8%	0.5293	78,399	5,082,658
MSM	-	-	-	98.9%	0.4843	99,892	2,207,688
KCM	5	5	-	68.2%	0.7231	99,544	7,644,424
		10	-	70.1%	0.7047	99,479	7,177,587
		15	-	73.2%	0.6763	99,426	6,291,849
		20	-	74.0%	0.6386	99,384	5,119,874
	7	5	-	68.0%	0.7169	99,304	7,500,139
		10	-	72.0%	0.6816	99,227	6,428,188
		15	-	77.5%	0.6334	99,145	4,892,615
		20	-	82.9%	0.5674	94,812	3,359,597
FPR	-	-	0.2	67.7%	0.6603	99,892	6,225,585
	-	-	0.3	69.0%	0.6319	99,892	5,458,018
	-	-	0.4	75.1%	0.6035	99,892	4,679,810
	-	-	0.5	80.8%	0.5746	99,892	3,890,961
	-	-	0.55	79.5%	0.5610	99,892	3,526,922
	-	-	0.6	80.2%	0.5453	99,892	3,133,376
LCNN	-	-	-	98.8%	0.5062	79,077	5,049,824
None	-	-	-	61.1%	0.7277	99,892	8,456,240