Experimental results of image retrieval on Google-Lanmark-V2 dataset

[zzc98]

Experimental results of image retrieval on Google-Lanmark-V2 dataset

Method: ResNet101 + GeM + LinearReduction(2048->512) + ArcfaceHead

set	optim	init lr	epochs	schedule	batch	image size	note	private	public
train clean	SGD	1e-3	10	cos	64	224*224		21.22	24.54
train clean	SGD	1e-3	10	cos	32	224*224		21.47	24.16
train clean	SGD	1e-3	10	cos	32	336*336		23.38	26.55
train clean	SGD	1e-3	10	cos	64	336*336		23.27	26.48
train clean	SGD	1e-3	15	cos	64	336*336		24.02	27.62
train clean	Adam	1e-4	40	cos	64	448*448		30.06	26.95
train clean	Adam	1e-4	40	cos	64*8	448*448		30.34	27.29
train clean	SGD	1e-2	10	cos	32	448*448	DOLG	24.85	28.10
train clean	SGD	1e-3	10	cos	32	448*448	DOLG	25.62	28.35
train clean	SGD	1e-4	10	cos	32	448*448	DOLG	18.74	21.59